The Economics of Quality Investment in Mobile Telecommunications

Patrick Kainin Sun

Submitted in partial fulﬁllment of the requirements for the degree of Doctor of Philosophy in the Graduate School of Arts and Sciences

COLUMBIA UNIVERSITY

Patrick Kainin Sun

This dissertation studies the U.S. mobile telecommunications industry, with a particular emphasis on the incentive to maintain antenna facilities, or base stations, to produce better signal quality. It combines insights from economic analysis to draw inferences from unique datasets for the state of Connecticut. Chapter 1 gives a broad overview of the industry and highlights the apparent importance of signal quality as a driver of demand. Publicly available information reveals that plan features, phone selection, and pricing seem to be less determinant of overall quality relative to the quality of the call network. Reduced form evidence from proprietary data on demand and base station location data from Connecticut confirm that signal quality is important and that base stations are important to signal quality. However, the analysis reveals base station numbers themselves are not the only thing important for signal quality but likely interact with other carrier characteristics such as spectrum, transmission technology, and network management. Given the importance of base stations, Chapter 2 asks what are the competitive incentives to provide them and how would these incentive change in proposed mergers between two of the four largest firms in this industry. To answer this question, I use proprietary demand data and base station locations to estimate a structural model of supply and demand in this industry. The analysis improves on the analysis in Chapter 1 by incorporating time- varying preference heterogeneity and by controlling for the endogeneity of quality through an instrumental variables strategy. I use a measure of land use regulation stringency from data on Connecticut zoning codes as instruments for the costliness of construction. Instrumenting matters, as it increases baseline estimates of quality sensitivity by a factor of 2. Overall, I find base stations to have important competitive implications, as they represent a significant proportion of costs. Findings also indicate base stations are strategic substitutes: if one firm increases the number of their base stations, their rivals will have a lower incentive to maintain base stations. Simulating mergers between AT&T and T-Mobile and Sprint and T-Mobile, I find these mergers to be generally consumer welfare reducing without efficiencies. In particular, and consistent with previous literature, I find the mergers induce increased differentiation between merging partners, so much so that one partner’s quality is severely degraded. However, the natural efficiency of being able to use a single network instead of two can make the mergers welfare-improving if resulting price increases are not too high. This result implies that merger reviews in industries with networks should investigate the scope of network integration as potentially important efficiency. Chapter 3 expands on the instrumental variable in Chapter 2 and explores how exactly land use regulation impacts the incentives of firms to invest across different jurisdictions. Through more costly requirements, land use regulations can discourage firm investment. But land use regulation can also encourage investment through reducing legal ambiguity and thus the risk of investment hold-up through legal technicalities. To test these ideas and to control for unobserved location heterogeneity, I use a border discontinuity approach that looks at sites placed near town borders and compares the relative stringency of regulation between bordering towns. To measure the stringency of regulation and to separate out different regulations, I use both researcher-coded measures from manual inspection of the regulations and measures derived from the computational linguistic technique of Latent Dirichlet Allo- cation topic modeling. I confirm that regulations that impose more requirements on things like landscaping, electric power and signal strength, and impose time limits for approvals are associated with fewer sites. I also confirm that regulations that improve clarity, particularly the acknowledgment of federal standards and regulations, are associated with more sites. This has the important policy implication that the Federal Telecommunications Act of 1996 may have helped reduce investment risk. A simple counterfactual shows that regulation has only a modest effect on the reallocation of facilities across towns, holding the number of facilities fixed. Overall, these chapters clarify the role and costs of an important kind of quality provision in a major industry. They contribute significant insight for policy in both merger review and land use regulation. The second chapter is the first paper to treat signal quality as an endogenous characteristic in a study of the wireless telecommunications industry. The last chapter introduces to the economics literature topic modeling in the analysis of regulatory effects and statutory clarity as a regulatory concern. Table of Contents

List of Figures ii

List of Tables iii

Acknowledgments iv

1 The Mobile Telecommunications Industry in Connecticut 1 1.1 Introduction ...... 2 1.2 Related Literature ...... 3 1.3 Industry Background ...... 4 1.3.1 Industry History ...... 4 1.3.2 Available Features and Services ...... 8 1.3.3 Pricing ...... 9 1.3.4 Handset Selection ...... 12 1.3.5 Signal Quality ...... 14 1.3.6 Base Stations and Sites ...... 14 1.3.7 Transmission Technology, Spectrum, and Network Management . . . 19 1.4 Connecticut Data ...... 22 1.4.1 Base Station Data ...... 22 1.4.2 Zoning Codes ...... 23

i 1.4.3 Proprietary Demand Data ...... 25 1.5 State-Wide Demand Analysis ...... 28 1.6 Sites and Base Stations ...... 36 1.7 Evidence on the Relationship between Signal Quality and Base Stations . . . 46 1.7.1 Data on Signal Quality and Base Stations ...... 47 1.7.2 Log Base Station Density ...... 51 1.7.3 Demand for Signal Quality and Base Stations ...... 53 1.8 Conclusion ...... 65

2 Quality Competition in Mobile Telecommunications: Evidence from Con- necticut 67 2.1 Introduction ...... 68 2.2 Competitive Effects of Quality ...... 73 2.3 The Industry Model ...... 83 2.3.1 Demand ...... 84 2.3.2 Supply ...... 89 2.3.3 Caveats ...... 92 2.4 Data ...... 94 2.5 Estimation and Results ...... 96 2.5.1 Endogeneity of Quality ...... 96 2.5.2 Demand Estimation Procedure ...... 97 2.5.3 Results: Individual Identified Demand Parameters ...... 100 2.5.4 Results: Quality Sensitivity Parameters and Brand-Year Effects . . . 104 2.5.5 Supply Side Estimation and Results ...... 113 2.6 Counterfactuals: Mergers of a Small Carrier ...... 118 2.6.1 Discontinue All Products from Purchased Firm (*) ...... 123

ii 2.6.2 Retain Products from Purchased Firm with Separate Networks (**) . 124 2.6.3 Retain Products from Purchased Firm with Fully Integrated Networks (***) ...... 132 2.7 Conclusion ...... 135

3 The Costs and Clarifying Effects of Regulation for Business Investment: Evidence from Cell Siting 137 3.1 Introduction ...... 138 3.2 The Impact of Land Use Regulation ...... 141 3.2.1 Burdens ...... 141 3.2.2 Clarification ...... 142 3.2.3 Causes of Inefficient Regulation Stringency ...... 146 3.3 A Model of Site Location Choice Under Regulatory Opportunism and NIM- BYism ...... 148 3.3.1 Setup of the General Model ...... 150 3.3.2 The Value of Zoning Codes: The One Town Case ...... 153 3.3.3 Zoning Codes in Equilibrium: Solution of the Multiple Town Case . . 158 3.4 Econometric Specification ...... 163 3.4.1 Conditional Logit ...... 166 3.4.2 Potential Omitted Variable Bias ...... 168 3.4.3 Border Discontinuity ...... 169 3.5 Data ...... 172 3.6 Measuring Stringency of Regulation ...... 175 3.6.1 The Model of LDA ...... 181 3.6.2 Estimation of LDA ...... 183 3.6.3 Estimated Topics ...... 187

iii 3.7 Estimation ...... 194 3.8 Results ...... 198 3.8.1 Manually Coded Results ...... 199 3.8.2 Topic Focus Results ...... 202 3.8.3 Topic Presence Results ...... 206 3.8.4 Combination of Regulation Measures Results ...... 211 3.8.5 Fifty Topics Results ...... 217 3.9 Discussion ...... 218 3.10 Conclusion ...... 226

Bibliography 228

Appendices 244

Appendix A Appendix for Chapter 1 245 A.1 A Simple Model of Quality and Base Station Density ...... 245

Appendix B Appendix for Chapter 2 249 B.1 Comparative Statics of the Example with More than 2 Firms ...... 249 B.2 General Comparative Statics of the Merger Scenarios ...... 251 B.3 Comparative Statics Under Multi-Product Logit Demand Model ...... 254 B.4 Additional Tables and Figures ...... 256

Appendix C Appendix for Chapter 3 259 C.1 LDA Algorithm Details ...... 259 C.1.1 Text Transformation ...... 259 C.1.2 Estimating Algorithm ...... 263 C.1.3 Town-Level Topic Distributions ...... 264

iv C.2 Additional Tables ...... 266 C.3 Results with 50 Topics ...... 270

v List of Figures

1.1 Consolidation in U.S. Telecom ...... 5 1.2 Average Revenue Per User by Carrier and Plan Type ...... 11 1.3 Average Handset Characteristics over the Sample Period ...... 12 1.4 UBS Estimated Total Sites by Firm ...... 18 1.5 From page 53 of 17th Annual Wireless Competition Report (2014)...... 20 1.6 Estimated Market Shares over the Sample Period ...... 30 1.7 Estimated Carrier Prepaid Shares over the Sample Period ...... 31 1.8 Estimated Carrier Shares by Income Brackets ...... 32 1.9 Nielsen Estimated Sample Market Shares by Five-Year Age Brackets . . . . 34 1.10 Site Ownership Fraction over Time ...... 38 1.11 Base Stations over Time ...... 40 1.12 2000 PUMAs - Subdivided into Block Groups Colored by 2010 Population Density ...... 41 1.13 Log Base Station Density on Quality Rating ...... 54 1.14 Quality Ratings on Log Shares ...... 55 1.15 Log Base Station Density on Log Shares ...... 56

2.1 Quality as Strategic Substitute and Strategic Complement ...... 77 2.2 Mean Quality on Base Station Density - Instrumented Pure Logit ...... 109

vi 3.1 Regulatory Application Timing ...... 152 3.2 Four Towns ...... 169 3.3 Histograms of the Core Index and the Full Index ...... 221

A.1 Cell Diagram - Macdonald (1979) ...... 246 A.2 Each regular hexagon can be divided into 12 right triangles...... 246

C.1 Value of tf-idf of Stem Over its Descending Ranking ...... 262

vii List of Tables

1.1 Primary Reason for Switching Carriers ...... 7 1.2 Top10Review.com - 2012 Features and Prices ...... 10 1.3 Top10Review.com - 2015 Reviews of Quality ...... 15 1.4 2008-2012 Unweighted Responses by Demographics Compared to 2008-2012 American Community Survey for Connecticut ...... 26 1.5 2008-2012 Carrier Shares ...... 28 1.6 2008-2011 Carrier Shares by Age Group ...... 33 1.7 2008-2011 Carrier Shares by Gender ...... 33 1.8 2008-2011 Carrier Shares by Family or Alone ...... 35 1.9 Site Statistics over Time ...... 37 1.10 PUMA Market Shares Summary Statistics over All Years ...... 42 1.11 Count and Density of Base Stations by Market ...... 43 1.12 Spearman Correlation Coeﬃcient of Unweighted Sample Market Shares over Markets ...... 44 1.13 Spearman Correlation Coeﬃcient of Base Station Density over Markets . . . 44 1.14 Rank Concordance of Base Station Density and Unweighted Sample Market Shares over Markets ...... 44 1.15 Individual-Level Quality Ratings from Nielsen Sample ...... 48 1.16 PUMA-Mean Quality Ratings from Nielsen Sample ...... 49

viii 1.17 Correlations Between Rating Measures ...... 50 1.18 First Stage: Log Base Station Density on Log Shares and Quality Measures . 60 1.19 Demand and Production System Estimates: Unrestricted Parameters . . . . 62 1.20 Demand and Production System Estimates: Restricted Parameters . . . . . 64

2.1 Example Model Results ...... 82 2.2 Individual Level Identified Coefficients from MLE ...... 101 2.3 Prepaid-Carrier-Year Fixed Effects from Pure Logit ...... 103 2.4 Mean Plan-Type-Carrier-Consumer Characteristic Effects over Years . . . . 105 2.5 Instrument Strength ...... 106 2.6 Signal Quality Sensitivity Estimates ...... 107 2.7 Median Quality Elasticities for Instrumented Pure Logit Specification . . . . 111

2.8 Marginal Monthly Base Station Cost - Fkmt ($1000) ...... 114 2.9 Total CT Costs as % of Variable Proﬁt by Year ...... 115 2.10 Cost Projected onto Covariates ...... 117 2.11 Dropping T-Mobile Product Line ...... 122 2.12 AT&T Buys T-Mobile, Separate Networks ...... 125 2.13 Sprint Buys T-Mobile, Separate Networks ...... 126 2.14 AT&T Buys T-Mobile, Single Network, AT&T Costs ...... 128 2.15 AT&T Buys T-Mobile, Single Network, T-Mobile Costs ...... 129 2.16 Sprint Buys T-Mobile, Single Network, Sprint Costs ...... 130 2.17 Sprint Buys T-Mobile, Single Network, T-Mobile Costs ...... 131

3.1 Facilities by Consideration Set ...... 173 3.2 Town Characteristics ...... 176 3.3 Coded Regulation Measures ...... 180 3.4 Perplexity of the Gibbs Sampler Chains ...... 187

ix 3.5 Estimated Topics ...... 188 3.6 Collocation Correlation in 2013 ...... 195 3.7 Results: Coded Regulation Measures ...... 200 3.8 Results: Topic Focus: Share of Tokens ...... 203 3.9 Results: Topic Presence: 97.5th Percentile ...... 207 3.10 Results: All Regulation Measures ...... 212 3.11 Measures of Regulation Stringency on Town Characteristics ...... 222

B.1 Hausman-McFadden Tests of Independence of Irrelevant Alternatives for the Demand Model ...... 256 B.2 Postpaid-Carrier-Year Fixed Eﬀects from Pure Logit ...... 257 B.3 Correlation in Mean Utility Attributable to Only Demographics and Years . 258

C.1 Multigram Replacement List ...... 260 C.2 Correlation Matrix of Coded Regulation Variables ...... 266 C.3 Correlation Matrix of Topic Focus ...... 267 C.4 Correlation Matrix of Topic Volume ...... 268 C.5 Correlation Matrix of Topic Presence ...... 269 C.6 Perplexity of the Gibbs Sampler Chains - 50 Topics ...... 270 C.7 Estimated Topics - K =50...... 271 C.8 Results: Topic Focus: Share of Tokens - 50 Topics ...... 273 C.9 Results: Topic Presence: 97.5th Percentile - 50 Topics ...... 275

x Acknowledgments

I am foremost thankful to my family for their support in my graduate studies. I especially thank my late father, who, upon learning about my selection of economics as my undergraduate major, first suggested the pursuit of graduate studies in the field. In addition, I am immensely thankful to my advisers and colleagues for their constant support during the creation of this dissertation and throughout my graduate education. For guidance and advice, I thank my sponsor Michael Riordan, Kate Ho, Chris Conlon, Andrea Prat, Debasis Mitra, Stephen Hansen, David Blei, Bernard Salanié,Christophe Rothe, Mikka Rokkanen, Awi Federgruen, Wojciech Kopczuk, Colin Hottman, Keshav Dogra, Michael Mueller-Smith, Ilton Soares, Zach Brown, Elliot Ash, Keeyoung Rhee, Alejo Czerwonko, Donald Ngwe, Hyelim Son, Ju-Hyun Kim, Jonathan Dingel, Jessie Handbury, Joe Hogan and all attendees of the Columbia Industrial Organization Colloquium. I thank Christopher Mueller-Smith for his insight into the technical aspects of wireless telephony. I also thank Christopher Wieczorak and Ken Schmidt for their insight into the wireless industry. I thank John Hodulik for providing the UBS wireless model. For assistance with the use of GIS, I thank Jeremiah Trinidad-Christensen and Eric Glass. I thank Kei Kawai and Matt Weinberg for comments on Chapter 2. I gratefully acknowledge the financial support of the research in Chapter 2 by the NET Institute, http://www.NETinst.org. I thank Victor Glass for first suggesting the use of computation linguistics for the analysis in Chapter 3. All errors in this dissertation are mine.

xi To my father, whom we will never forget.

xii Chapter 1

The Mobile Telecommunications Industry in Connecticut

Patrick Kainin Sun

1 1.1 Introduction

The wireless industry is an important part of the U.S. economy, with $189 billion of revenue in 2013.1 Not only does wireless service provide significant benefits to society from direct consumption, but also it improves the productivity of other economic activities by reducing information frictions.2 An important policy consideration for this industry is its concentration. Four firms, AT&T, Sprint, Verizon and T-Mobile, have approximately 93% national market share.3 It is not clear at this high concentration whether the carriers exert competitive pressure on each other to provide greater value to consumers and, if so, how great that pressure is. Moreover, the carriers appear eager to undergo further consolidation. AT&T attempted to merge with T-Mobile in 2011 before facing opposition from antitrust authorities later in the year. T- Mobile merged with the fifth largest carrier MetroPCS in 2013. During the middle of 2014 Sprint and T-Mobile discussed merging, but allegedly called off the effort due to expected antitrust opposition.4 To understand what competitive pressures are present in this environment it is important to understand what drives demand for wireless service. Therefore, I analyze in this first chapter public and proprietary data about carrier services and investments to understand

1From “CTIA-The Wireless Association, CTIA’s Wireless Industry Summary Report, Year- End 2013 Results, 2014.” See http://www.ctia.org/your-wireless-life/how-wireless-works/ annual-wireless-industry-survey.

2 A literature, surveyed by Aker and Mbiti (2010), shows how wireless telephones have signiﬁcantly reduced information frictions in markets in developing countries. R¨ollerand Waverman (2001) gives evidence on the impact of telecommunications infrastructure on aggregate productivity.

3See “6 years after the iPhone launched, just 4 big carriers are left standing,” http://venturebeat.com/ 2013/07/08/iphone-carrier-consolidation/, July 8, 2013.

4See “Sprint Abandons Pursuit of T-Mobile, Replaces CEO,” Wall Street Journal, August 5, 2014, http://online.wsj.com/articles/sprint-abandoning-pursuit-of-t-mobile-1407279448.

2 what consumers seem to care about and how firms have differentiated themselves to appeal to consumers. In particular, I match survey demand data to a unique dataset on the location and ownership of antenna facilities, or base stations. Using these datasets, I conclude that signal quality is an important determinant of demand for the four major carriers and that base station investment has a strong role in determining that aspect of quality. However, base stations alone do not seem to determine signal quality, as spectrum, network management, and transmission technology seem to be at least as important. However, these conclusions are tentative as there are issues of selection, precision, and endogeneity in the estimation of these effects. To overcome these issues, and to account for what appears to be changing tastes for smaller carriers and different plan types, I analyze demand in a more structural framework in the second chapter of this dissertation, which will also allow me to directly speak to the issue of consolidation.

1.2 Related Literature

Due to its importance, wireless has been the subject of numerous empirical studies. Among the earliest is Hausman (1999), which attempts to quantify the bias in the US CPI from the exclusion of mobile phones from the index. Busse (2000) and Miravete and R¨oller (2004) study the early U.S. industry in which the U.S. Federal Communications Commission (FCC) restricted each market to a duopoly. Macher et al. (2013) studies the substitution and complementarity of ﬁxed and wireless lines, important given how much wireless service has supplanted wireline in recent years. As the carrier-customer relationship is often mediated by contract, there is some recent literature using wireless phone data to test contract theory. Recent examples include Luo (2011, 2012) and Luo, Perrigne, and Vuong (2011). The long-term contracting environment also provides a laboratory for studying dynamic optimization. For example, Yao et al. (2012)

3 use mobile phone contracts to estimate discount rates, while Jiang (2013) and Grubb and Osborne (2015) show errors in dynamic optimization of minutes usage. The supply side of the industry also been used to study contracts: Zhu, Liu, and Chintagunta (2011) and Sinkinson (2014) both study the value of the exclusivity of the iPhone to AT&T.

1.3 Industry Background

1.3.1 Industry History

Wireless telephony began in the United States dating back to experiments in the 1950s.5 The phone system relied on a few long-range towers and cumbersome personal receivers that made usage impractical for all but the very rich. The most restrictive aspect of the system was that only a single call could occupy the same spectrum of transmission frequency - at most only a few hundred users in an area could use the spectrum before a carrier’s licensed bandwidth would become saturated. This restriction was greatly loosened through the realization in Macdonald (1979) that a ﬁrm could use the same bandwidth on separate base stations. Thus with many short- ranged base stations, one could reuse the bandwidth over and over and serve many more individuals. Service between these areas could be seamless using computer algorithms to switch base stations as callers move. Dividing one’s service areas into small “cells” around towers became known as “cellular” service, the name still used in the U.S. used for this kind of service today. In 1982, the FCC instituted a duopolistic industry structure by assigning spectrum to 1) a local wireline provider and 2) an entrant. In the 1990s, new technology called Personal Communication Services (PCS) allowed much more eﬃcient use of spectrum and made mass-

5Much of the following historical material is taken from Wikle (2002).

4 market wireless service truly possible. The 1995 spectrum auction allocated PCS compatible spectrum to 107 different companies, greatly expanding the market. Over the next two decades the market greatly consolidated. Today, there are four “national” carriers, which offer service to almost all markets in the United States. These are AT&T, Verizon, SprintNextel and T-Mobile. AT&T and Verizon are the largest, each capturing about 30% of the market. They are descended from the various “Baby Bells” resulting from the 1982 breakup of the AT&T telephone monopoly, and they generally have much of the original spectrum allocated 1982. SprintNextel and T-Mobile each capture less than 10% of the market and descend from entrants in the 1995 spectrum auction. The remaining market is split between regional carriers, like U.S. Cellular and NTELOS that do not serve the entire country, and Mobile Virtual Network Operators (MVNOs), like Straight Talk and H2O Wireless, which do not own their own base stations but instead rent from the other carriers. During the time period of my data samples, a fifth national carrier was active, MetroPCS, but differed significantly in its service offerings and resembled more a small regional carrier than the other four. T-Mobile merged with MetroPCS in 2013, but its brand is still active to this day. Since MetroPCS was active in Connecticut during the sample period, I will discuss MetroPCS when data on the firm is available, but this is unfortunately only some of time since MetroPCS was always a relatively small player. Given the complexity of wireless services, what drives the difference in demand for carriers is not obvious, but there is some anecdotal evidence. In a market research survey in 2008, respondents most frequently reported “Better Coverage” (22%) above “Lower Prices” (19%) as the reason for choosing their carrier.6 However, possible reasons for choosing one’s carrier still varied widely, as this survey also showed that things like “Family/Friends Subscribe to

6See Table 1.1.

5 Figure 1.1: Consolidation in U.S. Telecom - from “Sprint Abandons Pursuit of T-Mobile, Re- places CEO,” Wall Street Journal, August 5, 2014, online edition: http:// www.wsj.com/ articles/ sprint-abandoning-pursuit-of-t-mobile-1407279448 .

6 Table 1.1: Primary Reason for Switching Carriers

Percent of Survey Respondents

Primary Reason for Choosing Carrier Oct-Nov 2006 Feb-Mar 2008

Better Coverage 27% 22% Lower Prices 14% 19% Family/Friends Subscribe to Carrier 13% 17% Plan Features 9% 12% Promotional Oﬀer 8% 9% Better Minute Level Plan 9% 7% For a Speciﬁc Phone 4% 3% Other Reason 16% 11%

Taken directly from Comscore Wireless Report, Press Release March 31,

2008. See http://www.comscore.com/Insights/Press Releases/2008/03/Price

Increasingly Important Factor in Cell Phone Carrier.

7 Carrier” (17%) and “Plan Features” (12%) are important too. I examine more closely the various dimensions of wireless service in the following sections.

1.3.2 Available Features and Services

Service today consists of two components, the handset and the service plan. The handset is the actual phone, which serves as portable transceiver of electromagnetic signals. Modern phones are also portable personal computers and thus support intensive data services such as internet access and streaming video. Service plans are incredibly diverse and numerous: the number of possible plans was estimated by consumer advice website Billshrink to be approximately 10 million.7 There are two kinds of service plans, prepaid and postpaid. Prepaid plans are paid by the minutes used, day or month (or by megabyte in data usage). They are called “prepaid” since often one buys a fixed value of minutes or data that has to be replaced once depleted. In contrast, postpaid plans are structured as a three-part tariff: there is a fixed monthly fee, but if a certain amount of minutes or data is exceeded, the “overage” results in extra charges.8 Since the bills come at the end of the usage period, the plan is “postpaid.” In the United States, postpaid plans dominate, which is generally attributed to the “phone subsidy”: postpaid plans will give a discount on a bundled handset, which the prepaid plan does not. U.S. postpaid plans generally take the form of two-year contracts, which require an early termination fee to break. The postpaid plan also requires a credit check that many low-income consumers

7billshrink.com closed down in 2013. An archived February 4, 2011 press release with this estimate can be found at http://www.billshrink.com/blog/press-releases/americans-overpay-336-a-year-on-wireless/. More recently, a July 31, 2013 article in the Wall Street Journal, “Inside the Phone-Plan Pricing Puzzle,” notes there are 750 smart phone plans from the four major carriers.

8A postpaid plan is not a two-part tariﬀ, since in addition to the lump-sum subscription price there are two diﬀerent marginal prices-below the overage limit, the marginal price is zero, and over the limit the marginal price is positive.

8 cannot pass. Recently, consumers developed a greater taste for prepaid. This has led the national carriers to move away from the traditional postpaid contract model to unlimited data plans and no contract plans.9 Table 1.2 displays characteristics of service plans available in Connecticut during the sample period of 2006-2012.10 The services are divided into two major groups - the four national carriers with postpaid plans with many features and a higher overall rating, and other carriers specializing in cheap prepaid with lower features but no credit checks. This is roughly consistent with a vertical quality model as in Shaked and Sutton (1982), where firms differentiate on quality. Interestingly, Sprint maintains two prepaid services under the brands of “Boost Mobile” and “Virgin Mobile,” perhaps to take advantage of the benefits of quality differentiation.

1.3.3 Pricing

The diverse selection of postpaid plans and the fact that they are multi-part tariﬀs makes comparing prices diﬃcult in this industry. Table 1.2 shows that price for Verizon and AT&T standard unlimited plans for voice and text and the price for the standard just voice plan

are the same ($89.99 and $69.99). Meanwhile, T-Mobile has lower prices for the same plans ($69.99 and $59.99). This would be consistent with T-Mobile positioning itself as a low- price, low-quality provider and Verizon and AT&T as high-price, high-quality providers. In contrast, Sprint does not have either of these plans and specializes in a plan with also unlimited data with at $99.99. T-Mobile also has this plan and prices at the same level.

9See Federal Communications Commission (2014).

10The information was downloaded from http://cell-phone-providers-review.toptenreviews.com/ on March 5, 2012.

9 Table 1.2: Top10Review.com - 2012 Features and Prices

Carrier Verizon AT&T Sprint T-Mobile Boost Virgin MetroPCS TracFone

Owner/Network Verizon AT&T Sprint T-Mobile Sprint Sprint MetroPCS Am´erica M´ovil/ Various Overall Rating 9.58 9.38 8.38 8.33 7.73 7.3 5.45 5 Postpaid Plans Yes Yes Yes Yes Prepaid Pans Yes Yes Yes Yes Yes Yes Yes International Plans Yes Yes Yes Yes Yes Yes Yes Family Plans Yes Yes Yes Yes Yes Data Plans Yes Yes Yes Yes Yes Business Plans Yes Yes Yes Yes

Unlimited Voice & Text ($) 89.99 89.99 69.99 45.00 Unlimited Voice ($) 69.99 69.99 59.99 35.00 Unlimited Everything ($) 99.99 99.99 50.00 55.00 40.00

Phones & Features

Free Phones Available Yes Yes Yes Yes Free Mobile-to-Mobile Minutes Yes Yes Yes Yes Yes Free Nights and Weekend Minutes Yes Yes Yes Yes Free Domestic Long Distance Yes Yes Yes Yes Yes Yes Yes Yes Walkie-Talkie (PTT) Yes Yes Yes Yes Yes Refurbished Phones Yes Yes Night Start Time 9:00 PM 9:00 PM 7:00 PM 9:00 PM Rollover Minutes Yes

Additional Beneﬁts

Call Features† Yes Yes Yes Yes Yes Yes Yes Conference Calling Yes Yes Yes Yes Yes Roadside Assistance Yes Yes Yes No Credit Checks Yes Yes Yes Yes

Fees

Roaming Charges 0.19

Activation Fee ($) 35.00 36.00 36.00 35.00 Contract Length 2 Years 2 Years 2 Years 2 Years

Early Termination Fee ($) 175.00 150.00 200.00 200.00 Days to Return Phone & Cancel Service 30 30 30 30

Help/Support

Live Chat Yes Yes Yes Yes

Covers eight brands reported by Top10Reviews.com in the top 10 that were present in Connecticut in 2012.

† Combines the following set of features which are present in all carriers but Tracfone: Voice Mail, Caller ID, Call Waiting, Call Forwarding and 3-Way Calling.

From http://cell-phone-providers-review.toptenreviews.com/, downloaded March 5, 2012.

10 Figure 1.2: Average Revenue Per User by Carrier and Plan Type

Figure 1.2 shows “average revenue per user” (ARPU) for the different national carriers, estimated from the UBS model of the industry used to calculate numbers for their U.S. Wireless 411 reports.11 This measure is standard in the industry and combines revenues from all parts of the postpaid tariff. ARPU contrasts significantly with the service plan price of the unlimited plans from Figure 1.2 - here Verizon has the lowest postpaid “price” through the period. Sprint again is not positioned as a low-price postpaid provider: by the end of the sample period it has swapped positions with T-Mobile as the second most expensive postpaid provider. So unlike the market overall, the national-carrier segment for

11Thanks to John Hodulik of UBS for provision of the wireless model.

11 postpaid again seems hard to fit into a typical vertical model of quality differentiation. The prepaid ARPUs better fit the notion of AT&T and Verizon being higher-quality, yet higher-priced firms. AT&T and Verizon prepaid ARPUs exceed the prepaid ARPUs of T- Mobile and SprintNextel for most of the period. Those measures are a bit noisy (as they are backed out from prepaid revenue and subscription data) and strangely have Verizon prepaid being more expensive than postpaid in 2008. In 2013, the T-Mobile prepaid ARPU jumps to almost equal the T-Mobile postpaid ARPU - this is due to accounting complications from its merger with MetroPCS which had higher price prepaid plans.

1.3.4 Handset Selection

Handset selection is an important dimension of quality in this industry. Starting with the introduction of the Apple iPhone, so-called smart phones have attained dominance over the older “feature phones.” Apple famously signed an exclusive agreement with AT&T, though by 2013 the current version of iPhone was available for all national carriers. To give an idea of how handset quality has evolved over the sample period, I present data from PhoneArena.com. Sinkinson (2014) used this website for product characteristics of smart phones. Here I have collected all the data for all phones listed on the website. Con- veniently, PhoneArena maintains pages on all discontinued phones with introduction dates, so I am able construct a panel of handsets and their characteristics over time. Assuming the life of a handset model is 3 years, I take the mean of these characteristics over a year for the four national carriers and MetroPCS, and I report them in Figure 1.3.12 A sales-weighted mean would be more informative, but given I have no sales data, these means should still be informative about overall trends.

12There is no “discontinuation” date reported unfortunately, and technically a phone might be used in- deﬁnitely after purchase. So I need to assume some length of time after which use of the model becomes essentially negligible.

12 Figure 1.3: Average Handset Characteristics over the Sample Period

Overall, innovation in the quality of phones is remarkably similar, as storage capacity, camera pixels, display pixels, processor speed, and the percent of phones with touchscreens and accelerometers steadily increase in all carriers. This pattern is also true of talk time except for AT&T, which decreases starting in 2007 until it sharply increases after 2010. The weight of the phones slightly decreases until about 2009, reflecting greater miniaturization of feature phones, but afterward firms start switching heavily into heavier smart phones so weight greatly increases. There does not seem to be one firm that dominates a large number of characteristics, and the leaders in each of these characteristics rarely change. Available varieties are more or less stable over time for each carrier, though this is one characteristic where each carrier seems to be differentiated. AT&T always has the most

13 varieties; Sprint begins with second most but is supplanted by Verizon in 2009; and T- Mobile always has the fourth most. MetroPCS tends to be a laggard in all dimensions, including phone varieties. Thus in general, the typical handset available from any national carrier is going to be very similar, and the only difference is that carriers will stock different numbers of varieties. Thus if handset quality is important for determining market shares, it must be through only the characteristics of certain blockbuster handsets, i.e. the iPhone. This may have been true before, when the iPhone was exclusive to AT&T; however, examination of current market conditions suggest a much more level playing field since the iPhone and other popular phones are available from most of the national carriers. Table 1.3 is recent information from Top10Reviews.com about the quality of the network, phone selection and customer service on a grade scale.13 While there are differences amongst the four nationals in the “budget” phone selection, the “flagship” phone selections grades are all A’s and the “midtier” phones selections are all A- and B+. This implies the handset selections of the nationals are quite similar. Boost-Mobile and MetroPCS both have largely inferior selections, but of course these services are now simply prepaid brands of the four nationals. If these qualities are indicative of tastes for non-national brands, then there could be differentiation between national and non-national plans in phone quality.

1.3.5 Signal Quality

1.3.6 Base Stations and Sites

When a wireless call is made on a carrier’s network, the handset sends information to the nearest antenna that services that carrier over that carrier’s frequency band of the elec-

13This information from Top10Reviews was not available for my sample period for the later analysis, which is why I do not report 2012 versions of this information in Table 1.2.

14 Table 1.3: Top10Review.com - 2015 Reviews of Quality

Verizon T-Mobile Sprint AT&T Boost MetroPCS Mobile

Overall Rating 8 7.85 7.75 7.6 6.2 5.98 Network Provider Verizon T-Mobile Sprint AT&T Sprint T-Mobile Urban Coverage A+ C+ B A B C+ Rural Coverage A F C- B+ C- F Speed A A- D A- D A- Reliability A+ C+ B+ A B+ C+

Device Selection

Flagship Phones A A- A A+ D+ D- Midtier Phones B+ B+ A- B+ D+ F Budget Phones B D B B A D

Help & Support

Customer Service Score B- B+ D- C B B+

Covers brands reported by Top10Reviews.com in top 10 that were present in Connecticut in 2015.

Carriers like Tracfone no longer in top 10 are no longer reported.

From http://cell-phone-providers-review.toptenreviews.com/, downloaded April 9, 2015.

15 tromagnetic spectrum. These antennas are part of the carrier’s base stations, equipment facilities that reroutes the information through the landline telephone system. If the receiver of the call is also on a cell phone, the call will leave the landline network and be rerouted to the nearest base station to the receiver, and the base station will beam the call information to the target. Thus signal quality depends crucially on the ability of the base stations to form and maintain transmissions. The power of the transmission decreases with distance, so if no carrier base station is in range, then the signal power between the phone and the base station will be too weak to start a call. Even when a consumer is close enough to a base station to initiate a call, there can still be problems since random ambient interference might overwhelm the signal and disrupt it. This disruption ends the transmission of information, creating a “dropped call”. Accordingly, carriers are interested in building base stations to make sure their market areas are well covered and dropped calls are kept to a minimum. The more base stations in an area, the more likely a consumer will be in range and the less likely a call will be dropped. However, base stations are very costly. Aside from the costs of equipment, maintenance and power, base stations must be mounted on elevated structures. Therefore, a large tower must be built or space on a preexisting tall structure must be rented. Developing and acquiring these locations, or “sites,” requires signiﬁcant regulatory proceedings. In most U.S. states, the local town zoning board must clear the site. Towns are typically very concerned about potential radiation exposure and the impact of these sites on town views and real estate prices. Thus town objections can be very strong to proposed site developments, and negotiations between site developers and towns can last years. Such delays became so long that the FCC decreed a maximum delay time for responses to carrier inquiries about site development. Town objections to that order resulted in a 2012 Supreme Court Case: City of

16 Arlington, Texas v. FCC.14 In addition, many base stations end up camouflaged to assuage town aesthetic concerns. It is not uncommon to hide a base station in a church spire or to disguise a base station as a tree. Given these costs, it may not be surprising that different U.S. firms tend to collocate their base stations on the same site. Often a single tower or rooftop will accommodate multiple base stations from different firms. The right to collocate is obtained via a monthly payment to the owner of the sites. For the large industry players, the rates are set via national-level contracts. These contracts are secret for private firms, but are reputed to be around $1,000 to $3,000 a month. Public bodies also maintain sites for internal use in their emergency communications systems and will lease extra space to private firms. For example, a public schedule for site leasing fees is available from the Connecticut Department of Transportation.15 Their monthly fees start at $2,000 a month and increase each year until the twenty-fifth year, to $4,231. This is about a 3% average annual increase. Ownership of sites is not concentrated - while most of the sites were developed by the carriers in previous decades, most old sites have been sold to third-party tower management companies like American Tower, Crown Castle, and SBA. Most of the new sites are being built by the third party firms as well, resulting in a relatively low share of vertically integrated sites. The combination of high sharing and low vertical integration is unique to the U.S. context. In other nations, such as Canada or Brazil, laws had to be passed to mandate collocation as firms fiercely held onto both exclusivity and ownership of their sites.

14Another Supreme Court Case, T-Mobile South, LLC v. Roswell, GA, also involved a dispute about cell siting and town approval. T-Mobile South, LLC, a T-Mobile subsidiary, objected to Roswell’s denial of their application for a site since they believed the denial failed to include the FCC-mandated written reason for the denial.

15See http://www.ct.gov/dot/cwp/view.asp?a=1605&q=272692.

17 Figure 1.4: UBS Estimated Total Sites by Firm

Figure 1.4 shows estimates of the national owned site totals by quarter and carrier. These appear in the UBS U.S. Wireless 411 reports.16 There are some rather sudden jumps in site totals - Verizon jumps by over one thousand base stations from Q4 2008 to Q1 2009; T-Mobile jumps by over 2,700 from Q4 2012 to Q1 2013; and SprintNextel declines by over 28,000 sites from Q1 2012 to Q2 2013, before rebounding by over 17,000 to Q3 2013. These are explained by the estimates being based on primarily company self-reporting and so changes in a carrier’s counting methodology are not adjusted for, especially when a ﬁrm merges and gains new sites through the acquisition.

16These were compiled from UBS ﬁrst quarter Wireless 411 reports from 2009-2014, #32, 36, 40, 44, 48, 52. These do not appear in the UBS wireless model where I have prepaid and postpaid ARPUs.

18 So while these may not be the best measures of total base station numbers, there are several conclusions to draw. First, base stations are generally increasing over time to meet growing demand. Second, for most of the period, the ordering of site totals is SprintNextel, AT&T, T-Mobile and then Verizon. This does not reﬂect the national market share at all, so if signal quality in very important to demand, there must be more than simply the site totals that matter to consumers.

1.3.7 Transmission Technology, Spectrum, and Network Manage-

ment

Signal quality depends on the technology and spectrum available to different firms. In the United States, different firms use different technologies to encode their signals. AT&T and T-Mobile use variants of the GSM standard, in which each call is apportioned a different part of the carriers’ spectrum in that area. CDMA, used by Verizon and Sprint, interweave calls from all users over the carrier’s entire local spectrum. Theoretically, a CDMA signal will travel farther than a GSM signal, so a CDMA carrier might need less base station density to yield more quality. In addition, spectrum holdings are a signal quality concern in two dimensions. First, spectrum represents the amount of information that a base station can support in an area at any one time. A call can be dropped or switched to another base station if spectrum becomes so full a carrier with more spectrum may have less dropped calls. However, this concern seems minimal as industry sources I have spoken with characterize dropped calls due to capacity constraints as only 5% of all dropped calls, and dropped calls are themselves around only 1-2% of calls in general. Capacity is more of an issue when dealing with data, in which firms slow down data transfer to deal with congestion. For the purposes of this analysis, I will abstract from capacity concerns and assume firms have invested appropriately

19 Figure 1.5: From page 53 of 17th Annual Wireless Competition Report (2014).

in upgrading their base stations to mitigate capacity issues over my sample period. This approach is in line with news reports that characterize a spectrum shortage as a looming crisis, but note that the U.S. had “slight spectrum surplus” as of 2012.17 Second, and potentially more important, different parts of the electromagnetic spectrum have different properties. Frequencies under 1000 MHz propagate farther and therefore are more useful in rural areas. AT&T and Verizon have almost all this spectrum, since this was the first spectrum apportioned to firms. Other current carriers like Sprint and T-Mobile are descendants of entrants from the mid to late 1990s when most of the low frequency spectrum had already been distributed. Thus, Sprint and T-Mobile might yield less quality from base stations than their rivals. Figure 1.5 shows the spectrum distribution weighted by population of the license areas, as calculated by the FCC in their 17th Annual Wireless

17 See “Sorry, America: Your wireless airwaves are full,” http://money.cnn.com/2012/02/21/technology/ spectrum crunch/, February 21, 2012.

20 Competition Report.18 AT&T and Verizon have most of the sub-1000 MHz licenses classified as “Cellular” and “700 MHz,” and Sprint has the most frequencies. A final point is that the network managers are able to alter the actual shape of the cells to a certain extent. That is, when a call is made in an area in range of multiple base stations, the network can choose which base station serves the call. Through careful management, the network can dynamically adjust loads on base stations by shrinking or expanding the associated cells. The ability to do this is based both on software and on other technology used by the firm, and also the ability of the managers of the network. Verizon in particular is known for its ability to manage its network efficiently relative to its rivals. Table 1.3 includes grades for network quality, which Top10Reviews calculated from measures provided by website RootMetrics. RootMetrics generates signal quality ratings via independent testing of signals across markets.19 AT&T and Verizon have the best grades in signal quality nationally, with T-Mobile and Sprint having substantially worse grades. This is even though Sprint has a far higher absolute number of MHz relatively. Rural coverage is worse for each network, which is natural since more sparsely populated areas need more base stations to be similarly effective. The difference is substantially worse for T-Mobile and Sprint, which is consistent with the idea that AT&T and Verizon can more effectively serve sparsely populated areas with their low frequency spectrum. Notably, Verizon and Sprint, the CDMA carriers, are not strictly better than AT&T and T-Mobile, the GSM carriers, suggesting transmission technology by itself is not pivotal in signal quality determination. Overall, the national network quality measures match up with overall shares - Verizon,

18The ﬁgure is in fact directly taken from the report, Federal Communications Commission (2014).

19Exactly how Top10Reviews calculates these grades is not explained - though maps with rating information can be retrieved from RootMetrics’ website. That data, however, cannot be scraped and is too large for manual transcription, so I do not conduct a direct analysis of that data. After cursory examination, these maps do seem consistent with the Top10Reviews grades.

21 AT&T, Sprint and T-Mobile. There is, however, not a simple relationship between total spectrum holdings, transmission standard, and quality - the spectrum mix is important and the geographic distribution of sites may be important as well. In future sections I will examine geographic variation more closely using detailed data from Connecticut.

1.4 Connecticut Data

1.4.1 Base Station Data

The previous sections imply signal quality, and thus base stations, to be a potentially important component of product quality in the wireless industry. A further analysis requires locations of base stations, but this unfortunately is not available for most of the United States. The FCC does not collect comprehensive base station information.20 Such data is, however, available for Connecticut due to its unique laws and regulatory environment. In most U.S. states, getting a site cleared for development takes the form of an application that must be cleared by a town zoning commission. In addition to town zoning commissions, Connecticut has a statewide regulator of telecom facilities, the Connecticut Siting Council. Applicants to develop base stations still need to consult with towns, but if the application is to build a tower or to collocate on a preexisting site, then the official application is submitted to CSC, and they have final say with regards to clearance. Since 2006, Connecticut state law requires that the CSC maintains two datasets on base stations meant to be as comprehensive as possible and is therefore the best source of this kind of data in the US. The first records information for all proceedings between the CSC and site applicants. Thus for every tower and for every other site with more than one carrier, I have information

20The FCC has two databases. First, there is antenna data that is limited to only enough antennas to create license boundary maps. Second, there is site data that is mandatory only for installations over 200 feet tall and infrequently updated.

22 on when a base station was cleared for installation, geographic location, its owner and miscellaneous technical information like the site type and sometime comments about the type of equipment installed. The second dataset is taken from the towns that not only report the information in the first dataset, but also for sites on preexisting structures and only one carrier. A new rooftop site faces town regulation directly without involvement from the CSC at all. Thus the CSC only knows about the site from mandated reports from the towns themselves. This data is far less complete than the CSC original data, and generally only has the location and base station owners. This dataset uniquely reports about half the number of sites in the data, so I merge both datasets and use only the ownership and location variables, which are consistently reported across both. Further, the second dataset is continuously deleted and replaced with an update on a monthly basis, so older copies had to be retrieved using the Internet Archive.21 Archiving of websites is not done with perfect regularity, so the dates of the available website copies vary from year to year. Due to telecom sites being first recorded by the state regulator before construction and operation, I define the count of base stations for a year as the count of all base stations reported before January 1st of that year.

1.4.2 Zoning Codes

Since regulation is so important for determining base station location, I collected zoning codes for almost every town in Connecticut. When deciding to clear an application for a site or a base station, zoning boards take into account a variety of concerns, including aesthetics, noise and impacts on housing values. Many of their concerns and requirements are listed in so-called zoning codes, which are particular for each town in Connecticut. Zoning codes

21The Internet Archive (www.archive.org) is a website that archives other websites. By using the site’s “Wayback Machine” function, one can access old versions of websites that they have stored oﬄine.

23 thus explain the procedure for applications and what results in a successful application, and thus are an important summary of the stringency of the regulation. The local zoning codes in Connecticut matter less than other states since the final decision-making process for many of the sites lies with the CSC. The CSC may ignore the zoning codes of the town in question when making their decision. However, state law further requires applicants must send the application and consult with the town they plan to build in at least 60 days in advance of the submission to the CSC.22 The town then may send objections to the CSC, and the CSC takes the interests of the town into account in their final decision. Often multiple sites are offered as possibilities, and the CSC might choose the one least preferred by the carrier or insist on modifications to the site design to satisfy the town. In practice, the applicant and the town may negotiate for much longer than sixty days before submitting an application to the CSC so the town does not object or lobby for more conditions to the CSC when the application is finally made. For example, the tower company SBA informed the town of Bloomfield that it wanted to build a site on February 6, 2006. The town suggested a different site, causing further delays, until eventually SBA applied to the CSC over three years later on March 16, 2009. There are other times when the application is made, and there is no objection from the town at all and no negotiation. Thus regulatory variation at the town-level is still an important component in the siting process, even under the CSC-led system. Differences in zoning statutes across towns should still have a strong effect. Each zoning code was either collected from the town website or acquired from the town’s town clerk. In Connecticut zoning codes, there is usually a section called “Wireless Telecom- munications Facilities” that regulates cell sites. This details the goals of the zoning regu-

22Towns 2,500 feet away must also be informed - this is an important caveat of the analysis in Chapter 3. I discuss this point in more detail there.

24 lations from the town’s perspective, the requirements of the application process, what will be required for the application to be accepted, and restrictions on the design and operation of the site. Sometimes this section is absent and there are no special regulations for sites, and sometimes there are shorter regulations incorporated into other sections of the zoning code.23 I find something like this is present in 139 out of the 169 towns in Connecticut. One town, Cornwall, has its section missing from the dataset since that section is reported in an addendum that was not included with the main document. For now, I consider Cornwall as missing and will not use it for regulatory analysis. Of the sections I do have, the mean length is 13,611 characters and 2,487 words. For that same subsample, a section’s mean share of total zoning code is 4.84% in characters and 4.75% in words, so these sections represent a significant portion of the zoning code and presumably a significant concern for towns in general. I will leave further discussion of the zoning codes implications for Chapter 2 and 3, where this data is used extensively.

1.4.3 Proprietary Demand Data

To complement the base station data, I acquired consumer survey data from the Nielsen Company and Scarborough Research. The Nielsen data has been used in Sinkinson (2014) to study the wireless phone market, while the Scarborough data has been used by Gentzkow (2007) to study internet and print newspapers and Waldfogel (2008) to study the interaction between restaurant types and demographics. Both ﬁrms sample from the national U.S. population, collecting demographic, plan and carrier information, with total samples being

23In my application, I use all such passages, even those not separately broken out as a site regulation section. When an obvious section was not present, I searched for key words like “wireless,” “telecom,” and “sites.” I then manually inspected the results for potentially related regulations. I will abuse terminology for the rest of the dissertation by describing these also as “sections.”

25 in the hundreds of thousands. Nielsen is a somewhat more inclusive sample as it includes teenagers, who could be an important part of phone service demand. Nielsen is a web-based survey while Scarborough is a two stage phone-then-mail survey. For Scarborough, basic demographic questions are asked on the phone, but more detailed consumption questions are asked on a follow-up mail survey. For my purposes, this results in significant imputation problems in the Scarborough data, as the response rate for questions about carrier choice and plan type is not quite 50%.24 Nielsen does not suffer this issue since the web-based nature of the surveying prevents non-responses about these questions - a respondent cannot finish the survey if every question is not answered.25 In addition, one can see in Table 1.4, the Scarborough data oversamples women and the elderly much more than Nielsen does relative to the 2008-2012 American Community Survey estimates. This is consistent with Scarborough phone and mail methodology oversampling individuals who stay at home most of the day, i.e. housewives and retirees. In contrast, Nielsen oversamples families which are 81.5% of the sample rather than 72.4% in the ACS, possibly due to the way it recruits people into its surveys. However, the Scarborough data has a longer panel (2006-2012 rather than 2008-2012 for Nielsen) and more information for smaller carriers. In particular, it breaks out market shares for MetroPCS and the prepaid plans of Sprint. The Scarborough imputation is a nearest neighbor algorithm, which Chen and Shao (2000) and Chen and Shao (2010) have

24Gentzkow (2007) does not suﬀer from such high imputation rates since the questions about newspapers, print and web, are asked during the phone interview. He still reports a 6% imputation from non-response on the phone, though.

25To better represent certain populations that might ﬁnd the web survey hard to use, for example, non- English speakers, Nielsen samples some demographic groups over the phone. Since the survey is then entirely verbal, the survey-taker tries to obtain answers to as many questions as possible but often is unable to collect them all. This is a relatively small proportion (0.5%) of the Nielsen sample. I try to include these individuals as in as many calculations as possible; though, I remove them when the non-response is in the variable of interest. In addition, Nielsen respondents have the option of not reporting their income, which a signiﬁcant fraction (13%) chooses to do. This is treated as a separate income group in this dissertation.

26 Table 1.4: 2008-2012 Unweighted Responses by Demographics Compared to 2008-2012 American Community Survey for Connecticut

Nielsen Scar. ACS

Household Income† Less than $35k 21.8 22.0 25.6 $35k-50k 12.2 15.9 11.0 $50k-75k 22.0 15.6 16.7 $75k-100k 17.2 15.5 13.4 $100k+ 26.9 31.0 33.3

Household Size Single 18.5 24.6 27.6 Family 81.5 75.4 72.4

Age†† 18-34 Years 18.6 12.7 27.2 35-64 Years 65.1 56.5 54.3 65+ Years 16.3 30.8 18.5

Sex Male 43.4 38.9 48.7 Female 56.6 61.1 51.3

† Excludes individuals who did not report income, which is 13% of the Nielsen sample. †† Excludes teenagers, which are 6.3% of the Nielsen sample.

27 shown has good properties for generating aggregate summary statistics.26 So I will invoke the Scarborough data along with the Nielsen for aggregate statistics for Connecticut as a whole, but use Nielsen only to look at market level variation.

1.5 State-Wide Demand Analysis

To gain a rough idea of the nature of demand in this industry, I perform a qualitative analysis of the demand at the state level. I construct market shares for the whole state using the survey responses and weights and aggregate to the state level. I then break the data into different carriers or plan types and compare theses shares over years or demographics. Table 1.5 shows the market shares for the survey respondents of the two datasets.27 The Nielsen data has a shortcoming that only the four major carriers are identified, so all other carriers have to be aggregated in an “Other” category. This is a problem in that prepaid brands Virgin and Boost are not distinguished in the data. While this will not pose an issue in this chapter, both Virgin and Boost are owned by Sprint and use its network, so the calculations in Chapter 2 will be somewhat misspecified in the sense that Sprint will not have all of its customers included when calculating its profit. This discrepancy may not be so bad since the Scarborough dataset has approximately the same market share for SprintNextel also including Virgin and Boost - 11.1% in Nielsen versus 10.9% in Scarborough. Verizon is the market leader, followed closely by AT&T. Sprint and T-Mobile are distant also-rans, with less combined market share than AT&T alone. The aggregation of all other plans is

26A nearest neighbor algorithm replaces missing observations by randomly sampling known observations, with higher probability to those with more similar observables. The exact procedure to conduct imputation is not available from Scarborough, so I cannot back out the original responses.

27Here I use weights provided by the two surveys to re-balance the data to resemble the true population more closely. The Scarborough data is balanced on a state level, while the Nielsen is done on the national level. In practice, the results are similar to the unweighted versions.

28 Table 1.5: 2008-2012 Carrier Shares

Carrier Plan Type Nielsen Scar.

AT&T Postpaid 28.5 25.1 Prepaid 3.7 2.2 SprintNextel Postpaid 11.1 8.0 Prepaid - 2.9 T-Mobile Postpaid 6.6 4.6 Prepaid 2.1 1.3 Verizon Postpaid 31.1 31.5 Prepaid 2.0 1.5 Other Postpaid 0.7 2.7 Prepaid 7.8 5.0 None None 6.3 15.3

29 Figure 1.6: Estimated Market Shares over the Sample Period

8.5% in Nielsen and 7.7% in Scarborough. These other plans includes MetroPCS, which owns its base stations, and Mobile Virtual Network Operators StraightTalk and Tracfone, which license use of the network of other ﬁrms. Finally, postpaid plans dominate, with only 15.6% of respondents having prepaid plans in Nielsen and 12.8% in Scarborough. Penetration is very high in Nielsen, with only 6.3 percent without cell phone plans, but the Scarborough data has a very high 15.3% of individuals without plans even with weights. This is likely due to the imputation issues discussed earlier. The two demand datasets paint broadly the same picture in a dynamic context. As seen in Figure 1.6, AT&T and Verizon are the market leaders, followed by SprintNextel and T-Mobile and other smaller carriers. The main change over time is that Prepaid and the

30 Figure 1.7: Estimated Carrier Prepaid Shares over the Sample Period

Other carriers are growing more popular over the sample period. Looking at only prepaid plans in Figure 1.7 one ﬁnds that most of the prepaid growth appears to be due to increasing popularity of non-National prepaid plans. By demographics, there are consistencies in tastes for particular carriers and plan types. To focus on the demographic dimension, I construct full sample shares for all the years over the sample. I present the shares over income in Figure 1.8, in which each point is a full sample share displayed at the midpoint of the income bracket.28 Other, Prepaid, T-Mobile

28For the last bracket, which has no right endpoint, I display the datapoints at $175k for Scarborough, and $150k for Nielsen. This is because Scarborough has a higher left endpoint for this last bracket ($150k) versus Nielsen ($100k) and I take the left endpoint of these brackets as $200k for both for calculating a pseudo-midpoint.

31 Figure 1.8: Estimated Carrier Shares by Income Brackets

and no plan are more popular amongst the poor, while AT&T and Verizon are strongly preferred as income increases. I perform a similar exercise in Table 1.6 where I report full sample shares for diﬀerent age brackets. The results are similar, except for the share of “None” which is a whopping one third of elderly respondents in the Scarborough data and very likely the result of imputation bias. Though not as strongly, even in the Nielsen data those above 65 years of age are relatively less enthusiastic about phone service overall and more likely to be interested in prepaid and non-National plans. More detail on age can be seen in the Nielsen data, which actually reports age in years. Figure 1.9 plots shares over all ﬁve year brackets of age in the

32 Table 1.6: 2008-2011 Carrier Shares by Age Group

18-34 Years 35-64 Years 65+ Years Carrier Nielsen Scar. Nielsen Scar. Nielsen Scar.

AT&T 30.7 27.9 33.3 28.6 32.9 22.1 Verizon 35.1 33.0 34.5 36.1 30.0 23.5 SprintNextel 15.1 15.2 11.2 10.6 7.5 5.5 T-Mobile 10.3 7.7 8.4 5.5 8.5 4.2 Other 5.6 6.5 8.6 7.1 13.4 11.2 None 3.3 9.7 4.0 12.1 7.6 33.5 Prepaid 11.3 12.6 15.2 11.8 21.9 16.3

Nielsen sample.29 I use 5-year brackets rather than actual year age since year age results are very noisy. Levels resemble the aggregate market shares, but there are very nonlinear differences across age groups. Demand for T-Mobile is stable over the entire age distribution. AT&T and Verizon shares take a sharp drop for those over 70 years old, seemingly replaced by non-National carriers. SprintNextel shares have a sharp peak for those in their thirties. No plan and prepaid plans are preferred amongst teenagers and the elderly, creating roughly U-shape distributions. This is likely because the income saved from a having a prepaid or no plan is especially important to teenagers who have to rely on their parents for income and the elderly who are drawing down their retirement savings. There are very little differences over gender, as seen in Table 1.7. However, while the most important results are consistent across datasets, there are substantial differences in levels. Primarily, the Scarborough data seems to significantly overstate

29A few exceptions: Teenagers below 17 years of age do not have reported ages so are displayed at 15 years. Seventeen to nineteen years olds are displayed at 18 years. The highest bracket, 75+, is displayed at 77.5 years, as if it had a right endpoint of 80 years.

33 Figure 1.9: Nielsen Estimated Sample Market Shares by Five-Year Age Brackets

the share of no plan and to signiﬁcantly understate the share of prepaid and Other plans. This is consistent with Scarborough oversampling people who have landline telephones and thus have less taste for unusual plans types, unpopular brands, and mobile phones use in general. Instead of replacing missing values with comparable people, the biased replacement sample imputes the missing values to be too much like the oversampled sub-populations.30 Thus the levels of shares are likely distorted in all previously mentioned cross sections of demographics, though the general trends across demographics are mostly preserved. This is

30This could be remedied by a good distance measure to sample very similar individuals more strongly, but it appears that the observables-based distances metric used to construct the sampling weight for the Scarborough data is not accurate enough.

34 Table 1.7: 2008-2011 Carrier Shares by Gender

Male Female Carrier Nielsen Scar. Nielsen Scar.

AT&T 34.3 26.8 30.2 27.6 Verizon 31.7 33.6 34.4 32.5 SprintNextel 10.9 11.3 11.3 10.6 T-Mobile 7.9 5.9 9.6 5.8 Other 8.7 7.0 8.4 8.2 None 6.5 15.4 6.1 15.3 Prepaid 17.0 12.3 14.3 13.3

Table 1.8: 2008-2011 Carrier Shares by Family or Alone

Family Alone Carrier Nielsen Scar. Nielsen Scar.

AT&T 32.3 27.8 31.1 23.6 Verizon 33.7 34.7 28.7 22.0 SprintNextel 11.4 11.6 9.0 6.6 T-Mobile 8.7 6.0 9.4 5.2 Other 7.9 7.2 13.3 10.7 None 6.0 12.8 8.6 31.8 Prepaid 14.9 12.4 21.4 15.3

35 even true for the cross sectional difference across single and multi-person households, seen in Table 1.8. Here, the qualitative share distributions look quite different, though if one looks only at the inside shares the results are actually quite similar. More important is the effect on the dynamic character of the data. As seen in Figure 1.6, the share of no plan seems to be sharply decreasing in the Scarborough data where it seems to be flat in the Nielsen data. Since base stations are increasing over time this could lead to spurious findings of the importance of quality, and the scope for the benefits of market-expansion. So for the rest of the paper, I will focus the Nielsen data as it is less problematic. However, an important contribution of the Scarborough data is its breakout of the MetroPCS market share. An entrant to the Connecticut market along with fellow small carrier, Pocket, in 2009, it will invest a substantial amount in base stations. MetroPCS greatly increased its base stations at the end of 2010 as Pocket sold its base stations to MetroPCS after it decided to leave the market. Thus one might expect sizable effects on market shares from the two entries and the merger. While this narrative would be consistent with the Nielsen data, as the market share of the national carriers decreases slightly over the sample from 2008-2012 and the Other category, which subsumes MetroPCS, grows at the same time, the Scarborough data shows that MetroPCS market share is very small with 0.04% in 2009, 0.09% in 2010, 0.59% in 2011 and 0.83% in 2012.31 While the share is likely underestimated because of the imputation issues, with such small numbers it is hard imagine that it could be more than a few percent. In general, it suggests that even a firm with the combined quality-producing capital of two entrants has a difficult time penetrating the market. Thus the barriers to entry in this industry are

31Given the small market shares, I do not report MetroPCS in the other sub-samples.

36 Table 1.9: Site Statistics over Time

Year Total Sites Site Growth(%) Base Stations/Site Towers (%) Carriers (%) Tower Co (%)

2006 1,407 - 1.64 49.5 31.4 21.3 2007 1,447 2.8 1.68 49.3 29.5 21.8 2008 1,503 3.9 1.72 49.6 29.8 23.2 2009 1,513 0.7 1.82 49.7 29.1 24.5 2010 1,548 2.3 1.89 49.8 27.7 27.1 2011 1,679 8.5 1.84 47.8 28.2 26.8 2012 1,693 0.8 1.84 47.8 27.9 26.8 2013 1,710 1.0 1.87 48.2 17.2 41.0

Base stations include only wireless telecommunications equipment.

Tower companies include Crown, SBA, American Tower, and some smaller companies. likely to be large, and the required amount of capital like base stations and spectrum to be competitive is prohibitive. This suggests something like the endogenous sunk cost theory of Sutton (1991, 2001); in which incumbent ﬁrms produce a socially excessive amount of quality or advertising to discourage entry. Unfortunately, given the small estimated market shares of MetroPCS (which are based on less than a hundred out of over 41,000 Scarborough survey responses), it is hard to come to a more deﬁnite conclusion. In general, it seems safe to conclude that the prepaid and non-national plans are gradually gaining popularity over the sample period. In addition, preference for brands seems to vary with income and age, but in highly nonlinear and idiosyncratic fashions. This creates complicated relationships between demographics and preferences for signal quality.

37 1.6 Sites and Base Stations

Table 1.9 gives basic information about sites over time. Sites have grown from 1,407 to 1,710 during 2006 to 2013. There does not seem to be a consistent year to year growth rate as it varies from 0.7% to 8.5%. Collocation of sites increased from 1.63 base stations per site to 1.87, which is consistent with increasing scarcity of sites. The percent of towers remains relatively stable between 47.8% and 49.5%, which suggests that this scarcity is not due to inability to build towers. The percent of vertically integrated sites remained stable between 27.7% and 31.4% from 2006 to 2012, but drops in 2013 to 17.2% because Sprint and T-Mobile sold oﬀ much of their portfolio to tower company Crown Castle. In contrast, the number of locations owned by professional tower companies was actually lower at that point, between 21.3% and 26.8%. It is only with the large sales thereafter does the percentage of tower companies increase to 41.0%. To breakdown ownership in greater detail, Figure 1.10 displays the shares of ownership by major owners. This includes the four national carriers, but not Pocket and MetroPCS since they did not own sites; the three major tower companies, Crown Castle, SBA and American Tower Corporation; small tower companies, called “Small” in the ﬁgure; towns and the state of Connecticut, which build sites for their own emergency services and lease collocations for extra revenue; and power companies Connecticut Light and Power and United Illuminating, who can easily mount base stations on their electricity transmission towers and have their own sites for metering purposes.32 The white space at the top represents the remainder, hundreds of small rooftop sites owners, which includes churches, schools, water companies and any other private real estate owner that has decided to lease space for a site. This also includes Alltel, which was active in this market to a very small degree up until 2010 but

32The “small” tower companies include national company TowerCo, and regional companies Message Center Management, North Atlantic, New England Site Management, Pinnacle Towers, Optasite, Ten Thirty, Wireless Edge and Wireless Solutions.

38 Figure 1.10: Site Ownership Fraction over Time

was too small to break out as its own share on this graph. When much of Alltel was sold to Verizon in 2010, its base stations were transferred to AT&T in the antitrust settlement which permitted that transaction to go forward. Alltel was eventually completely bought by AT&T in 2013. The most notable feature is the large increase in the share of Crown Castle from sales with Sprint and T-Mobile. Other than that, shares are relatively stable. Interestingly, Sprint owned a large number of towers relative to its market share due to assets it acquired from the Nextel merger. One can see over time that it divested itself of these properties. The other noticeable point is that Verizon both has a relatively small number of owned towers relative

39 Figure 1.11: Base Stations over Time

to its rivals, but that this amount grew signiﬁcantly during the sample. Since carriers rarely buy ownership of already existing towers, this means that Verizon was expanding disproportionately by building towers rather than collocating. A look at the data directly shows this is mostly because Verizon disproportionately expanded in the rural northwest corner of the state, where there were not many other sites to collocate on anyway. Figure 1.11 shows total increases in base stations by each carrier over time. Verizon has the least number of base stations, even though it has both the most market share on the national and local levels, and a better reputation for quality. Verizon and T-Mobile grow signiﬁcantly in the number of base stations over time, such that they have both overcome

40 Figure 1.12: 2000 PUMAs - Subdivided into Block Groups Colored by 2010 Population Density

9 10 1 4 8 3 5 7 6 19 11 14 15 13 21 16 12 22 20 17 18 24 23 25 23

Sprint by 2012 in base station counts. AT&T and Sprint grow far more slowly in base stations, which might be explained by gains of many base stations from mergers of rough equals from earlier in the decade - AT&T with Cingular in 2004, and Sprint with Nextel in 2006. The other outstanding feature is the number of MetroPCS base stations by the end of the period, which is really the sum of two entrants into the market as Pocket divested its base stations to MetroPCS in 2010. MetroPCS has about half or almost half of the totals of its rival, but according to the Scarborough data, MetroPCS has a miniscule market share. This shows that entry into this market is hard and that simply having adequate coverage in some areas is not suﬃcient for success in this industry. The many other components of quality mentioned in Section 1.3, spectrum, transmission technology, network management, phone selection, price, etc. - are all potentially important as well.

41 Table 1.10: PUMA Market Shares Summary Statistics over All Years

Carrier Mean SD Min 25pct Median 75pct Max

AT&T 30.2 7.3 21.9 25.0 29.5 34.9 39.0 Verizon 32.1 10.5 17.4 23.4 34.0 40.3 44.3 Sprint 8.4 3.9 4.0 6.0 7.5 10.0 13.7 T-Mobile 7.8 4.7 2.5 4.3 6.2 10.9 14.3 Other 9.6 5.3 3.5 6.1 8.5 12.0 17.5 None 12.1 5.3 5.7 8.3 11.1 15.9 20.0

To facilitate analysis of geographic variation, I will separate the Connecticut into distinct geographic markets. I deﬁne a market as the PUMA, the smallest level of geography in the Public Use Microdata Sample (PUMS). Each of the 25 Connecticut PUMAs has at least 100,000 people in it so that the identities of sample respondents are protected. According to U.S. Census documentation the PUMAs are designed to represent existing communities whenever possible with similar characteristics.33 I therefore use the PUMAs to approximate travel patterns. The 2010-2011 Regional Household Survey records detailed information about travel behaviors in the New York commuting area, which includes Fairﬁeld and New Haven counties in Connecticut. The data show that 53.6% of trips taken by Connecticut respondents are intra-PUMA, so the PUMA represents a fair approximation of travel patterns. Due to PUMA 23 bisecting PUMA 25, I merged 23 and 25 to maintain contiguity in markets. Table 1.10 shows, consistent with national level shares, AT&T and Verizon have the most market share on average, while Sprint and T-Mobile have far less. However, there is quite a

33See “A Compass for Understanding and Using American Community Survey Data,” February 2009.

42 Table 1.11: Count and Density of Base Stations by Market

Carrier Type Mean SD Min 25pct Median 75pct Max

Count AT&T 32.7 15.1 11 18 35 41 69 Sprint 26.7 12.0 12 19 23.5 30 67 T-Mobile 25.5 12.3 10 19 25.5 32 79 Verizon 25.3 12.8 5 17 22 34 58 All 27.9 13.4 5 18 24 36 79

Per 1000 km2 AT&T 1.22 1.06 0.17 0.67 0.87 1.40 5.11 Sprint 1.16 1.18 0.20 0.48 0.71 1.24 4.52 T-Mobile 1.20 1.22 0.11 0.42 0.74 1.47 5.79 Verizon 0.97 0.96 0.20 0.42 0.62 0.95 4.67 All 1.14 1.11 0.10 0.46 0.75 1.32 5.79 bit of variation as, for example, Verizon has as few as 17.4% for a particular market while T-Mobile has a much as 14.3% in another. Thus it appears there are significant geographic shifters of quality. Examination of market level variation seen in Table 1.11 shows that AT&T has on average the most base stations per PUMA, and Verizon has the least. Verizon is the market leader in the data and in the nation as a whole, and has a reputation for high signal quality. Thus much of the overall quality in Verizon’s case must be either explained by aspects other than base station placement like pricing or phone selection or by higher average productivity per base station via technology differences, superior network management, and different spectrum. These facts persist even when looking at densities of base stations per PUMA land area. As explained later in Section 1.7.2, base station density determines distance to base stations

43 Table 1.12: Spearman Correlation Coeﬃcient of Unweighted Sample Market Shares over Markets

2008 2012

AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon

AT&T 1.00 - - - 1.00 - - - Sprint 0.10 1.00 - - -0.13 1.00 - - T-Mobile -0.04 0.24 1.00 - -0.29 -0.20 1.00 - Verizon -0.55 -0.243 -0.52 1.00 -0.35 -0.48 -0.18 1.00

Table 1.13: Spearman Correlation Coeﬃcient of Base Station Density over Markets

2008 2012

AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon

AT&T 1.00 - - - 1.00 - - - Sprint 0.96 1.00 - - 0.96 1.00 - - T-Mobile 0.95 0.95 1.00 - 0.95 0.97 1.00 - Verizon 0.95 0.96 0.96 1.00 0.96 0.95 0.95 1.00 which determines signal strength.34 To examine if firms strongly differentiate geographically via base station deployment, I use the Spearman rank correlation of base station density between carriers.35 As a rank measure, it will control for base station deployment between different firms being potentially higher or lower on average. This measure is very highly positive, revealing that the markets with the most (and least) investments are similar for all carriers.

34This land area measure, which will be used throughout this dissertation, does not include area in the PUMA covered by water.

35Spearman’s rank correlation simply takes the correlation between the integer rank of diﬀerent variables across all the observations.

44 Table 1.14: Rank Concordance of Base Station Density and Unweighted Sample Market Shares over Markets

Kendall’s W 2008 2009 2010 2011 2012 All

Market Share 0.66 0.79 0.79 0.78 0.78 0.73 Base Stations 0.46 0.30 0.25 0.17 0.18 0.24

In contrast, the rank correlation of Nielsen market shares varies considerably over the years, with both positive and negative values. Spearman correlation between Verizon market share and its rivals are consistently negative, suggesting that particular markets favor Verizon relative to all its rivals. T-Mobile and Sprint have a consistently positive rank correlation, which implies that they are more successful in similar markets and might be attractive to a similar clientele but do not strongly compete with each other. AT&T’s rank correlation with T-Mobile and Sprint varies from positive to negative depending on the year, so it is hard to characterize how those market shares vary. A look within markets shows that there is actually signiﬁcant rank variation in base station density, and this has changed over time. For each year, I calculate Kendall’s W which measures the concordance of list of rankings of the same objects.36 I consider each market a ranking in terms of base stations, and then through Kendall’s W measure how well these rankings match each other. In 2008, it starts at a rather high 0.72, but by the last period Kendall’s W has fallen to 0.18. A look at the data reveals this is mostly due to construction on Verizon’s part: while

36Formally, Kendall’s W is a transformation of the sum of squared deviations in rankings of the objects. I use the tie-corrected version of the formula due to a large number of ties in the data. Given Rk as the sum of the ranks of Qkm over m, I have

12 P R2 − 3 ∗ M 2 ∗ K(K + 1)2 W = k∈K k (1.1) M 2 ∗ K(K2 − 1) − M ∗ T where T is a correction term for the number of ties in the data.

45 Verizon is last place in base stations for most markets at the beginning of the sample period, over time it aggressively invested to become first, second or third in many markets. This is, however, not met with a similar reordering of the market leadership - the same exercise with market share shows that the rank concordance increases over time to 0.86 by 2012. A look at the data reveals this is largely due to Verizon already being the market leader for most markets in 2006. Verizon quality improvements only either cemented its dominance or improved its ranking with regards to share, so that by the end of the sample period the markets looked more similar, with Verizon on top. Overall, this suggests that while geographic variation is in the data, the effect is going to be relatively subtle and observed primarily through modest absolute differences in shares rather than drastic changes in ordering.

1.7 Evidence on the Relationship between Signal Qual-

ity and Base Stations

A general model of the relationship between demand, signal quality, and base stations is suggested by the anecdotal evidence. Demand is increased by own signal quality, but rival quality decreases demand. However, signal quality is not the only thing that determines quality, as price, phone selection, and service characteristics matter, too. A simple linear model of demand is

K X Dkmt = γkQkmt + δk,hQhmt + αkt + kmt (1.2) h6=k

k indexes each carrier, m indexes each market and t indexes the year. Qkmt is signal quality by carrier, market and year. γk is carrier-specific own quality effect, while δk,h are carrier- rival specific competitive effects. αkt represents carrier-year specific characteristics of the

46 carriers. kmt is carrier-market-year specific shock. Signal quality is in turn increased by increased numbers of base stations, but the impact of these varies by firm, through differences in spectrum, transmission technology and network management. A simply model of this is:

Qkmt = Fkm(Nkmt) + φkt + ψkmt (1.3)

The number of base stations Nkmt enters into quality via the function Fkm. Carrier and year variation in this production function due to spectrum, transmission technology and network management are represented by the carrier-specific component of Fkm and the fixed effect

φkmt. ψkmt is an idiosyncratic carrier-market-year-shock. I will discuss the speciﬁcs of Fkm, including the variation in this function by market, when I discuss how I bring Equation (1.3) to the data later.

1.7.1 Data on Signal Quality and Base Stations

To examine the relationship between demand, signal quality and base stations directly, I use the measurements of each from the Nielsen data and CSC data. For quality, I use customer ratings of their carrier that are included with the survey responses. There are four types of ratings: Voice, Network, Data, and Overall. Voice reports the quality of voice service; Network reports the quality of the network; Data reports the quality of internet and other data intensive activities; and Overall reports quality on all dimensions, even those not mentioned in the other questions. The ratings are on a 1 to 10 scale. Unlike the other survey responses, reporting of these measures is not mandatory - data quality is in particular not reported very often. This creates a non-reporting bias that might cause some individuals to report less based on unobservables. These signal quality measures have also the other selection problem that they represent the preferences of those who chose

47 those networks. They are thus likely biased upwards as these are individuals who have the best ratings for their networks.37 Thus conclusions based on the quality ratings must be cautiously interpreted. Tables 1.15 and 1.16 show the distribution of quality ratings across individuals and markets. Market level ratings are created by taking the weighted mean over the non-missing ratings. Ratings are skewed left, with most of the weight on 8, 9 and 10. As a result, the medians are all either 8 or 9, and the mean scores are very similar to each other. AT&T, SprintNextel and T-Mobile all have similar distributions with similar means. Verizon and the ratings for the other Carriers have notable higher ratings, with median and mean scores in voice, network and overall quality. At the individual level ﬁrms have a median score of 8 in data quality, and similar means. This result contrasts with the Top10Review grades, which suggested AT&T and Verizon together have superior signal quality. This also suggests that non-national carriers might have better quality because of less voice traﬃc, or since they lease network space from other carriers, are able to choose the best carrier in a location, i.e. in Connecticut, they disproportionately use Verizon’s network because it is the best. Table 1.17 shows the ratings are correlated. In general, the ratings are all very positively correlated. This is consistent with all signal quality dimensions relying on common techno- logical factors, or that the selection bias causes consumers to choose the options with high random quality shocks in every dimension. The lone exception is data quality, which seems less correlated than the other three. This may be because data relies more on equipment and spectrum while voice and network relies more on buildout, but it may also be because data is reported only about half the time and there is selection bias in the resulting subsample.

37The ideal ratings would be from an objective third party who meters all carriers equally, like RootMetrics. As noted before, while their data is public and online, it is so voluminous and coded in such a way that it resists automated and manual collection.

48 Table 1.15: Individual-Level Quality Ratings from Nielsen Sample

AT&T SprintNextel T-Mobile Verizon Other

Voice Mean 8.01 7.69 7.9 8.301 8.36 Median 8 8 8 9 9 SD 1.95 2.11 2.06 1.75 1.9 Obs 5078 1043 1202 5904 1726

Network Mean 7.71 7.21 7.36 8.36 8.13 Median 8 8 8 9 9 SD 2.1 2.31 2.27 1.69 2.01 Obs 5079 1402 1203 5904 1723

Data Mean 7.6 7.3 7.45 7.8 7.4 Median 8 8 8 8 8 SD 2.07 2.67 2.37 2.13 2.44 Obs 2132 637 453 2828 359

Overall Mean 7.77 7.42 7.75 8.24 8.28 Median 8 8 8 9 9 SD 2.1 2.18 2.13 1.83 1.97 Obs 5079 1407 1202 5905 1725

Total Observations 5,081 1,408 1,203 5,905 1,726

49 Table 1.16: PUMA-Mean Quality Ratings from Nielsen Sample

AT&T Sprint T-Mobile Verizon Other

Voice Mean 8.03 7.63 7.82 8.31 8.11 Median 7.97 7.79 7.97 8.31 8.28 SD 0.47 1.04 1.07 0.44 1.23 Obs 120 120 119 120 118

Network Mean 7.71 7.14 7.22 8.37 7.97 Median 7.73 7.28 7.52 8.33 8 SD 0.53 1.11 1.3 0.42 0.85 Obs 120 120 119 120 118

Data Mean 7.73 7.09 7.44 7.82 7.1 Median 7.76 7.36 7.86 7.86 7.46 SD 0.85 1.53 1.82 0.93 1.98 Obs 119 114 107 107 89

Overall Mean 7.78 7.36 7.58 8.26 8.1 Median 7.77 7.4 7.78 8.21 8.17 SD 0.51 1.1 1.08 0.41 1.08 Obs 120 120 119 120 118

Total Observations 120 120 119 120 118

50 Table 1.17: Correlations Between Rating Measures

Individual Level PUMA Level

Voice Network Data Overall Voice Network Data Overall

Voice 1.00 - - - Voice 1.00 - - - Network 0.74 1.00 - - Network 0.71 1.00 - - Data 0.53 0.54 1.00 - Data 0.37 0.41 1.00 - Overall 0.65 0.74 0.56 1.00 Overall 0.69 0.74 0.41 1.00

Due to diﬀerent levels of non-reporting there correlations are based on diﬀerent sample sizes: Voice and Network, Voice and Overall have sample sizes on the individual level of 15,305, and on the PUMA level of 597; Network and Overall 15,308 and 597; and any correlation with Data 6,399 and 597.

1.7.2 Log Base Station Density

Given the preceding discussion of signal quality, I assume signal quality to be a function of the log density of base stations in one’s local market area, i.e. Fkm(Nkmt) = ln(Nkmt/Am) in

Equation (1.3), where Am is the market’s land area. While imperfect, the use of this log base station density can be motivated by considering the very simple world where the following assumptions are true:

Assumption 1. Base stations are distributed uniformly across space.

Assumption 2. A consumer at any given time is at any given point in a ﬁnite travel/market area with equal probability.

Assumption 3. A consumer’s signal quality at a point is a decreasing function of the distance between that point and the nearest base station.

Assumption 4. A consumer’s utility is concave in signal quality.

51 Assumption 5. A consumer’s expected utility for signal quality is the expected value of utility from signal quality over all the locations.

In Appendix A.1, these assumptions imply many identical subdivisions in the market where the average distance is only a function of the relative size of those areas. As those sizes are determined by how many subdivisions are made in a fixed area, there is a linear relationship between the area per base station and square of the average distance. As distance increases, the power of electromagnetic transmissions drops off at an inverse-square rate or worse, so quality should be a function of the inverse of the area per base stations, i.e. the base stations per a given unit of area. In addition, I show in Appendix A.1 that this function of base station density is concave under these assumptions. Assumptions 1 and 2 are not strictly true since there is a lot of bunching in both base station location and human travel patterns. Björkegren (2013) fully accounts for the non- uniform distribution in his study of the Rwandan wireless phone industry, as he has access to phone record data from the national quasi-monopoly and can estimate the distribution of consumer locations based on their calls. Even without individual travel data, one can bring in aggregate traffic data to help estimate location distributions, as in Houde (2012). Unfortunately I have neither kind of data, so I cannot explicitly model utility in this way.38 However, both types of bunching tend to be in the same population dense areas. Carriers may be getting close to a geographic distribution of base stations that matches the distribution of consumers travel, so that Assumptions 1 and 2 may not be so far from the truth. Thus as a starting point, I use the fact that the utility function in the simple world is concave and will likely remain concave in the real world. Better locations will likely be chosen first so each subsequent base station should be less effective. To provide concavity

38Connecticut does have detailed traffic data - the Traffic Log, but this dataset only includes flows of traffic on segments of highways, and the distribution of the endpoint of trips cannot be inferred.

52 with parsimony, I use the log. Given a flexible intercept B and a flexible slope A, the loglinear function Y = B +Aln(X) provides a reasonable approximation of a strictly concave monotonically increasing function which asymptotically approaches −∞ at 0 and is defined over R+. Parameters analogous to the slope A and intercept B in the formal model will also be allowed to vary by firm to control for the variation in spectrum transmission technologies and network management efficiency. Finally, to make demand in one market independent from quality in other markets, I make the following assumption.

Assumption 6. The set of travel/market areas is ﬁnite and travel/market areas do not overlap.

This is not strictly necessary for reduced form regressions in this chapter, since those only purport to show correlations and do not control for all factors. However, this assumption will be useful in the next chapter’s game theoretic analysis of the industry. This assumption will likely be close to reality when the spillover eﬀects of base stations in nearby markets are very small. This approach to study facilities investment has precedent in the ATM literature. Ferrari, Verboven, and Degryse (2010) also assumes that consumers have concave utility for the density in the local area. That paper assumes that consumer utility for an ATM network is based on the average travel cost to the nearest ATM and considers cost to be linear in distance traveled. Using a derivation from an earlier paper on ﬁre engine response times by Kolesar and Blum (1973), Ferrari, Verboven, and Degryse (2010) models the average distance to be the inverse square root of the density of ATMs in distinct postal code zones. Ishii (2007) is also similar, but she uses the count and not a function of density.

53 Figure 1.13: Log Base Station Density on Quality Rating

54 1.7.3 Demand for Signal Quality and Base Stations

To relate the signal quality measures to firm investment in base stations, I compare the market average quality ratings over the sample consumers to the log density of base stations.39 I use area in 1000 KM2 so that the log density is between 1 and 10, and thus always positive. As base station density increases, the signal quality should increase and demand for that carrier should increase. Figure 1.13 which plots each quality measure on the log density of base stations illustrates that the relationship does not seem especially strong, and from this perspective seems very flat.40 Figures 1.14 and 1.15 show the scatterplot of the log base station density on log shares. Log shares are less skewed from data perspective and theoretically often the log share is the variable of interest.41 From the figures, the quality ratings seem to have a strong positive impact, while log base station density has a smaller but still positive effect. This is consistent with the idea that log base station density only indirectly impacts the utility though actual signal quality. To be more precise, I conduct multiple equation estimation using versions of Equations

(1.2) and (1.3). I modify Equation (1.3) using the log density of base stations for Fkmt. For ¯ Equation (1.2), I use as my measure of quality the weighted average quality ratings Qkmt and log share ln(Skmt) as my measure of demand. Since parameters can vary for the four different firms, each equation actually has four different versions. Thus each of the four firms

39The rating average I use is the weighted average using the included weights from Nielsen.

40The plot points are weighted by the number of observations used to construct the average.

41For example, Berry (1994) shows that in a multinomial discrete choice logit model with a linear indirect utility function, a function of log shares is linear in the product characteristics.

55 Figure 1.14: Quality Ratings on Log Shares

56 Figure 1.15: Log Base Station Density on Log Shares

57 k has two equations used for estimation:

K ¯ X ¯ ln(Skmt) = γkQkmt + δk,hQhmt + αkt + kmt (1.4) h6=k ¯ Qkmt = λkln(Nkmt/Am) + φkt + ψkmt (1.5)

This results in an eight equation system, where γk represents the own beneﬁt of quality by

firm, δk,h represents the impact of rival h’s quality on the demand of carrier k, and λk is the productivity of log base station density on quality. ¯ Clearly, Qkmt are endogenous variables in this system that the firm sets to change demand. ¯ Moreover, there is also certainly measurement error since Qkmt are simply sample (weighted) means. To address this, I fit this problem into a multiple equations instrumental variables format, which will also remain flexible enough to allow correlation in errors across equations for the same market and year.

Deﬁne the matrix of demand covariates for each market year as Xmt. There are four demand equations with an own quality term, the four potential rival quality terms and the

five year fixed effects, so this results in Xmt being an 8 × 36 block diagonal matrix. Let the width of Xmt be denoted by L. Let demand equation errors for the market-year be the matrix mt. Define the matrix of demand covariates for each market year as Zmt. Theoretically, I can then get consistent identification of the parameters of the demand equations, under the following assumptions that make Zmt valid instruments for Xmt:

Assumption 7.

E[mt|Zmt] = 0 (1.6)

58 Assumption 8.

0 RANK(XmtZmt) = L (1.7)

Assumption 7 assumes that the instruments Zmt are not correlated with the errors of the demand function. Assumption 8 assumes that there are at least as many instruments as there are coefficients to estimate. The fixed effects are clearly exogenous and will instrument themselves, but I need instruments for the four quality terms in the demand equations. The obvious candidate for these is log density of base stations by firm. If all log base station densities of all firms are assumed exogenous to all market shares of all firms, then the Assumptions 7 and 8 will be satisfied as this will provide 4 excluded instruments for the 4 endogenous quality parameters in each demand equation. Assumption 7 is likely violated since the base stations are endogenously deployed by the firms to produce quality. The estimates therefore could be even more biased than they would have been without instrumenting. However, estimates under this incorrect assumption will shed light how the estimation can be improved, so I will work under this assumption until the end of this chapter, and then comment how in the next chapter these results can be improved. While this problem could be estimated under a variety of Generalized Method of Moments estimators (Hansen (1982)), the use of the same instruments for different equations fits nicely into the 3 Stage Least Squares (3SLS) framework introduced by Zellner and Theil (1962).

To apply 3SLS I make a further assumption of conditional homoskedasticity of the kmt in each equation:

59 Assumption 9.

V ar(mt|Zmt) = V ar(mt) (1.8)

Under this assumption, the 3SLS will be more efficient than a more flexible GMM esti- mator. It is hard to validate this assumption ex ante, though it does simplify the estimation procedure substantially. A final point before presenting the results is that I have multiple quality dimensions

(Data, Network, Voice and Overall), which could all feasibly depend on ln(Nkmt/Am). Since these are quite correlated with each other, and I do not have enough instruments if I were to include all quality measures simultaneously in demand, I simply estimate four different systems using each one. Table 1.18 reports the results of the first stage regressions, which are equation by equation regressions for each endogenous variable used in any system. Here I regress the full set of excluded instruments, the log base station densities of all firms, on each firm’s market share and mean quality ratings. The strength of excluded instruments is mixed, as measured by the F-Statistic testing the difference in explanatory power between a regression with the excluded instruments and without. While there seems to be a strong impact of the instruments for particular regressions, no system has F-Statistics larger than 10 for all equations, the standard rule of thumb suggested by Stock, Wright, and Yogo (2002).42 Interestingly, the instruments seem to matter more in the regressions for log share, rather than quality. This might imply that the log base station density may have explanatory

42Technically, the Stock, Wright, and Yogo (2002) rule-of-thumb is for the single equation, single variable context, but given possible cross equation dependencies, one should hope that F-Stat should be even larger than 10 in each equation. Angrist and Pischke (2009, p. 217-218) suggests a multiple endogenous variable test statistic for the single equation case, but currently my understanding is that there is no widely agreed-upon test for the multiple equation case.

60 Table 1.18: First Stage: Log Base Station Density on Log Shares and Quality Measures

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Dependent Shares Voice Network Variable: AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon

AT&T ln(N/A) 0.02 -0.07 0.12 0.12 0.13 0.06 -0.62 0.00 0.17 0.35 -0.51 0.08 (0.07) (0.24) (0.33) (0.13) (0.20) (0.56) (0.31) (0.29) (0.25) (0.53) (0.49) (0.24) Sprint ln(N/A) -0.12 0.64* 0.27 -0.58 -0.18 1.00 0.84 0.20 -0.33 0.97 1.36*** -0.14 (0.16) (0.26) (0.35) (0.28) (0.26) (0.63) (0.50) (0.27) (0.28) (0.63) (0.26) (0.23) T-Mobile ln(N/A) 0.05 -0.58* 0.55** 0.34* 0.05 -0.91* -0.50 0.13 0.14 -0.92** -1.02** 0.05 (0.14) (0.21) (0.20) (0.14) (0.14) (0.39) (0.33) (0.30) (0.22) (0.22) (0.30) (0.15) Verizon ln(N/A) 0.01 0.25 -0.53 -0.06 0.20 -0.23 0.58 -0.36 0.26 -0.21 0.76 0.07 (0.08) (0.26) (0.25) (0.11) (0.18) (0.48) (0.50) (0.35) (0.22) (0.50) (0.40) (0.31)

Observations 119 119 119 119 119 119 119 119 119 119 119 119 R-squared 0.03 0.21 0.48 0.22 0.09 0.05 0.09 0.04 0.10 0.04 0.16 0.01 F-Stat 26.72*** 7.274*** 70.03*** 81.58*** 8.915*** 3.208 66.13*** 26.28*** 17.28*** 5.648*** 72.86*** 1.654 61

(13) (14) (15) (16) (17) (18) (19) (20) Dependent Data Overall Variable: AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon

AT&T ln(N/A) -0.18 0.87 0.37 0.49 0.13 0.42 -0.34 0.17 (0.29) (0.73) (1.18) (0.24) (0.20) (0.61) (0.55) (0.09) Sprint ln(N/A) 0.35 -0.07 0.42 -1.16 -0.03 0.80 0.25 -0.28* (0.39) (0.28) (1.43) (0.73) (0.12) (0.64) (0.54) (0.13) T-Mobile ln(N/A) 0.14 -0.34 0.32 0.06 0.13 -0.78 -0.01 0.21 (0.42) (0.46) (1.35) (0.19) (0.20) (0.54) (0.18) (0.10) Verizon ln(N/A) -0.09 -0.28 -0.91 0.67 -0.00 -0.40 0.55 -0.14 (0.05) (0.59) (0.84) (0.57) (0.22) (0.40) (0.42) (0.25)

Observations 101 101 101 101 119 119 119 119 R-squared 0.17 0.01 0.04 0.08 0.13 0.03 0.12 0.03 F-Stat 12.35** 0.812 22.49* 8.000** 230.2*** 0.885 248.1*** 2.344 Year fixed effects included in all regressions. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Observations vary since some ratings not observed for any respondent in particular market-years. power than is not captured in the quality ratings alone. Of course, the performance of individual coefficients is poor, as very few are statistically significant and many do not have the expected signs from theory. Own log base station density should have a positive impact on the independent variables as improving quality, but the associated coefficients are often negative. The relatively poor performance of the instruments, which, as discussed earlier, is not surprising given log base station densities are not likely exogenous to the demand errors. The poor performance continues in the results for the actual 3SLS, which I report in Table 1.19. The parameters of the demand equations are particularly poorly estimated, as they are often not statistically significant and often have the wrong sign. Own quality should have a positive impact and rival quality should have a negative impact, but in each system these parameters often have the opposite sign, often with statistical significance. The production system seems to have somewhat more precise results, as the impact of base stations on quality for AT&T and T-Mobile are usually very statistically far from zero. However, while the T-Mobile effect is twice as large as the AT&T effect in the Network Quality System, the effects are never statistically different from each other in any system.43 In addition, Sprint and Verizon base stations never have a statistically significant effect and are sometimes negative. Thus the fully unconstrained systems with firm specific coefficients do not seem to be well identified, either because of the aforementioned endogeneity issues or simply low sample size. As an alternative, I make a restriction. If measured in the same units across carriers then quality could have roughly the same impact to the log share across carriers as well. So for a second set of system estimates, I estimate 3SLS under the restriction:

• Restriction: γk = γ, δk,h = δ

43That is, two eﬀects are never simultaneously outside each other’s 95% conﬁdence intervals.

62 Table 1.19: Demand and Production System Estimates: Unrestricted Parameters

System by Quality Measure: Voice Network Demand Equation: (1) (2) (3) (4) (1) (2) (3) (4) Carrier k AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon

γk -0.11 0.87*** 0.54* -0.66*** -0.27 0.42 0.60 -0.53 (0.35) (0.29) (0.32) (0.20) (2.07) (0.32) (2.10) (10.59)

δk,AT &T 0.58 1.59 -0.32 -5.10 12.76 1.24 (0.73) (1.23) (0.59) (3.35) (28.04) (3.98)

δk,Sprint -0.09 -0.61* -0.50** 0.16 -3.54 0.31 (0.16) (0.36) (0.25) (0.21) (2.36) (0.37)

δk,T −Mobile -0.11 0.66*** -0.13 0.95*** -0.79** (0.13) (0.24) (0.22) (0.37) (0.37)

δk,V erizon -0.30 -0.58 5.17*** -0.60 1.75 16.00 -52.45 (0.28) (0.55) (0.68) (0.46) (5.96) (10.14) (65.58) F-Stat 0.827 6.507*** 7.461*** 5.468*** 3.109 0.917*** 3.110*** 0.876***

Production Equation: (5) (6) (7) (8) (5) (6) (7) (8)

λk 0.19*** -0.07 0.22** 0.00 0.21*** 0.14 0.41*** 0.04 (0.06) (0.11) (0.11) (0.05) (0.07) (0.12) (0.13) (0.05) F-Stat 2.843** 1.082 1.745 1.560 3.109*** 0.917 3.110*** 0.876

Observations 119 119

System by Quality Measure: Data Overall Demand Equation: (1) (2) (3) (4) (1) (2) (3) (4) Carrier k AT&T Sprint T-Mobile Verizon AT&T Sprint T-Mobile Verizon

γk -0.07 1.58*** 2.22*** 0.90*** -4.51** 4.59*** -47.43*** 18.22*** (0.23) (0.48) (0.48) (0.33) (2.14) (1.20) (8.91) (5.71)

δk,AT &T 2.22 0.08 -1.54* -12.62*** 73.07*** -9.97*** (1.41) (1.16) (0.87) (4.24) (13.14) (3.50)

δk,Sprint -0.16 -1.07** -0.80** 1.31** -22.05*** 2.60** (0.22) (0.46) (0.38) (0.63) (3.85) (1.05)

δk,T −Mobile -0.03 -2.22*** 1.33*** 2.93** 8.64*** 6.45*** (0.22) (0.45) (0.36) (1.44) (2.69) (2.45)

δk,V erizon 0.08 -1.42*** 0.97** 7.42** 17.57*** -114.01*** (0.16) (0.46) (0.48) (3.45) (6.51) (20.34) F-Stat 4.479 1.650*** 3.377*** 1.193*** 0.713 1.491*** 4.564*** 1.867***

Production Equation: (5) (6) (7) (8) (5) (6) (7) (8)

λk 0.29*** 0.08 0.26 -0.08 0.25*** 0.02 0.37*** -0.05 (0.07) (0.17) (0.18) (0.12) (0.06) (0.12) (0.11) (0.05) F-Stat 4.479*** 1.650 1.068 1.193 3.575*** 1.491 2.742** 1.867*

Observations 101 119

Year fixed effects included in all regressions. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Observations vary since some ratings not observed for any respondent in particular market-years. 63 This implies that the effect of own quality and rival quality is the same for all firms - aside from fixed effects the equations are the same.44 This restriction results in only two demand parameters to be identified rather than 16, resulting in an over-identified system. The restricted estimates are reported in Table 1.20. For the demand equations, the precision markedly improves, and the parameters always have their theoretically predicted signs. Own quality improves log share with statistical significance under any quality measure, and rival quality reduces log share with statistical significance in the Network quality and Overall quality systems. The production equations in those systems also precisely identify their carrier-specific productivity parameters for all carriers beside Verizon, whose production parameter is not statistically different from 0. In those specifications, the AT&T, Sprint and T-Mobile productivity parameters are statistically significantly different from each other and can be ranked in productivity in that ascending order. The productivity estimates for the other systems are less precise: in the Voice quality system, Sprint productivity is not statistically significantly different from 0, and Verizon is found to be negative with statistical significance. That the Network and Overall systems have results more congruent with the model may be due to base stations being more strongly connected with generating quality in network coverage and Overall quality also including network coverage. Voice and Data quality may have more to do with spectrum and thus may not reflect base stations as strongly. Taking the results as a whole, while base stations seem to matter for network quality and quality seems to matter for demand, carrier specific effects are not precisely estimated. As these seem to be important to account for other aspects of quality besides base stations, this is a drawback of the analysis. Moreover, while endogeneity may be especially important

44I estimated systems with restrictions in between the fully ﬂexible system and the fully restricted system. For example, with δk,h = δh or δk,h = δk. None of these resulted in as precise estimates as the fully restricted systems, so I do not report them here.

64 Table 1.20: Demand and Production System Estimates: Restricted Parameters

System by Quality Measure: Voice Network Data Overall Dependent Variable: ln(Share) ln(Share) ln(Share) ln(Share) Equation Parameter (1-4) γ 0.51*** 0.73*** 0.71*** 0.76*** (0.12) (0.09) (0.10) (0.11) δ -0.04 -0.13*** -0.02 -0.15*** (0.03) (0.02) (0.04) (0.05) F-Stat (AT&T) 3.653*** 2.309*** 8.628*** 1.209*** F-Stat (Sprint) 3.974*** 5.852*** 8.706*** 7.923*** F-Stat (T-Mobile) 4.140*** 10.57*** 6.879*** 8.157*** F-Stat (Verizon) 3.803*** 0.758*** 3.265*** 2.195***

(5) λAT &T 0.09* 0.12** 0.08 0.11** (0.04) (0.05) (0.06) (0.05) F-Stat 1.479 10.20** 1.768 7.762

(6) λSprint 0.07 0.34*** 0.23*** 0.28*** (0.08) (0.07) (0.07) (0.07) F-Stat 1.184 10.40*** 3.585*** 4.781***

(7) λT −Mobile 0.38*** 0.60*** 0.49*** 0.49*** (0.08) (0.08) (0.09) (0.07) F-Stat 5.667*** 11.48*** 8.742*** 9.779***

(8) λV erizon -0.06 0.00 -0.22*** -0.07 (0.05) (0.05) (0.07) (0.04) F-Stat 1.886* 10.390* 8.678*** 7.913*

Observations 119 119 101 119

Year ﬁxed eﬀects included in all regressions. Robust standard errors in parentheses, clustered on carrier-year. *** p<0.01, ** p<0.05, * p<0.1 Observations vary since some ratings not observed for any respondent in particular market-years.

65 there, the instruments for signal quality cannot be plausibly expected to be exogenous. Further, these shares are based on very small samples. The signal quality ratings are based on subsets of those samples, and suffer from selection bias. There is likely variation in consumer tastes that could have strong effects, which suggests a more disaggregated estimation approach that takes advantage of the demographic information I do have. So while the current results are suggestive of a connection between base stations, signal quality and consumer demand, a more precise estimate is needed. A restriction on coefficients seems to help with identification, so a fully structural system of supply and demand, with a different strategy for identification, may be useful.

1.8 Conclusion

Using a combination of national level and detailed local data, I have analyzed the market for wireless telephone service and the accompanying physical capital investments. While smaller carriers could be differentiating themselves from the four national carriers through cheaper plans but inferior plan features and phone selection, differentiation between the national carriers seems to be driven by differences in signal quality. Analyzing detailed data on demand and base stations for the state of Connecticut, I find that base stations by themselves do not explain the market share differences. However, controlling for firm specific mean quality which may represent differences in transmission technology, spectrum, network management ability and price, I find that base station density does have explanatory power for measures of signal quality and that signal quality has explanatory power for market shares. However, these regressions do not control for endogeneity well, the imprecision in the measured covariates, or selection bias. I also do not exploit individual level variation in the value of plans. The later seems to be especially important as analysis of demand over time implies strong secular growth in the demand for

66 the non-national plans and prepaid plans over the sample period. To that end I propose a more careful analysis of demand, which includes demographic controls and time-varying demand for diﬀerent brands and plan types, and controls for the endogeneity of base stations. I explore this approach in the next chapter, which I then use to also answer the important policy question of how consolidation in this industry would aﬀect quality provision.

67 Chapter 2

Quality Competition in Mobile Telecommunications: Evidence from Connecticut

Patrick Kainin Sun

68 2.1 Introduction

There has long been interest in the impact of market structure on product characteristics. Works such as Salant, Switzer, and Reynolds (1983), Perry and Porter (1985), Deneckere and Davidson (1985) and Farrell and Shapiro (1990) examined equilibrium price effects of mergers and showed that they depend on more than just industry concentration. Given these ambiguous effects, later economists began to use new empirical techniques to estimate the potential effects of mergers. Early examples, like Werden and Froeb (1994), Nevo (2000), and Town and Vistnes (2001), focused on price effects as the theory literature had, but later works, like Draganska, Mazzeo, and Seim (2009) and Fan (2013) also looked at the effect of other product characteristics. These studies match recent theoretical work in Gandhi et al. (2008) and Mazzeo, Seim, and Varela (2014) that show repositioning of products in a merger can have welfare impacts equal to or larger than price. Such research has been in line with antitrust policy which now considers price and quality effects in their merger analyses.1 As demonstrated in Chapter 1, an important example of product quality is signal quality in the wireless industry. As base stations become more common in an area, the signal quality improves as the average distance between consumers and the antennas decreases. A carrier can increase its market share by building more base stations in a local market, but base stations are costly in terms of materials, power, maintenance and regulatory compliance. Carriers must build their own base stations to serve their customers, so base station investment is a competitive activity.2 Unlike price, quality improvements in one firm may not induce other firms to improve quality in kind to compete - rather they may be discouraged from competing directly with the now stronger rival and reduce their quality response. Thus

1See U.S. Department of Justice and Federal Trade Commission (2010) and European Commission (2004).

2Carriers occasionally share a single base stations in extraordinary situations.

69 a natural question is how competition affects the incentive to provide signal quality in this economically important industry. Moreover, given this industry is quite concentrated and inclined to consolidation, the impact of a merger on quality provision is of paramount importance. Since repositioning in signal quality might counteract or reinforce negative price effects from a merger, the merger effect on incentives to provide signal quality may be pivotal to whether a merger is consumer welfare reducing or not. Competitive analyses of market structure changes, both in antitrust practice and the academic literature, generally focus on price changes, holding all other quality dimensions fixed. In contrast, I conduct an analysis treating signal quality as an endogenous variable under the control of firms, which allows firms to reposition their products after the merger. Consumer utility for a given carrier’s network is modeled as a function of the density of that carrier’s base stations. Design of the demand system is done in the light of the findings from Chapter 1, and incorporates difference in brand and plan preferences across demographics and across time. I also control for the endogeneity of signal quality using local land use regulation as a cost-based instrument. I estimate this model using two unique datasets for Connecticut where I directly observe base station location and ownership and consumer carrier choices in the state from 2008- 2012. I combine my demand estimates with a model of quality competition to recover the costs of maintaining base stations. I then use the parameter estimates and full model to run counterfactual simulations of the proposed AT&T and Sprint acquisitions of T-Mobile. Across carriers, markets and years, I find that a 1% increase in the observed level of log base station density results in a median market share gain of 0.17% for the investing firm and median losses of 0.04% for each rival firm. These small marginal effects are not unexpected since a firm invests in base stations until the return for the marginal base station is low. In contrast, inframarginal effects implied by expenditures on all base stations are

70 substantial. For example, T-Mobile is estimated to spend about 60% of variable profits on its base stations. Further, the demand estimates and model of competition imply that base stations are locally a strategic substitute, in the sense of Bulow, Geanakoplos, and Klemperer (1985): rival increases in base stations reduce the incentives to provide own-base stations. Mergers may cause firms to adjust base stations in one direction, but that will in turn cause rival firms to adjust base stations in the opposite direction. Thus the overall sign of welfare effects is ambiguous without accurate parameter estimates. Simulation of mergers between a “small” carrier (T-Mobile) and two of its major rivals (AT&T and Sprint) have two main results. First, I confirm findings in both the theory and the empirical literature that show that mergers have a tendency to increase differentiation in merging firms’ products and that this benefits non-merging rivals. In simulations that retain both networks of the merged firms, the signal quality of T-Mobile’s network is reduced significantly relative to its merger partner. This comes from differences in diversion ratios with respect to signal quality: for the same percentage change in signal quality, T-Mobile steals significantly more share relative from its merging partner than its merging partner does from it. Thus a merger between the two results in the joint firm degrading the relative signal quality of T-Mobile as that results in less cannibalization. The non-merging firms benefit as their as T-Mobile is positioned much farther away from the other firms. Second, I find the scope for integration of the two merging firms’ networks is crucial for consumer-welfare improving conduct. Eliminating the acquired carrier and its network entirely results in the remaining firms increasing quality due to strategic substitutability, but not enough to make up for the consumer welfare loss from the decrease in variety. Keeping the acquired product line around instead but also keeping the networks separate results in a decrease in signal quality by the two merged firms since now they internalize the

71 negative impact each network has on the others’ product lines. Only in the counterfactuals where the consumers can use their phone on both networks are consumer gains realized. A base station can serve consumers who have a horizontal taste for either product line of the merging firms, whereas before they could only serve one or the other. This spillover across product lines makes marginal investments in base stations more effective in terms of attracting consumers. This translates into a greater incentive to provide quality relative to its cost. Under reasonable assumptions about prices and costs, consumers benefit as effective quality improves even if the total number of base stations decreases. These effects are qualitatively similar whether the acquiring firm is the AT&T or Sprint, though the negative impacts of mergers are blunted somewhat when the acquirer is the smaller Sprint. Thus from the perspective of competitive and telecommunications policy, merger reviews in the wireless and other network-based industries should require detailed evidence from applicants about the potential and plans for network integration. My treatment of base station density as an endogenous non-price characteristic is very similar in model to the previously mentioned Draganska, Mazzeo, and Seim (2009) and Fan (2013). Like those papers, this means the analysis also belongs to the discrete choice demand estimation literature which controls for endogenous product characteristics, such as Berry (1994) and Berry, Levinsohn, and Pakes (1995). This study also contributes to the literature on the economics of wireless service, surveyed in Section 1.2 of Chapter 1. This paper belongs to the branch of that literature that uses discrete choice demand systems to estimate wireless operator incentives. Often these papers include signal quality as a component of consumer utility, but only as an exogenous control. For example, Zhu, Liu, and Chintagunta (2011) includes averages of online ratings of signal quality while Sinkinson (2014) uses proprietary Nielsen measures.3 Similarly, Macher et al.

3These Nielsen quality measures are in fact diﬀerent from the measures used in Chapter 1. Instead of survey responses, Sinkinson (2014) uses direct signal quality readings, or “drive tests,” so-called as they are

72 (2013) studies the substitution and complementarity of fixed and wireless lines, and include the total national number of cell sites in their demand system to proxy for improving quality of cell service overall. Miravete and Röller(2004) also includes cell sites in their analysis, though they do not include it as a quality proxy. Rather, they use it to proxy demand since they assume each site serves some fixed number of customers. My paper is distinguished from the above as its focus is the carriers’ incentives to change signal quality so signal quality cannot be assumed exogenous. In this respect, the most similar paper in the literature to mine is Björkegren (2013), who looks at the Rwandan wireless quasi-monopoly to estimate positive demand externalities consumers have on each other. As he has access to the Rwandan operator’s private data, he also has information about base station location and includes coverage as an endogenous component of utility. Given the complexities of his model, he cannot fully simulate equilibrium coverage provision even for the monopoly, but does partial equilibrium counterfactuals about base station deployment in response to a government program.4 In contrast to Björkegren (2013), my model is greatly simplified, but provides a unified framework for policy experiments taking into account the strategic aspect of quality decisions. In the remaining sections of this chapter, I illustrate how I implement this framework. I explain how my model captures the aspects of this industry relevant to signal quality provision and the results from estimation of that model. I then implement a variety of counterfactual simulations using my results to explore mergers in this industry and then conclude with an overview of the findings. However, since the incentives behind signal quality provision may not be obvious, I start with a simple example model to illustrate the sampled by surveyors driving around the country.

4Speciﬁcally, Bj¨orkegren (2013) removes the 10 base stations with the lowest revenues to simulate demand but for a policy imperative for carriers to serve rural areas.

73 intuition.

2.2 Competitive Eﬀects of Quality

The welfare effects from a market structure change are largely determined by whether quality is a strategic complement or a strategic substitute, as defined by Bulow, Geanakoplos, and Klemperer (1985). In a game, strategic complements are a set of control variables for the players such that if a change in one player’s variable induces rivals to change their variable in the same direction. For example, price is a strategic complement in Bertrand competition. In that game, the downside of cutting price is that while lower prices brings new consumers, old consumers who would have bought at the original price are now given a discount. If a rival decreases price, there are fewer old consumers so the gross loss via the discount to these consumers is smaller and price cutting is less costly. Analogously, strategic substitutes are control variables that when changed induce changes of rivals in the opposite direction. Quantity in Cournot competition is a strategic substitute. If a rival expands their demand, then the market price goes down. Thus own demand expansion is less beneficial since there is less revenue per consumer. If quality is a strategic complement, signing the welfare effect of a merger is straightforward, since the effect the merger has on the quality of the merging firms would be reinforced by like quality changes of the non-merging firms. Price is generally a strategic complement, so the impact of a merger through price is straightforward: a merger that would induce price increases holding non-merging firm’s prices fixed must be anti-competitive since full equilibrium would only imply more price increases by rivals. But if quality is a strategic substitute, then the sign of the welfare effect of the merger is ambiguous, since any effect on the quality of the merging firms might be completely canceled out by the changes of the

74 non-merging firms. Therefore, analysis of mergers taking endogenous quality into account needs to determine both whether the merger will induce merging parties to change quality and the direction of the response of non-merging rivals. In the model I take to the data, it turns out that strategic complementarity and substitutability depend on the shape of the demand function and where the relative utilities of the plans put the different carriers on that demand function. To illustrate the forces at work in the estimated model, consider the following simple example model. Let there be two carriers, indicated by k ∈ {1, 2}. Each offers a single product. In a first stage, the carriers set national prices. In the second stage, they set local signal quality Qk by adjusting the number of their base stations. Consumers choose the carrier which gives them the most utility or an outside option k = 0. Utility of the two products are

Uik = Qk + ik (2.1)

ik is a mean-zero random shock, which is independently and identically distributed over ik and explains why all consumers do not just choose the carrier with highest mean quality. For simplicity, I normalize the outside option to always have utility 0.

The market share is determined by a function Sk(Q1,Q2) of the signal quality. Assuming a market population and constant markups normalized to 1 and cost function φ(Qk), proﬁt is

πˆk = Sk(Q1,Q2) − φ(Qk) (2.2)

Taking Qk as continuous and φ(Qk) as suﬃciently convex, then a pure strategy Nash equi-

75 librium exists and the necessary ﬁrst order condition is:

∂π ∂S (Q ,Q ) ∂φ(Q ) k = k 1 2 − k (2.3) ∂Qk ∂Qk ∂Qk

i.e. marginal variable profit for quality equals marginal quality cost. Invoking Topkis (1978) and Milgrom and Shannon (1994), the comparative statics depend on the sign of this cross partial derivative of profit since the control variables are assumed continuous. The cross partial of the example profit function of k with respect to rival quality h depends entirely on the share/demand function, since rival signal quality does not enter the cost function.5 Thus the pivotal factor in determining strategic substitutes or complements will be how a change in rival quality affects the number of marginal consumers from a quality increase. If the marginal consumers increase in number, then quality is a strategic complement; if marginal consumers decrease, then strategic substitutes. What makes a consumer marginal for 1? A consumer must be indifferent between 1 and 2 or the outside option, which implies she must get certain levels of shocks such that she has the same utility for 1 as 2 or the outside option. Denote the identical CDFs of these shocks as G and their PDFs as g.

Consumer i chooses good 1 if Q1 + i1 > 0 and Q1 + i1 > Q2 + i2. If Q2 + i2 < 0 ↔

i2 < −Q2 then the outside option utility is always greater than that of good 2. Demand is then equal to how often the utility of good 1 is greater than the utility of the outside option

5With a different cost function, this implication might change. For example, Chu (2010) studies quality provision in the form of channels offered by cable companies. In his case, quality costs do not enter separately from demand, since channel contracts payments are per subscriber. Thus it is possible in that setting, even without the heterogeneity he includes in his specification, to have entry of satellite competition or rival improvements in quality induce own quality improvements since the resulting loss of demand reduces marginal consumer costs. In the wireless industry, marginal consumer costs should, if anything, go down with more own base stations, since it might be less costly to maintain calls with a smaller territory associated with each base station. That assumption would imply strategic substitutability of base stations even more strongly, since now entry or rival quality improvement decreases own demand and thus decreases total cost per marginal consumer.

76 - the probability that Q1 + i1 > 0 ↔ i1 > −Q1. Analogously, if Q2 + i2 > 0 ↔ i2 > −Q2 then the good 2 utility is always greater than that of the outside option. Demand is then equal to how often the utility of good 1 is greater than the utility of good 2 - the probability that Q1 + i1 > Q2 + i2 ↔ i1 > Q2 − Q1 + i2. Demand of good 1 is therefore expressible as the following integral over the distribution of the normalized shocks:

Z −Q2 Z +∞ S1(Q1,Q2) = (1 − G(−Q1|i2))g(i2)di2 + (1 − G(Q2 − Q1 + i2|i2))g(i2)di2 −∞ −Q2 (2.4)

An infinitesimal change in Q2 has an infinitesimal impact on the demand of good 1 when good 2 is the worse than the outside option since all substitution happens between 1 and 0 there. The only effect is to infinitesimally reduce the support on which this is the case. So

Q2 has effectively no impact on marginal return from that first part of the equation. Thus the entire effect in the cross partial will be on the range where good 2 is better than the outside option, i.e. where consumers are substituting directly between 1 and 2. The resulting cross partial is:

2 Z +∞ ∂ S1 0 = g (Q2 − Q1 + i2|i2)g(i2)di2 (2.5) ∂Q1∂Q2 −Q2

This expression makes sense - the overall cross partial is the conditional expected value of how fast the density of marginal consumer between 1 and 2 grows as the relative quality of the rival good 2 grows. If greater rival relative quality increases the number of marginal consumers, then the PDF grows, and quality must be a strategic complement. Analogously, if greater rival relative quality reduces the number of marginal consumers, quality is a strategic substitute.

77 Figure 2.1: The lines separate the distribution of consumers, represented by the shading, into those who buy Product 1, 2 and the Outside Option 0. The origin is set to (−Q2, −Q1). The Green line represents consumers indiﬀerent between 1 and 0, the Blue between 0 and 2, and the Red between 1 and 2. The shift in lines represents a change in the quality of 2.

i1 i1

1 1

i2 i2

0 ∆Q2 2 0 ∆Q2 2

(a) Q1 is a strategic substitute for Q2 (b) Q1 is a strategic complement for Q2

Figure 2.1 diagrams the example model in two diﬀerent cases. Here I represent the

distribution of consumers by plotting the 1 and 2 space and by shading the background of the graph to represent denser parts of the space. I assume the shocks are unimodal, which is equivalent to assuming consumers are less common the more extreme their predisposition to either goods 1 and 2. The carriers and outside option split the space of consumers: the outside option taking anyone who does not have at least shocks greater than −Q1 or −Q2;

1 taking remaining consumers with high i1 and low i2; and 2 taking remaining consumers with high i2 and low i1. The Red line represents consumers indifferent between 1 and 2, the Green indifferent between 1 and the outside option, and the Blue the indifferent between 2 and the outside option. The marginal incentives to invest are represented by the density of marginal consumers who would switch between firms by an infinitesimal quality increase. For 1, this is equal to the consumers along the Green and Red lines, and for 2, the Blue and Red lines.

The diagram shows the effect on those lines after an increase in Q2. There are fewer consumers who are indifferent between 0 and 1 since some now prefer 2, so the green line gets shorter, but while this looks substantial in the diagram, as explained earlier with infinitesimal

78 rival quality changes this effect also becomes infinitesimal. The first order change is that the Red line gets shifted left, meaning that the consumers who substitute between 1 and 2 must be more biased in terms of shocks relative to 1. In the left subfigure, one can see that the shift moves the Red line into a less dense region of the graph, so the Red line must

contain fewer consumers. Thus the Q2 increase decreases marginal consumers, so quality is a strategic substitute for the example drawn above. The marginal consumer between 1 and 2 must now be more predisposed to 1, but this means that consumer is less common. However, this does not mean quality is a strategic substitute for all quality levels. If the starting points of the lines were elsewhere in the graph, say more to the right as in the right subﬁgure, then the same shift would result in the Red line going from a region of few consumers to the center region with more consumers. Then quality is a strategic complement since the Red line, the marginal consumers between 1 and 2, has more density.

This corresponds to the case where the qualities are not similar: Q1 is much higher than Q2, so 1 has high market share and has captured most of the market. The marginal consumers for 1 are thus people who are actually very predisposed not to buy 1, so they are few in number. When Q2 increases, 2 is a better product and thus those consumers will switch to 2. The marginal consumers between 1 and 2 are thus now less predisposed to 2, and are thus more numerous. Given unimodal shocks in each dimension, a general intuition for N-products is suggested. Depending on the exact shape of the distribution and holding other qualities ﬁxed, there is

k some level Q1 where above Q1 is a strategic substitute for a given Qk, and below Q1 is a strategic substitute for Qk. When a product has far superior quality relative to the other option, then quality is strategic complement because the relevant marginal consumers are in the decreasing parts of the multidimensional “hump”. Relative quality decreases brings the margin back to the dense center and implies strategic complements. Otherwise, the margin is in the increasing part of the hump, and relative quality decreases push the margin away

79 from the dense center and implies strategic substitutes. I do not have general conditions on the shape of the taste shock distributions for this “two-toned” comparative static, but it currently appears that the logconcavity of the joint shock distribution is important, which is a common assumption for discrete choice models.6 In Appendix B.1, I explore this intuition by examining a generalization of the derivatives of the N-product case in slightly different notation. To add further concreteness, I now consider what would have happened in certain merger scenarios in a slightly modified version of the toy model. Assume that there are now five options to choose from, two large firms (like AT&T and Verizon) and two small firms (like T-Mobile and Sprint). The large firms have an added, additional component to mean utility

ηBig = 1 to represent exogenous quality diﬀerences that causes the larger ﬁrms to have more market share. This can be things like better phone selection, national coverage or pricing.

A ﬁfth option is the outside good, which has 0 mean utility but also a shock i0. All ﬁrms

1 2 share the convex cost function φ(Qk) = 2 Qk. I assume the shocks are i.i.d. type 1 extreme value which makes demand a multinomial logit. The shock difference joint distribution is unimodal, so the general intuition of the toy model carries through. In addition, it turns out that the second order condition implies strategic substitutes below a constant market share of 50%.7 Given the parametrization, all the big firms have individually 34% market share and the small firms 11%, so all qualities are locally strategic substitutes.

6With shocks on all options, including the outside option, the relevant distributions are convolutions of the shocks with respect to the item in question. Convolutions of logconcave shock remain logconcave by An (1998), and logconcave distribution are unimodal. Type 1 Extreme Value, Normal, Exponential and Uniform are all unimodal distributions.

2 7 ∂ Sk This is due to the cross partial being equal to = −SkSh(1 − 2Sk). As the ﬁrst two terms are ∂Qk∂Qh simply market shares and thus positive, the only ambiguity is in the last term, which is positive only when Sk < 0.5. Since the entire derivative is pre-multiplied by a negative, this means that quality is a strategic 1 substitute below market share of 2 and a strategic complement above.

80 I consider three types of scenarios for the merger. I denote the merging firms 1 and 2 and collectively call them the “insiders”. The non-merging firms I call “outsiders”. I summarize the qualitative effects of these scenarios below, which I work out in detail in Appendix B.2. First, consider when the merged firm discontinues one of the insider’s products entirely after the merger. I denote this scenario by “∗”. Discontinuation might happen if a product has fixed costs associated with it that cannot be justified ex post the merger, such as separate advertising for separate brands. Analytically, dropping a product is analogous to an infinitely large decrease in the mean quality of that product. Since the above comparative statics are general to rival mean quality and not just signal quality, there is a strong incentive to increase signal quality by all remaining firms. Next, consider when the insiders keep all their products and nothing else changes except for the joint control. Denote this scenario and the joint firm as “∗∗”. Joint control causes the insiders to not only care about how much improvements in quality of 1 increases demand for 1, but also how it steals demand from product 2, and vice versa. Thus the incentive for quality provision decreases for both insiders, which in turn leads to higher incentives to provide quality for outsiders due to the strategic substitution. One can also show that strategic substitutes are stronger for the insiders, so it could be the case that the quality level of one of the insiders increases in equilibrium because the incentive to decrease the other insider’s quality is so strong. Finally, consider when in addition to joint control, there are efficiencies from the merger in the form of network integration. That is, if a consumer chooses carrier 1, that consumer can use 100% of the quality of 1’s network, but also some fraction ρ of 2’s network quality. Denote this case and the merged firm by ∗ ∗ ∗. For simplicity, I only consider the case of 100% spillover, ρ = 1, but in principle one could argue it could be less due to incompatibility of handsets with some base stations, since installed technology varies from base station

81 to base station, and from firm to firm.8 The spillover makes each base station effectively cheaper, since each base station can now serve multiple product lines. These are more consumers than would be served by a single product line with equal amount of quality since the multiple brands capture consumer with different horizontal tastes. This effectively lower cost counteracts the lower incentives for quality provision from the internalization carried over from Scenario ∗∗, so overall incentives for quality provision are higher relative to Scenario ∗∗. Because of the strategic substitutability, the outsiders will have an incentive to lower their quality. The spillovers also happen to increase the strategic substitutability of insiders even relative to Scenario ∗∗, because now the firms have to consider how rival quality affects the size of the spillovers. Thus the equilibrium result is even more ambiguous than Scenario ∗∗. Table 2.1 shows the distribution of network quality and the consumer welfare impact under the above merger scenarios. I also report permutations with the size of the insiders for a total of 8 counterfactuals. As shown in McFadden (1978) and Small and Rosen (1981), expected welfare for a consumer in the logit model is the log of the sum of the exponents of mean utility of all the products available:

X ln(1 + exp(δk)) (2.6) k∈K

In general, when a carrier is lost completely, welfare decreases even when network quality of all the remaining firms increases due to the loss of the utility from variety built into the logit. Even keeping the products, if there is no network integration, consumers will be worse off as the internalization of the cannibalization effects causes network quality losses that exceed compensating investment by rival firms. When there are 100% spillovers, the result

8ρ = 0 is simply Scenario ∗∗.

82 Table 2.1: Example Model Results

Variable Carrier Size (1) (2) (3) (4) (5) (6) (7) (8)

Qk % Change 1 Big 5.3 11.3 0.9 0.1 -49.6 -1.4 -0.9 -2.7 (Pre-Merger: 2 Big 5.3 - -17.1 0.1 -49.6 10.9 -0.9 -2.7

QBig = 0.22 3 Small 10.7 43.6 1.6 -13.0 7.2 -2.3 83.4 0.4

QSmall = 0.10) 4 Small - 43.6 -38.4 -13.0 7.2 153.7 83.4 0.4

πk % Change 1 Big 12.7 53.4 1.8 0.3 1.9 -2.7 -1.9 5.9 (Pre-Merger: 2 Big 12.7 - 0.1 0.3 1.9 5.5 -1.9 5.9

πBig = 0.31 3 Small 11.9 51.4 1.7 0.1 8.0 -2.6 7.7 0.4

πSmall = 0.11) 4 Small - 51.4 0.6 0.1 8.0 11.3 7.7 0.4

CS Change (1/100 SDs of k) -8.3 -30.2 -1.3 -0.2 -5.7 1.9 1.3 -0.3

(1) Discontinue Small (Carrier 4) (*)

(2) Discontinue Big (Carrier 2) (*)

(3) Merge Small/Big (Carriers 2 and 4) - No Integration (**)

(4) Merge Small/Small (Carriers 3 and 4) - No Integration (**)

(5) Merge Big/Big (Carriers 1 and 2) - No Integration (**)

(6) Merge Small/Big (Carriers 2 and 4) - Full Integration (***) †

(7) Merge Small/Small (Carriers 3 and 4) - Full Integration (***) †

(8) Merge Big/Big (Carriers 1 and 2) - Full Integration (***) †

† For the merged firms the differences are calculated with respect to the total of both firms.

83 is markedly better for consumers as merging firms increase joint network quality significantly relative when there are no spillovers. However, the benefit depends on whether a merging firm is large or not - smaller firms merging is less harmful since smaller firms contribute less to consumer welfare and have smaller cannibalization effects. When a big firm is involved, these effects are much stronger, so that if the two large firms merge the resulting joint firm reduces its network quality on net since the cannibalization effects are so large. In summary, the only cases with net benefits to consumers are the mergers with spillovers and involving the small carriers. The above results are only here to illustrate the range of possible outcomes and are based on a particular set of parameters. Of the various forces at work, the one that wins out in equilibrium depends on the true parameters. Moreover, while the assumption of unimodal shocks and the comparative statics they lead to seem reasonable, they are fairly restrictive. The results illustrated in Figure 2.1 depend on the unimodal distribution that is dense in the middle of the shock space. Given an arbitrary multimodal distribution of consumers, there would be no strong prediction about the strategic complementarity or substitution of quality. Thus, accurately assessing the merger welfare implications of network quality in the mobile phone industry requires accurate estimation of parameters and a flexible demand system that admits potentially multimodal consumer heterogeneity. I explain how I do this in the context of the cell phone industry in the following sections.

2.3 The Industry Model

In the following subsections, I explain how signal quality is modeled in terms of base stations, the nature of the estimated demand model, and how both these interact with ﬁrms’ incentives to invest in quality. As in Chapter 1, I will treat separate PUMAs as separate markets with quality that is

84 equal to that market’s log density of base stations. This has the added benefit to this analysis since the ultimate aim of this chapter is to run counterfactuals under alternative market structures. I will need a tractable industry game and this in turn requires a simplification for my measure of signal quality. If I make the more realistic assumption that each consumer has a unique area in which they travel and these areas overlap, then a new base station will have an effect on demand for all market areas. A new base station will cause some nearby consumers to switch carriers, and this will change the incentives for carriers to invest in base stations in adjacent areas. Base stations thus change in these adjacent areas, and then they affect their adjacent areas, and so on, until all areas are affected. Carriers would then be playing an oligopoly location game with N-dimensional location strategies, where N is the number of all the possible locations a firm might place a base station. Even with a relatively coarse discretization of locations, this kind of model clearly has multiple equilibria, and thus sharp counterfactual predictions would not be possible.9 Thus I opt to impose Assumption 6 from Chapter 1, under the belief that with the relatively large market areas of PUMAs, cross market interactions are weak, so the equilibrium produced in this model are reasonably close approximations to the multiple equilibria of a more general model. As a result, demand and supply will depend on different quality levels that are specific to PUMAs.

2.3.1 Demand

As in Section 2.2, I assume a static model of consumer utility to model the eﬀect of signal quality on demand. A static model is not ideal given the importance of long-term contracting

9The closest one has come to dealing with this situation is Panle Jia’s analysis of Walmart vs. Kmart store placements (Jia (2008)). The game in Jia’s model is supermodular so she can ﬁnd and characterize an optimal-for-Walmart equilibrium and an optimal-for-Kmart equilibrium. She focuses on these two salient equilibria for counterfactuals. Unfortunately, the supermodularity is conditional on two players so her approach is not applicable in my case.

85 for the US market, but given that my data is relatively thin at the local market level that I study, I am unable to incorporate demand dynamics as does Sinkinson (2014).10 As a result, there may be downward bias in estimated quality sensitivity as some consumers under contract would like to change carriers, but are unwilling to pay the early terminations fees to do so. The overall effect of this is to understate importance of quality to welfare, as the implied demand responses to quality changes would be similar but implied welfare effects would be underestimated. As in the example model, I adopt a discrete choice model of consumer demand. This model is adopted in lieu of the difficulty in estimating the effects of rival quality on demand in Chapter 1. With a discrete choice model, rival indirect utility directly impacts own demand through causing different choices for the same taste shocks. Thus given identification of own effects, the assumptions of the model pin down the rival signal quality effects, which are crucial to my policy experiments. Moreover, the structure of the model also allows other components of quality of carrier’s service to have an effect on quality, which otherwise may be an important omitted variable. The downside of the structural model is that many aspects of the substitution structure are assumed ex ante - the results will be biased and the counterfactuals misleading if substitution is more complicated. Thus I will spend significant time in this section explaining how the model’s substitution works and how the data allows the substitution to be relatively flexible. Formally, index each consumer by i. In each year, t, they have to choose which wireless plan to use, which is a combination of the carrier k and a plan type j. Indirect utility for a

10Sinkinson (2014) defines the market at a multi-county level and has data for the entire United States, so he is able to discretize time into the monthly level. I instead work with markets smaller than the county so to estimate market specific variables I have to aggregate time at the year level. My supply data is also reported at (approximately) yearly level. So while I cover more time than Sinkinson (2014) and the same data source, I only have five periods (years) while he has twenty-six (months).

86 plan jk given a consumer with characteristics Wi in market m and year t is

Uijkmt = (γk + γcity1(city)m)Qkmt + Xjktα(Wi) + Likβ + ηkt + ξkmt + ijkt (2.7) where

Qkmt(Nkmt) = ln(Nkmt/Am) (2.8)

Qkmt is signal quality as deﬁned as the log fraction of the number of market base stations,

Nkmt and the market land area, Am. γk + γcity1(city)m is the consumer sensitivity to the signal quality. γk represents the quality sensitivity which varies by carrier due to their different technologies, network management ability and spectrum holdings. The consumer sensitivity can also vary by a city effect, γcity which is applied if the indicator for a highly urban environment, 1(city)m, is equal to 1. This captures the potential for interference to be greater in these areas due the presence of tall buildings that interfere with signal propagation. Out of the markets, I designate PUMAs 8, 19, 20 and 24 as the “city” markets. Respectively, these PUMAs are downtown Waterbury, Hartford, New Haven and Bridgeport, which are the densest PUMAs by population. PUMA 23, downtown Stamford, would normally qualify as well, but as noted earlier has been merged with PUMA 25 since PUMA 23 bisects PUMA 25. This merged PUMA thus represents both urban and suburban areas, so I do not count it as a city. Relative to the two equation model of Section 1.7.3, I am essentially directly substituting the production function for quality into the demand equation. This is motivated first by the inherent bias in Chapter 1 quality measures because they are selected from people who chose particular carriers and are likely have gotten particular large idiosyncratic taste shocks for those carriers. Second, the first stage regressions of Chapter 1 showed that log base station

87 density is potentially a stronger predictor of variation in market share relative to the quality measures, or at the very least represents an impact on market shares that is not captured by the survey quality measures. Thus I am essentially assuming that even if consumers do not know exactly where the base stations are, they do know the actual signal quality from word of mouth, the internet and ﬁrm advertising.11 To the extent that base station density is a good proxy for quality, this strategy is valid. There are still possibly unobserved components in signal quality which I handle in two

variables. First, ηkt captures all carrier specific characteristics over time, such as changes in phone selection, phone pricing, national coverage, and national advertising, and also transmission technology, spectrum, and network management ability that could affect quality additively. There are still likely some unobserved components of quality that varies by market, year and carrier, either being measurement error or omitted variables that impact base station effectiveness, like unique geography or local advertisement. All this variation is represented by ξkmt. I will later use instrumental variables to control for the endogeneity of this term.

Xjkt are plan-type-carrier-year fixed effects, whose effects vary by consumer characteris-

tics Wi. I choose to use this instead of instead of explicitly using pricing and plan characteristics since these vary little over time and not at all over markets. In particular, I eschew estimating the intensive use of phone minutes in response to the fee structure since this is

12 only possible with usage and speciﬁc plan structure data. Lik is the great-circle distance between the consumer location (in practice their population weighted zip code centroid) and the nearest store that sells a carrier’s plans, which matters as consumers may be more likely

11There are various websites where individuals can post ratings of their quality levels, such cellrecep- tion.com and signalmap.com. More recent sites such as opensignal.com and rootmetrics.com use readings directly from phones using a mobile phone app.

12For an example of what can be done with such data, see Jiang (2013).

88 to buy a plan if they have to travel a shorter distance to initially obtain or service the plan.

β is thus the sensitivity to distance to the nearest store. Inclusion of Lik is potentially important as it might explain geographic variation in carrier selection that might otherwise be attributed to base station placement. If this is correlated with base station construction, this might over estimate signal quality sensitivity. ijkt is an idiosyncratic i.i.d. random shock, which will rationalize consumer adoptions of plans that are lower in deterministic indirect utility. Deﬁne the mean (i.e. deterministic) part of utility as

δijkmt = Uijkmt − ijkmt (2.9)

I assume a type 1 extreme value distribution of the shock. Thus the model is similar to the example in Section 2.2, but there is added heterogeneity in terms of options (prepaid and postpaid) and in consumer characteristics. This formulation yields the familiar logit formula for adoption probability of plan jkt for consumer i:

exp(δijkmt) Sijkmt(δimt) = P P (2.10) k0∈K j0∈J exp(δij0k0mt)

With ex ante identical consumers, this would be almost exactly the same model as Section 2.2 where the only heterogeneity is in the random taste shocks. However, consumers are not identical since I observe their characteristics, and I allow this to affect their utility. Construed broadly, overall taste shocks are now a combination of the logit shock and the consumer- plan fixed effects. Thus the observed heterogeneity allows the model to flexibly accommodate strategic complements and substitutes at arbitrary mean utility levels and shares, since now the overall taste shocks in utility may be multimodal. While the elasticities are functions of market shares within groups of observably identical

89 consumers, they are not over the entire population in a market. Rather, the overall elasticities are a mixture of the group-level elasticities. As is well known, any discrete choice model with independent shocks has an independence of irrelevant alternatives property (IIA) - the rate at which two goods are substituted between each other by the same decision maker is independent of other options. Thus substitution from an option, A, is most strong with the option with the highest probability and implied mean utility, B. This is even though option A may be extremely similar (even identical) to option C. In the context of the logit, this translates into the elasticity of substitution for an individual being completely proportional to a function of the probability of that decision maker choosing each option. For a given population of identical consumers, population elasticity becomes then a function of market shares, but since I differentiate consumer utilities by observed characteristics, this does not hold in my model. Extensions of logit that weaken the IIA property by adding unobservable heterogeneity are possible and widely used in the literature. These extensions would also weaken the two- toned comparative statics from the toy model by adding even more taste heterogeneity. I report later two alternative specifications - a nested logit taking the nests as the plan types, and a random coefficient on quality. Nested logit can be thought of introducing a nest specific shock term that when added to the option-level logit shock creates a nest-level logit shock term.13 Random coefficients, on the other hand, turn one or more of the coefficients on the explanatory variables into a random variable itself. Nested logit is in some sense a variant of random coefficients - the nest specific shock can be thought of as a random coefficient on a nest-specific fixed effect. The idea of both these approaches essentially is to add correlation into the unobserved parts of the utility, so that ex post all shocks the utilities are closer for certain goods.

13See Cardell (1997) for a full treatment.

90 2.3.2 Supply

The industry game assumed for estimation is very similar to the example model in Section 2.2. Each year t, the headquarters of ﬁrm k sets national level prices simultaneously for all their

products, Pjkt. Their engineers then simultaneously set the number of base stations Nkmt at

the market level and the firm incurs marginal costs Fkmt of quality. In this industry, I find this timing more realistic than the usual modeling assumption where quality is changed first, since an individual engineer is unlikely to consider the small price effect his local building decision has on the incentive to change national price levels. There are no adjustment costs in this model, which could be considered unrealistic in this context since there are additional costs in construction when a base station is first installed. Given the limited amount of data it is unfortunately not possible to estimate a fully dynamic model of oligopoly quality investment.14 In a growing market like wireless sunk costs are less important since the option value of waiting is limited. Also, the carriers tend to treat their capital investments in annualized terms - they treat the initial cost as part of that year’s borrowing, and the costs are spread over more than a decade in repayments. Assuming that demand is static, actions year to year do not affect each other, so each year can be thought as an isolated two stage game.15 The lack of dynamics has further benefit since I do not have data for the entire United States. Without national level data I will not be able to simulate equilibria for the pricing aspect of the game. But since the quality setting stage for one year has no effect on later

14Since demand is estimated at the year level, supply can only be estimated at the year level as well. In addition, while some of the data for supply reports dates for base stations to the day, these dates represent the day the base station is reported or approved by the Connecticut state government. Other data is from collection from archives which were collected randomly and do not have exact dates associated with. Given this level of imprecision, my aggregation to the year level seems to be prudent.

15Formally, I am making the assumptions that ﬁrms do not play history dependent strategies. Given my assumptions, repeated plays of the static equilibrium are then an equilibrium for the inﬁnite horizon game.

91 periods, I can examine each year’s quality stage alone taking prices as given.

Let Pjkt be plan speciﬁc prices, Ckt be constant carrier-speciﬁc per consumer costs, and

Nmt be the vector of all base station counts. Deﬁne also the demand Djkmt as the total sum of probability of adoption of a carrier’s plan in a market-year over all consumers. Market proﬁts are equal to markups times demand, or

X πkmt(Nmt) = (Pjkt − Ckt)Djkmt(Nmt) − φk(Nkmt) (2.11) j∈J

As in Section 2.2, this is a normal-form game of complete information with a pure-strategy Nash Equilibrium. The implied necessary condition of the equilibrium is

∂πkmt(Nmt) X ∂Djk(Nmt) ∂φkmt(Nkmt) = (Pjkmt − Ckt) − = 0 (2.12) ∂Nkmt ∂Nkmt ∂Nkmt j∈J

As in the example model of Section 2.2, the cross partial of demand still determines the monotone comparative statics of the model. These are explicitly derived in Appendix B.3, but in short, the model without any heterogeneity in consumers would be almost exactly the same as the model in Section 2.2 and would also have strategic substitutes for all the market structures observed in the data. The consumer heterogeneity does allow for strategic complements though, but this is dependent on having high enough amounts of consumer heterogeneity such that ﬁrms have very high market shares for particular segments of the population. Thus the comparative statics depend on the heterogeneity parameters estimated in the demand system.16

16Random coefficients or nested logit specifications could also introduce strategic complements since these segment markets by consumers with unobserved variation in tastes for particular goods based on either their coefficient draws or by nests, respectively. As I will present later, random coefficient and nested logit versions of the model do not have very different results from the pure logit model with heterogeneous effects, implying that the heterogeneous effects explain almost all of the variation.

92 2.3.3 Caveats

In addition to the previously mentioned limitations, the model has several other shortcomings. First and foremost, the model only allows adjustment to signal quality and price. All other characteristics are absorbed in the fixed effects. Average price will be later adjusted for the counterfactuals, but I will have to hold fixed all other characteristics such as phone selection, customer service levels, network management, transmission technology and spectrum. These features of service are generally hard to change in the short term. Phone selection is determined by long-term contracts between the carriers and suppliers. Improvements in customer service require retraining of employees and changes in management practices. Transmission technology changes require complete changes to network architecture and replacement of equipment and software. Network management depends on not only on software and implemented algorithms, but like customer service would also require changes to the managers and their strategies. Spectrum increases only come with the FCC auctions, though there are rare sales or swaps of spectrum by the carriers. Thus counterfactual analysis could be considered reasonably close to the short term outcome of a merger, though the long term outcome could be considerably different. How depends on consumer demand for these other characteristics, and in particular whether these features are strategic substitutes or complements. While I have very little knowledge of how most of these characteristics would evolve, in the years since the end of the sample the nature of transmission technology has begun to change. While the carriers in the United States are separated between GSM and CDMA carriers, all carriers are moving towards versions of LTE technology, which are more or less an improved CDMA standard. Given CDMA has some quality advantages relative to GSM, this would imply that the GSM

93 carriers would improve relative to their CDMA rivals. This would result in the predictions of the counterfactual being distorted relative to what would happen in a merger within next few years.17 The one feature that would be potentially problematic for the short-term assumption is plan structure, which can be changed more or less instantly. Due to the 2-year length of contracts, incumbent consumers would retain their old plans for some time, but new plans could be introduced with a much more generous structure. This in fact happened in the last few years since the end of the sample period, as the industry has entered into essentially a price war.18 The industry has moved towards unlimited data plans and away from contract plans to compete with each other, which suggests that plan structure is a strategic complement. This implies that my later counterfactuals may understate the welfare losses from a merger, as the reduction of competition would decrease the incentive to provide more generous plans. Another caveat is the assumption of the PUMA as the market deﬁnition may be too small. If consumers care about signal quality in other PUMAs, then signal quality sensitivity

17 The distortions depend on how transmission technology enters into the utility of consumers, though in general overall welfare changes are ambiguous. The model is formally agnostic about this since the effect could either be in the carrier specific signal quality coefficients or the carrier fixed effects. On the one hand, if the move to LTE makes each base station more efficient (a higher γk), then the diversion ratio between CDMA and GSM carriers would become more favorable towards the GSM carriers. As I will show later, the diversion ratios between Verizon, the main CDMA carrier, are already more favorable to its GSM rivals, so the technology change would exacerbate this difference. The effect on everyone else is a bit more ambiguous since the bias of the diversion ratios is not always in favor of the other firm, so the overall effect is hard to sign. On the other hand, if LTE results in a level shift in signal quality irrespective of base stations (a higher ηkt), then the level of substitution increases overall and diversion ratios become less unequal. This makes the changes of the merger stronger as now there is greater incentive to change signal quality as now signal quality steals more share. I will later find signal quality a strategic substitute so merger-induced reductions in quality between firms would be compensated by greater increases by non-merging rivals. Within merging firms, the impact is ambiguous as mergers will result in greater incentive to reduce quality but less incentive to degrade one carrier by a lot since diversion ratios are now more equal.

18See La Monica, Paul R. “Wireless war: Consumers win, investors lose.” CNN Money December 12, 2014. http://money.cnn.com/2014/12/12/investing/wireless-carrier-price-wars/

94 would be under-attributed to local quality changes by the model, since such signal quality is negatively correlated with own market signal quality.19 Then the counterfactuals would be biased by giving too little new demand for increases in signal quality, though the sign of the bias is hard to tell since the model also implies the cost of base stations by the underestimated signal quality sensitivity. The estimated costs would be decreased so that would somewhat cancel out the decreased incentive to provide quality. Finally, the wireless market is subjects to major disruptive changes. A recent example is the iPhone, which made data-intensive smart phones the new standard handset consumers expect. Aside from revolutionary changes in demand that would invalidate the entire estimated demand system, carrier expectations are important and may be an important consideration when deploying base stations. For example, carriers may expect one day a major increase in traﬃc could occur which would require many more base stations. Thus the marginal beneﬁt from base stations is higher than what I estimate, and thus marginal costs are lower that what I estimate, as well. Unfortunately, I have no data to estimate carriers’ expectations of this margin, so I must assume that such concerns are minor.

2.4 Data

To estimate my model, I use the following subset of data from Chapter 1. I use the Nielsen Mobile Insights Survey for Connecticut respondents from 2008-2012, which as noted in Chapter 1 is relatively well balanced in comparison to the actual population of Connecticut.20 I will use the plan type, zip code and consumer demographics of the 17,235 respondents. Given the data I have, I will simplify and say each carrier oﬀers one of two

19Log base station density has correlation of -0.228 with the average of log base station density in adjacent PUMAs.

20See Table 1.4.

95 composite plans, prepaid or postpaid. I use income, household size, age and gender in the estimation as they are likely to be especially important for taste variation in cell phone use. Income is likely to affect price sensitivity; household size will proxy for the value of family plans that are very popular options; age will proxy for the affinity for new technology; and sex might capture variation in calling patterns across genders. I use the base station data from the two Connecticut Siting Council datasets, but as the demand dataset only goes from 2008-2012 I will limit myself to those years for the base station data as well. Using the PUMA, I will match demographic information from the PUMS as needed. Additional information I use in this chapter that I did not use in Chapter 1 is store location information taken from ReferenceUSA.21 In autumn 2013, I recorded the locations of all stores in Connecticut that contained “cellular” or “mobile telephone” in their Standard Industrial Classification (SIC) title. I further hand cleaned this list and determined carrier selection via web searches when possible. Clearly, this measure is imperfect since I am including only store locations from after my sample period - there will be stores I include that will not have opened yet and some currently closed stores that were active I cannot include. I also use the zoning code data from Connecticut towns reported in the Chapter 1. I use this as an instrument, so I defer explanation of its use to the next section on estimation.

21I declined to feature this in Chapter 1, as there was not much variation in general here across carriers: they all located stores in the urban center of the state. With a more a disaggregated approach in this chapter, this may be more important though.

96 2.5 Estimation and Results

2.5.1 Endogeneity of Quality

Typically economists worry about the endogeneity of price in demand estimation due to unobserved demand shocks. In the application of wireless telephony that is less of a concern because pricing is done at the national level. As noted earlier, I eschew estimating price elasticity directly and absorb all the corresponding variation in ﬁxed eﬀects. Instead, there is a need to correct for the endogeneity of signal quality to the unobserved component of demand. Formally:

E[Qkmtξkmt] 6= 0 (2.13)

That is, since base station placement is endogenous the carriers may have placed base stations according to some unobserved components of demand. For example, some areas might have especially high interference due to unique geography or buildings conﬁgurations. Thus a carrier might place more base stations to yield to same signal quality, thus biasing the estimates of quality sensitivity downwards. Alternatively, a ﬁrm might decide to advertise new base station deployment in a market - which would boost demand but be confounded with the increase in base stations, biasing the quality sensitivity upwards. Traditional methods for dealing with endogeneity in demand systems employ instruments that are infeasible in this setting. Berry, Levinsohn, and Pakes (1995) use product characteristics of rival products to instrument for price, under the rationale that attractiveness of rival products would shift demand for the good in question. In that context, product characteristics were assumed exogenous due to the long product development cycles in auto- mobiles. Alternatively, Hausman, Leonard, and Zona (1994) and Hausman (1996) use prices

97 in other regional cities as instruments for price, citing some unobserved regional component of a firm’s costs common to all markets. Neither of these can be implemented here since the product assortment and pricing for all markets across all times is the same. Quality does vary by market, but the carrier-year fixed effects use all the variation that could be attributed to the Hausman-style instruments. Instead, I use a cost side instrument that would influence a firm’s incentives to build base stations: the fraction of a town’s zoning regulations that are telecommunications related. As explained in Chapter 1, industry sources note the primary difficulty with siting is the cost and delay in proceedings with local zoning authorities, which is greatly hampered by long and ambiguously-worded statutes. If a town devotes more space to telecommunications facilities then they must be more worried about it relative to other kinds of zoning. I use the ratio of the number of characters used rather than just the characters in the telecom sections since this corrects for the fact that some towns might simply have longer, wordier regulations. Since there are multiple towns in a PUMA, I use the population-weighted average. With the firm and city interactions, I need four additional instruments. I therefore also use the interaction of regulation with firm and the city dummy. Regulation has a major drawback as an instrument in that it does not vary by firm, but only by market. Regulations further do not vary by year since I collected them over the period of 2012-2013, and thus the regulations reflect current law. However, there does not seem to have been radical changes in the telecom sections of the zoning codes as many zoning codes include references to amendments and their dates. Also the use of a weighted average mitigates any potential unobserved change by a particular city.

98 2.5.2 Demand Estimation Procedure

Even with instruments, dealing with the endogeneity is not straightforward. ξkmt cannot be estimated as a fixed effect because it is not separately identified from quality. The typical procedures for handling endogeneity in demand estimation, introduced in Berry, Levinsohn, and Pakes (1995), requires population market shares. The sample is not large enough for me to confidently use the shares found therein as proxies for population shares. I have 17,235 survey responses, which collapsed to the 480 carrier-market-years would have too much noise to be used in this way. For example, some markets are as small as 29 individuals in a year.22 I instead adapt a suggestion made as an aside in Berry (1994) and most prominently applied in Goolsbee and Petrin (2004), in which fixed effects soak up all the variation at the carrier- market-year level in a first step, and then covariates of interests are regressed on these fixed effects in a second step.23 The second step allows for linear instrumental variables regression since the endogenous error terms enter linearly into the fixed effects.24 Define the variable that absorbs all carrier-market-year-variation as

ζkmt = (γk + γcity1(city)m)Qkmt + ηkt + ξkmt (2.14)

22The sample size by market-year varies from 29 for New Haven in 2010 to 331 in 2008 for the Windsor Locks area.

23Technically, Goolsbee and Petrin (2004) do use the procedure in Berry, Levinsohn, and Pakes (1995), which takes the observed markets shares as given to imply unique values for the fixed effects. Like me, however, they break their estimation into two parts, and do not simultaneously estimate the parameters of the endogenous variables, as in Berry, Levinsohn, and Pakes (1995). They also note that they could have estimated the fixed effects rather than use the procedure in Berry, Levinsohn, and Pakes (1995), but simply chose not to, presumably for computational concerns.

24An alternative would be to use a control function, as in Petrin and Train (2010), though I decline to do so due to the strong assumption of the independence of the instruments with ξkmt, rather than just no correlation.

99 so

δijkmt = ζkmt + Liktβ + Xjkmtα(Wi) (2.15)

I then conduct maximum likelihood over the observed individual choice probabilities by solving the following objective function:

X arg max ln(Sijkmt(θ|Qkmt,Am,Likt,Xkmt,Wi)) (2.16) θ={ζkmt,γk,γcity,β,α} i∈I

In practice Xjkmtα(Wi) is simply diﬀerent for every plan-type, carrier, year and characteristic

combination so I estimate a corresponding ﬁxed eﬀect. Once ζkmt is recovered, I estimate γk

via instrumental variables using (2.14) as the estimating equation and Zkmt as instruments.

While there is error in the measurement of ζkmt, the linear form allows that error to be

absorbed into ξkmt. I weight using the standard errors for ζkmt from the maximum likelihood step for efficiency reasons.25 Identification in this model depends on variation in choices over the different markets and time. Identification of the quality sensitivity terms are identified across markets within a carrier-year, as I employ carrier year effects. The distance terms are identified from variation across markets and across individuals within markets, as distance is zip code specific. The product-demographic specific terms are identified from the relative share of products in the sample for that demographic across markets, and brand-year-market effects are identified from the shares for that brand in that market year. For comparison, I also present results from the nested logit specification and a random

25I could do the maximum likelihood and the linear steps in a single-step GMM procedure in which the moments are the score of the maximum likelihood and the exogeneity conditions of the instruments. This would be similar to Berry, Levinsohn, and Pakes (2004), although in that paper they also exploit data on second-best choices.

100 coefficients specification. The nested logit uses the plan types - none, prepaid and postpaid - as nests and estimates a single dissimilarity parameter, λ, which approaches 1 as the model approaches pure logit. Plan type was used because prepaid customers might be different from postpaid customers on some unobservables since they prefer a plan that they can make cheaper on average through lower utilization. In addition, postpaid plans require a credit check which some consumers might not pass. This means that different carriers are not directly dissimilar since every carrier (except Sprint) has a product in prepaid and postpaid, but since T-Mobile has much more successful prepaid product than AT&T or Verizon, the overall substitution between the firms should differ. For the random coefficients specification, I assume no nesting and that the quality sensitivity coefficient is distributed normally with some standard deviation σ. With normally distributed quality sensitivity, the mean of the coefficient is additively separable and is absorbed into ζkmt. Thus the quality sensitivity can still be instrumented for in the second step and while the standard distribution of the distribution can be recovered from the first step by integrating over an interaction between the quality and a random variable with the standard normal distribution. Rather than simulate, I use numerical Gauss-Legendre quadrature on 15 points, which is both computationally simpler and more accurate than simulation. The high accuracy of quadrature obviates the need for correction of the standard errors due to simulation error.

2.5.3 Results: Individual Identiﬁed Demand Parameters

In the MLE step, I estimate ζkmt along with all parameters that capture variation at the consumer level. This includes the plan-consumer-characteristics-year eﬀects and the β store distance sensitivity term. For the nested and random coeﬃcient terms, λ and σ are also presented, respectively.

101 Table 2.2: Individual Level Identiﬁed Coeﬃcients from MLE

Pure Logit MLE Nested Logit MLE † RC Logit MLE

β (KM to Store) -0.003 -0.003 -0.003 (0.005) (0.005) (0.005) λ( Nest Parameter) 1.09 (0.13) σ(S.D. of Rand. Co.) 0.08 (0.16) Observations 17,235 17,235 17,235 Log Likelihood 29,607 29,607 29,607 McFadden’s Pseudo-R2 0.254 0.254 0.254

***, **, * indicate 1%, 5% and 10% signiﬁcance, respectively. † Nested logit estimates are divided by λ for comparison with other speciﬁcations.

102 The results from all three specifications are nearly identical, with a McFadden’s Pseudo- R2 of 0.25. There is in fact only a difference of less than 0.2 log points between any of the three models, and clearly a likelihood ratio test fails to distinguish between them. For the nested logit specification, λ is 1.09, which is generally not consistent with utility maximization.26 However, the value is not significantly different from 1, reflecting the already established fact that the nested logit model does not explain any further variation than the pure logit model. The standard deviation of the random coefficient estimated is also relatively small at 0.08 and not significant. A formal test of the multinomial logit from Hausman and McFadden (1984), in which the model is re-estimated on data without one of the options to see if the estimate parameters are the same, is presented in Table B.1. The test statistics suggest no evidence that the true model is not multinomial logit.27 One can think of the nesting and random coefficients as adding more ex post heterogeneity to the logit - observationally identical populations of consumers have different distributions of ex post utilities for options which will cause their substitution to differ from the markets shares of the total population. Given consumer characteristics, product and year specific fixed effects, there is not much variance that this unobserved heterogeneity can explain. Moreover, there are not many options (10 in total), which further lessens the available variation. Given the lack of difference between the models, and that there is substantial heterogeneity in utility given by the fixed effects that would break the IIA property at the market level, I continue with the pure logit with individual level heterogeneity as my preferred specification.

26I say “generally” since B¨orsch-Supan (1990) shows that dissimilarity parameters greater than 1 may be possible in a utility maximization framework given certain values of the covariates.

27Three out of ﬁve of the test statistics are actually negative, which is odd given these are Chi-square tests. However, in practice this test statistic is often negative, as in the original application of Hausman and McFadden (1984). They took negative values to be a sign of no evidence of a diﬀerence between the multinomial logit and the true model, and I follow that in line with the subsequent literature.

103 Table 2.3: Prepaid-Carrier-Year Fixed Eﬀects from Pure Logit

AT&T T-Mobile Verizon Other

2008 -2.23*** -1.26*** -3.42*** 2.76*** (0.30) (0.39) (0.46) (0.61) 2009 -2.49*** -1.48*** -3.07*** 4.35*** (0.29) (0.36) (0.35) (0.71) 2010 -2.25*** -2.44*** -2.48*** 3.70*** (0.29) (0.61) (0.29) (0.60) 2011 -2.19*** -0.80** -2.87*** 1.92*** (0.25) (0.36) (0.35) (0.36) 2012 -1.88*** -0.87** -2.61*** 2.56*** (0.22) (0.34) (0.31) (0.41)

***, **, * indicate 1%, 5% and 10% signiﬁcance, respectively. The above represents the diﬀerence in mean utility of prepaid plans relative to postpaid plans, for women between the ages of 35 and 64, in multiple-person house-

holds(families) that earn from $50-75 thousand annually.

In all specifications β is negative as one would expect, though it is very small and not significant. The implied own elasticity of travel distance to stores is essentially zero, so I ignore this aspect of demand in the subsequent analysis. The estimates of the plan-type, carrier, year and characteristics effects are too numerous to report completely, so I will report them in part. The year specific product effects for the prepaid products are in Table 2.3. These represent the difference in utility from prepaid products relative to postpaid products for every year and carrier. These results imply that

104 the value of prepaid products generally grew over the sample period for all carriers. The results also show that prepaid is an inferior product relative to postpaid - except for Other which has large positive estimates revealing that its prepaid plans are actually significantly better than its postpaid ones. The remaining coefficients are the difference in plan utility relative to that of a 35 to

65 year old woman in a multi-person household that earns between $50 and $75 thousand a year. Instead of reporting all 444 remaining estimates and standard errors, I report in Table 2.4 the mean and in parentheses how often the estimate was estimated to be different from 0 with 95% significance. About 37% of these estimates are significant at the 95% level. Prepaid value effects are estimated with less accuracy in general given the lower number of prepaid purchases. In general, the estimated effects are quite small and do not vary strongly across firms. The same trends are true for most of the firms - the poor, seniors and those living alone have less value for phone service. In particular, seniors seem to prefer postpaid plans in the Other category more than other demographics. Also, the value of prepaid plans of the Other composite brand is actually greater for poor individuals.

2.5.4 Results: Quality Sensitivity Parameters and Brand-Year Ef-

fects

Given the MLE results for the preferred specification, linear instrumental variable estimation can proceed with the carrier-market-year fixed effects. Table 2.5 examines the strength of the regulation instrument. Weighted regression of the instrument on Qkmt is very significant, and remains so on subsamples divided by the different carriers and for the city markets only. The weights used are the same for instrumental variables regression itself, the estimated variance of the carrier-market-year fixed effects. Using the multiple-endogenous F-Statistic suggested in Angrist and Pischke (2009, p. 217-218) yields very large values that are greatly

105 Table 2.4: Mean Plan-Type-Carrier-Consumer Characteristic Eﬀects over Years

Postpaid Prepaid

AT&T Sprint T-Mobile Verizon Other AT&T T-Mobile Verizon Other

Less than $25K HHI -0.98 -0.55 -0.65 -1.44 0.36 -0.61 -0.47 -0.47 0.10 (5/5) (3/5) (4/5) (5/5) (1/5) (2/5) (1/5) (1/5) (1/5)

$25k-50k HHI -0.56 0.04 -0.38 -0.72 -0.11 -0.78 -0.47 -0.36 0.04 (3/5) (0/5) (1/5) (4/5) (0/5) (1/5) (0/5) (1/5) (0/5)

$75k-100k HHI 0.36 0.45 -0.20 0.52 0.81 0.35 -0.10 0.61 -0.05 (1/5) (1/5) (0/5) (2/5) (1/4) (1/5) (0/5) (0/5) (2/5)

$100K+ HHI 0.65 0.91 0.29 0.83 0.42 0.10 0.22 0.11 -0.44 (3/5) (4/5) (2/5) (3/5) (0/4) (0/5) (0/5) (0/5) (2/5)

Declined to Report Income -0.44 -0.37 -0.57 -0.29 -0.20 -0.24 -0.28 -0.39 -0.07

(2/5) (1/5) (3/5) (2/5) (0/4) (0/5) (0/5) (0/5) (0/5)

Single -0.57 -0.62 -0.83 -0.43 -0.11 -0.84 -0.34 -0.54 -0.44

(5/5) (4/5) (5/5) (4/5) (0/5) (5/5) (0/5) (1/5) (4/5)

Minor -1.01 -0.72 -0.80 -1.08 0.40 0.01 -0.49 -0.06 -0.78

(4/5) (2/5) (3/5) (5/5) (0/2) (0/5) (1/5) (1/5) (3/5)

Between 17 and 35 Years Old 0.44 0.80 0.72 0.60 0.43 0.31 0.23 0.25 -0.21

(3/5) (5/5) (3/5) (4/5) (1/5) (1/5) (1/5) (0/5) (1/5)

More than 65 Years Old -0.66 -0.83 -0.92 -0.55 0.50 -0.12 -0.68 -0.73 -0.21

(5/5) (5/5) (4/5) (5/5) (0/5) (1/5) (3/5) (2/5) (1/5)

Male -0.14 -0.21 -0.37 -0.32 -0.22 0.08 -0.21 -0.11 -0.17

(1/5) (1/5) (2/5) (3/5) (2/5) (0/5) (1/5) (0/5) (0/5)

The number of estimates at 95% significance over total years estimated listed in parenthesis. Total years sometimes less than five since some year no one of that demographic chose that option - effect then assumed to be zero.

106 Table 2.5: Instrument Strength

Weighted OLS Regression

Dependent Variable: Qkmt Full Sample Just Sprint Just T-Mobile Just Verizon Just City

% Telecom Regulations -37.70*** -38.04*** -47.19*** -32.00*** -17.28***

Brand-Year Eﬀects? Yes Yes

Year Eﬀects? Yes Yes Yes

R2 0.54 0.53 0.53 .46 0.23

Observations 478 120 118 120 80

Testing Identiﬁcation for Each Interaction

Qkmt interacted with Constant Sprint T-Mobile Verizon City

Multivariate F-Stat 106.82 106.84 123.58 75.25 89.64

***, **, * indicate 1%, 5% and 10% signiﬁcance, respectively.

Weights from pure logit specification. above the rule-of-thumb value of 10 suggested by Stock, Wright, and Yogo (2002). Estimates from both an instrumented regression and an un-instrumented regression demonstrate significant differences between the firms in quality sensitivity. Relative to AT&T, T-Mobile and Sprint have significantly higher quality sensitivity and Verizon a negative one. The city effect on quality sensitivity, as expected, is significantly negative. In- strumenting matters: the baseline sensitivity γAT &T of 0.15 doubles to 0.30, while the other effects all become more positive. The downward bias correction is especially important for the Verizon parameter, since weighted OLS implies it is slightly negative. Though not significant, this would imply improvements in Verizon signal quality reduce market share. In the case of cities, the total coefficient would be even more negative. After instrumenting, this is no longer the case: the net effect of Verizon quality becomes 0.16 outside cities and 0.02 within. Nevertheless, Verizon ends up being problematic - the sensitivity in either

107 Table 2.6: Signal Quality Sensitivity Estimates

(1) (2) OLS IV

γAT &T 0.15*** 0.30*** (0.05) (0.08)

γSprint − γAT &T 0.20*** 0.24** (0.07) (0.09)

γT −Mobile − γAT &T 0.38*** 0.41*** (0.06) (0.09)

γV erizon − γAT &T -0.17* -0.14 (0.09) (0.16)

γCity -0.05*** -0.13*** (0.02) (0.03)

Endogeneity Test 10.98*

Carrier-Year Eﬀects? Yes Observations 478†

***, **, * indicate 1%, 5% and 10% significance, respectively. † Five markets-years had no observations for any carriers; two of these times were Sprint, two were Other and once was None. In those cases, a carrier-market-year fixed effect could not be estimated, so the second stage regression lacks 2 of the 480 observations that would be potentially possible. The Endogeneity Test is the difference between the Sargan-Hansen statistics of the exogenous and endogenous values.

108 case is not significantly different from zero nor is it significantly different from AT&T. In contrast, marginal effects from signal quality can be differentiated between AT&T, Sprint and T-Mobile both with and without instrumenting. Estimates from both an instrumented regression and an un-instrumented regression demonstrate significant differences between the firms in quality sensitivity. Relative to AT&T, T-Mobile and Sprint have significantly higher quality sensitivity and Verizon a negative one. The city effect on quality sensitivity,

as expected, is significantly negative. Instrumenting matters: the baseline sensitivity γAT &T of 0.15 doubles to 0.30, while the other effects all become more positive. The downward bias correction is especially important for the Verizon parameter, since weighted OLS implies it is slightly negative. Though not significant, this would imply improvements in Verizon signal quality reduce market share. In the case of cities, the total coefficient would be even more negative. After instrumenting, this is no longer the case: the net effect of Verizon quality becomes 0.16 outside cities and 0.02 within. Nevertheless, Verizon ends up being problematic - the sensitivity in either case is not significantly different from zero nor is it significantly different from AT&T. In contrast, marginal effects from signal quality can be differentiated between AT&T, Sprint and T-Mobile both with and without instrumenting. The results have a curious implication - the firms with higher market share, AT&T and Verizon, have less marginally productive base stations than the smaller carriers, T-Mobile and Sprint. This compounds the earlier puzzle that Verizon has the least base station density on average even though it is the leader by share in most markets in the sample. This apparent contradiction can be resolved by recognizing that a more concave production function of signal quality in base station density will have a smaller γk but a higher intercept ηkt. As noted earlier Verizon and AT&T have more and better spectrum than their rivals, and as early entrants may have access to better site locations compared to Sprint and T-Mobile. Thus they may have reached the flat part of their production functions much earlier than Sprint and T-Mobile, resulting in the observed relationship between base station density and

109 Figure 2.2: Mean Quality on Base Station Density - Instrumented Pure Logit 2

0 AT&T Sprint Utility T-Mobile Verizon −1 AT&T-City Sprint-City T-Mobile-City Verizon-City −2 0 0.1 0.2 0.3 0.4 0.5 0.6 Base Stations per KM2 market level quality to be very ﬂat relative to their rivals. To illustrate this idea, Figure 2.2 plots the implied mean quality on base station density for the year 2012. The curves are equal to

(γk + γcity1(city)m)ln(Nkm2012/Am) + ηk2012 (2.17)

As you can see, while the plots are quite ﬂat for AT&T and Verizon for most of this range, the average level is much higher than Sprint or T-Mobile. Thus an average AT&T and Verizon base station is estimated more productive than a T-Mobile or Sprint base station;

110 it is just that this benefit is very front-loaded in the former case.28 To examine the economic magnitudes, I calculate the percent change in demand given a 1% change in log base station density, the elasticity with respect to the quality proxy, using the observed number of base stations, estimated demand implied from the model, and the analytical derivative of the estimated demand. This estimated demand for a given brand is the sum of the probabilities of adoption of that brand for each individual in the population for the appropriate year and PUMA. To approximate these populations I use the PUMS responses for these years and PUMAs and the given weights.29 The derivative of demand is simply the derivative of the probabilities, summed over all the individuals. I report medians of demand elasticities of log base station density across markets and years in Table 2.7, in terms of what a 1% change in log base station density, the signal quality proxy, would do to market share.30 A 1% change Verizon’s signal quality has the smallest own-effect, giving Verizon 0.13% more of the market for the median market-year. AT&T has the highest effect (0.25%) while the effect is intermediate for Sprint (0.16%) and T-Mobile (0.18%). That AT&T is the highest despite its relatively low quality sensitivity parameter suggests the importance of its other quality dimensions, such as spectrum and

28 A difficulty with this interpretation is that the carrier effect also contains all other carrier specific differences, so the estimated intercept of this function is not separately identified from things like phone selection, pricing, spectrum holdings or even branding. So while I can interpret some of the difference in fixed effects as differences in the production function, some of the difference is certainly due to other aspects of the carriers. Thus the true signal quality gap between the carriers is probably not as dramatic as displayed in Figure 2.2.

29For 2012, the PUMS uses new PUMA border deﬁnitions that are not consistent with the 2000 PUMAs deﬁnitions used in the rest of this paper. To compensate, I create a new PUMS pseudo-population by taking the distribution of characteristics seen in the 2011 data and scaling up all weights so the implied total population is the same the one reported in the 2012 PUMS. Also, I need to assign zip codes to each consumer so I can calculate their store distances, but this is not reported in the annual PUMS data. I assume the distribution of consumers characteristics are invariant within a PUMA, and then assign the PUMA population according to the proportions in the 2010 Decennial Census data.

30The median 1% change in the proxy is equivalent about a 1.2 base stations change over all the market years.

111 Table 2.7: Median Quality Elasticities for Instrumented Pure Logit Speciﬁcation

1% Change in Signal Quality Proxy of... Full Sample AT&T Sprint T-Mobile Verizon

...Results AT&T 0.25 -0.06 -0.06 -0.06 in Change Sprint -0.03 0.16 -0.02 -0.02 of Market T-Mobile -0.03 -0.01 0.18 -0.01 Share % of... Verizon -0.13 -0.06 -0.06 0.13

1% Change in Signal Quality Proxy of... 2012 AT&T Sprint T-Mobile Verizon

...Results AT&T 0.27 -0.05 -0.04 -0.06 in Change Sprint -0.03 0.15 -0.01 -0.01 of Market T-Mobile -0.02 -0.01 0.16 -0.01 Share % of... Verizon -0.12 -0.06 -0.06 0.13

The matrices do not represent any particular market. Rather, each entry is the median across market-years for that particular ﬁrm.

112 the exclusivity of the iPhone. As a result of the relative unimportance of the consumer characteristics-based heterogeneity, consumers are estimated to largely agree about carriers’ mean utilities. Thus, people who selected options providing less mean utility, like Sprint and T-Mobile plans, must have gotten a large idiosyncratic taste shock for that option. As a result, substitution away from those firms is low when rivals increase quality - these vary between -0.1% to -0.3% for the median market year. In contrast, consumers of AT&T and Verizon are not particularly loyal since most consumers chose them because of their high mean utility - if another option increases in mean utility they are more likely to switch. Thus substitution from AT&T is on average -0.06% of the market and from Verizon between -0.13% to -0.06%. In terms of diversion ratios, AT&T and Verizon are very substitutable with their rivals so have the most share poached from quality gains: AT&T and Verizon steal half of their gain from each other when they improve quality, while Sprint and T-Mobile steal about a third from AT&T and T-Mobile, each. Notably, Sprint and T-Mobile are not very good substitutes - they steal only about 10% of their gain from each other. In any case, the cross effects of quality are relatively small, with no effect in any market year exceeding -0.27% market share per base station.31 Finally, the estimates imply that base stations are strategic substitutes, since there seems to be no dimension in which a firm dominates the market, i.e. even conditional on consumer characteristics, market share is never greater than 0.5. As mentioned earlier, characteristics do not seem to add a great deal of heterogeneity into tastes, so carriers do not split up the market into segments which they individually dominate.32 While there is variation in market

31Interestingly, a combination of changes in preferences for products and increased quality overall results in elasticities going down for Sprint and T-Mobile over the sample period. See Table 2.7.

32One can note Table B.3, which shows correlation across products in the value of mean utility in the sample attributable to consumer demographics and year only.

113 specific factors like base stations across firms, firms only have more than 50% market share three times in the Nielsen sample. Using the estimates the predicted market share for the pseudo-population of PUMS respondents, 99% of the implied probabilities of adoption for each option by each individual are below 0.5. Thus the aggregate cross partial of demand with respect to signal quality is negative, implying strategic substitutes.

2.5.5 Supply Side Estimation and Results

I back out a cost for each ﬁrm in each market-year for each carrier using the ﬁrst order condition:

X ∂Dkmt(Nmt) ∂φkmt(Nkmt) (P − C ) = (2.18) jkt kt ∂N ∂N j kmt kmt

Calculation of the left hand side of this equation (the marginal variable proﬁt) requires prices and marginal quantity costs to be known or estimated. Selecting a price to use for

Pjkt is problematic in my application since I aggregate over many products and products have a usage aspect which means each individual could potentially pay a diﬀerent eﬀective price. Instead of arbitrarily selecting a price for a particular focal plan, as Sinkinson (2014) does when he uses the introductory smart phone plan fee, I use the ARPU reported in the UBS’s analysis of the US wireless industry, as I did in Chapter 1. ARPU is the main revenue measure used by industry participants. It is also the variable used by the BLS to construct their price index for cellular phone service. Given that I need Pjkt purely as a marginal revenue number, ARPU seems like a sensible proxy.

Marginal consumer costs Cjkt would usually be estimated via a price ﬁrst-order condition, but I do not have the nationwide data to do so. In principle, I could estimate it as a free

∂Dkmt(Nmt) parameter for the ﬁrst order condition of base stations, but in practice the term Pjkt ∂Qkmt is nearly collinear with ∂Dkmt(Nmt) . I instead take the “Costs of Wireless Service” reported ∂Nkmt

114 Table 2.8: Marginal Monthly Base Station Cost - Fkmt ($1000)

Full Sample

Carrier Mean SD Min 25pct Median 75pct Max

AT&T 11.8 3.7 4.8 8.8 11.6 11.6 22.3 Sprint 8.6 3.6 2.4 5.5 8.2 11.0 21.1 T-Mobile 8.5 5.5 0.2 4.5 7.4 11.2 29.5 Verizon 7.1 3.3 0.8 4.6 7.0 9.2 15.1

2012

Carrier Mean SD Min 25pct Median 75pct Max

AT&T 12.5 3.6 6.5 9.3 12.4 15.4 18.5 Sprint 9.3 4.5 3.2 5.5 8.0 11.7 21.1 T-Mobile 6.9 3.5 0.8 4.3 6.4 8.9 16.6 Verizon 6.9 3.2 0.9 5.0 6.6 8.9 14.0

by the UBS analysis, and divide by the total number of consumers. This implicitly assumes cost is constant across markets, which seems reasonable given carriers do not oﬀer specialized plans or phones by market.

The median base station cost per month is $8,147, which is about twice as much the numbers assumed by the engineering paper by Claussen, Ho, and Samuel (2008) or Björkegren (2013). Those estimates were based on pecuniary costs alone and did not include costs that come from regulatory costs from negotiating with towns or meeting particular zoning codes. This implies that non-pecuniary costs, i.e. delays caused by regulatory proceedings or negotiations, are significant drivers of economic cost. Costs do appear different by firm, with the median across all markets and years being

$11,605, $8,230, $7,354 and $6,970, for AT&T, Sprint, T-Mobile and Verizon, respectively. Variance overall is quite high, with a standard error of $5,263 across all market-years and

carriers. Given the very high variance, I elect to use the entire term Fkmt as the cost, rather

115 Table 2.9: Total CT Costs as % of Variable Proﬁt by Year

Year AT&T Sprint T-Mobile Verizon

2008 19.26 44.76 58.90 8.90 2009 18.85 45.56 59.20 8.93 2010 19.05 46.45 59.87 9.28 2011 18.67 44.86 61.71 8.81 2012 18.96 45.57 61.87 9.19 a mean over year or markets. Doing so allows me to retain the high cost heterogeneity in the counterfactuals. The size and variation in base station costs lead costs to be an important strategic consideration, and shows how important inframarginal signal quality is. For example, in 2012, the estimates imply that T-Mobile spent 62% of its variable profit in Connecticut on base stations. Ratios are smaller but still large at other carriers: with 46% (Sprint), 19% (AT&T) and 9% (Verizon).33 Thus signal quality is both 1) important enough to firms to commit significant resources to and 2) where gains in efficiencies could be very beneficial to the firm.34

As a point of comparison, AT&T and Verizon spent $6.6 billion and $9.4 billion on licenses in the spectrum auction of the 700 MHz band in 2008. This was the only auction any of the four national ﬁrms won spectrum during my sample period: Sprint and T-Mobile did not win any spectrum. Using the UBS data, these expenditures would amount to 19.2% and 22.2% of 2008 national variable revenue of AT&T and Verizon, respectively, or 3.2% or

33The variation in ratios is largely due to the diﬀerence in elasticities with respect to base stations and own demand - with a higher elasticity the greater the relative gain from investment will be. Accordingly, these proportions correspond closely to those elasticities, which are not reported but available upon request.

34See Table 2.9 for ratios over multiple years.

116 3.9% for the entire sample period, respectively. Thus base stations costs are quite signiﬁcant even relative to spectrum costs. To investigate how my regulation instruments compare with cost I decompose the marginal quality cost in two ways:

∂φkmt(Nkmt) = Hkmtψ + νkmt (2.19) ∂Nkmt and

∂φkmt(Nkmt) Am = Hkmtψ + νkmt (2.20) ∂Nkmt

The former is self-explanatory: the marginal quality cost is a linear function of regressors

Hkmt plus an error term. The second specification posits that costs are constant in density, and not base stations, which may be plausible given that as firms continue to build in the same area, the costs of finding new suitable locations would increase. Hkmt should include all possible cost shifters for firms, but for my purposes I include fixed effects for firms besides AT&T, interactions of these terms with the regulation instrument, and a fixed effect for the city markets. I then regress Hkmt on the LHS of the above equations. The constant returns in base stations model yields estimates that imply the regulation variable decreases the cost of base stations for all firms, which contradicts the rationale for using the instrument. However, the alternative model implies that cost per base station density does increase with regulation. I thus prefer the density model as the explanation of how the regulation variable contributes to costs. Looking at the other results for that model, the intercept of density costs of AT&T is much lower than that of its rivals, while AT&T is affected much more by regulation. This may be because of differences in the skill of a carrier’s regulatory staff or because a larger firm like AT&T or Verizon has more resources

117 Table 2.10: Cost Projected onto Covariates

$1,000 per: Base Station Base Station per 1000 km2

Constant 13.05*** -1.85* (0.75) (1.02) Sprint Dummy -1.75* 2.08* (0.92) (1.13) T-Mobile Dummy -0.94 3.33*** (1.39) (1.06) Verizon Dummy -5.07*** 3.06*** (1.00) (1.02) % Telecom Regulation -56.78** 388.57*** (27.02) (52.81.31) % Telco Regulation * Sprint -61.56* -256.54*** (37.53) (54.69) % Telco Regulation * T-Mobile -100.99*** -256.54*** (37.14) (54.69) % Telco Regulation * Verizon -61.56* -174.00*** (33.14) (58.80) City Dummy 17.30 -225.00*** (34.20) (52.95)

R2 0.26 0.64 Obs 480.00 480.00

***, **, * indicate 1%, 5% and 10% signiﬁcance, respectively.

118 to devote to regulatory issues. Presumably, there are many omitted variables in this regression, as site leasing fees, construction, backend, and power fees that should all contribute to the cost of installing and running the base stations. As a result, I am not conﬁdent that the model estimates of ψ are robust enough to use in counterfactuals, which is another reason why I prefer to use the entire terms ∂φkmt(Nkmt) which includes the large amount of variation from market speciﬁc ∂Nkmt variation that the current regression cannot explain.

2.6 Counterfactuals: Mergers of a Small Carrier

To learn about the impact of market consolidation in this industry, I examine two merger proposals that recently have been pursued: the attempted acquisitions of T-Mobile by AT&T in 2011 and then Sprint in 2014. The AT&T attempt got quite far in the approval process until it was ultimately abandoned in December of 2011 after the Department of Justice decided to oppose it. The Sprint attempt only got as far as discussions when it was abandoned in August of 2014. Allegedly, this was due to concerns that the merger would also be ultimately opposed as well, demonstrating government concerns about mergers in this industry overall. While these proposed mergers actually failed and thus are unlikely to be attempted again, they should resemble the most likely mergers to be proposed in the future - acquisitions of small regional carriers like Alltel or Pocket, which in their service areas might have comparable market share to T-Mobile. Counterfactuals are conducted using the state of Connecticut, as the costs so far estimated are speciﬁc to the markets I have supply data for. As with the cost estimation, I used the PUMS sample with the Census-assigned weights to simulate population level demand. To correspond closest with current conditions and long-run outcomes, I set the counterfactuals

119 in 2012 and use the corresponding parameters and PUMS data.35 Inferences will be limited to Connecticut, but this should give a good idea of what would happen nationally given that the market shares are not wildly different from reported national levels, and individual level characteristics (which are different from the national average) are not overwhelmingly important in determining demand.36 I find equilibrium levels of quality by use of a fixed point algorithm, a la Morrow and Skerlos (2011). One can rewrite the first order condition (2.18) explicitly using the chain rule:

X ∂Dkmt(Qmt) dQkmt ∂φkmt(Nkmt) (P − C ) = (2.21) jkt kt ∂Q dN ∂N j kmt kmt kmt

Conveniently, base station count appears as the denominator of the derivative of signal quality due to the log speciﬁcation, so I can write:

−1 X ∂Dkmt(Qmt) ∂φkmt(Nkmt) (γ + γ 1(city)) (P − C ) = N (2.22) k city jkt kt ∂Q ∂N kmt j kmt kmt

So I only need to calculate the right hand side of the equation given a guess for Nk0mt for all ﬁrms, which produces a new guess, which produces a new right hand side, and so on until convergence. Uniqueness is an issue, since it is not guaranteed in this setup with strategic substitutes. I have tried various starting values for the counterfactuals and have found no other equilibrium, but it may be possible that other equilibria exist. The players are very asymmetric, so there

35As noted earlier, I use the 2011 PUMS data scaled up to the 2012 Connecticut population for 2012 since geographical deﬁnitions of the PUMAs changed.

36 Otherwise, there may be issues: Connecticut is the 4th densest state and the wealthiest.

120 may be less of an issue with multiplicity than if they were very similar.37 Pricing is certainly important to judging counterfactual situations in mergers and was the main focus of the AT&T-T-Mobile merger review. I have abstracted from pricing for the most part in this paper, since national pricing makes identification of price sensitivity difficult, and I cannot match the national-level first order condition to my state level data. Up to this point, all the estimates are valid given the assumption that pricing is done in a first stage before the base station decisions, but a merger between the firms looked at in the counterfactuals would be at the national level and so pricing incentives would change. Given I cannot exactly model the equilibrium price adjustment, the counterfactuals would all likely lead to higher equilibrium prices as I do not assume any marginal consumer cost efficiencies. Either there are fewer products, or cannibalization effects of lower prices are internalized, so the incentive to price higher increases. I therefore run each counterfactual with both no price adjustment and a 5% price adjustment for all firms. I chose 5% since it seemed that if merger authorities expected a price increase any higher they would have blocked the merger, regardless of any quality adjustments. One further wrinkle is that the effect of a merger is mediated by the price impact on utility. I did not estimate this, so I appeal to the literature for guidance, using previously estimated own-price elasticity for wireless phone plans to calibrate the counterfactuals. Unfortunately, there is significant variance in the elasticities estimated - for example, Sinkinson (2014) reports own elasticities of price of -1.4 from Verizon and -1.5 for AT&T.38 In contrast, Jiang

37In particular, with identical players there can be asymmetric equilibria in which players play diﬀerent strategies but there is no guidance on which player will play which strategy.

38As he is using a sample from the same survey I am, Sinkinson (2014) also has no market variation in price, and very little variation across time since his panel is short. He instead relies on product characteristic variation in both service and phones and does not use product level eﬀects as I do. Price then is used to explain product-year level variation in shares, while controlling for as much of the demand variation as possible. In particular, he has Nielsen data from drive tests that actually measure dropped call rates, providing variation on roughly the MSA level, which is much larger than the market I examine.

121 (2013) reports much higher elasticities for the fixed fees of contracts: -5.33, -6.92, -5.09 and -4.78 for AT&T, Sprint, T-Mobile and Verizon, respectively. Jiang (2013) also reports an industry own-price elasticity (with respect to having no phone at all) of -0.61, which is much higher than Miravete and Röller(2004)’s report of -0.13, the lowest estimate that I know of in this literature. Casual empiricism implies that the Sinkinson (2014) is more believable for my sample. Under Nash-Bertrand pricing (and assuming no other endogenous variables), the equilibrium percent markup is equal to the negative inverse of the own elasticity, the so-called Lerner In- dex.39 For postpaid plans, the implied elasticities under this rule are between -1.11 and -1.36 in my sample, and for prepaid plans between -1.15 and -2.47. The Jiang (2013) elasticities are clearly much higher and might suggest that elasticities have changed significantly over time. Jiang (2013) looks at a sample period from 2000 to 2001, while Sinkinson (2014) is much more recent, looking at 2008 to 2010. I therefore find the price coefficients that match Sinkinson (2014)’s price elasticity for AT&T and Verizon in my data for the years 2008 to 2010 and then take their average as the coefficient I use in estimation. I also run counterfactuals holding the actions of non-merging firms fixed. I call this the “Unilateral” case, and I do this to examine how much the actions of non-merging firms have on the equilibrium. When I allow firms to adjust prices in the Unilateral case I only allow the merging firms to do so. I use the three scenarios from Section 2.2 again, and since base stations are strategic substitutes, then comparative statics from the logit example still hold. I explain the theoretical forces at work in detail in Appendix B.2. I review the results of each below, organized by the scenarios.

39See Lerner (1934).

122 2.6.1 Discontinue All Products from Purchased Firm (*)

Here the counterfactual is the same whether AT&T or Sprint buys out T-Mobile - T-Mobile products leave. The only difference is how much the lump sum transfer is and who is paying it, but that is outside of the scope of the model. Given my finding of base stations as strategic substitutes, I find that when T-Mobile leaves, the remaining firms increase their base station density. For example, holding prices fixed AT&T would increase base stations by 2.90%, Sprint 8.15% and Verizon 2.46%. In the unilateral cases actions for the firms allowed to move resemble quite closely the full equilibrium case. For example, holding prices fixed and allowing only the acquirer to move, AT&T would increase base stations by 3.14% and Sprint 8.75%.

However, even when holding prices fixed there is a net consumer welfare decrease of $1.35 per consumer. Given no price change, this must come from the loss of T-Mobile variety and the substitution of some former T-Mobile consumers to the Other composite carrier and the outside option. This only gets worse as prices increase 5%, as per capita monthly consumer welfare losses increase to $3.42. When all prices are allowed to adjust, quality increases are actually higher, as the price increases marginal revenue from each new consumer. However, the price increase of 5% seems to be too high for Sprint moving alone, since the profit is not as high when price is held fixed, and quality actually reduces slightly. The AT&T merger with a 5% price increase also has lower profit for AT&T without accommodation, which suggests that the unilateral price increase for this merger is actually less than 5%. Profit of the acquirer is larger than sum of their pre-merger profit and T-Mobile’s in only some of these counterfactuals. These mergers can be rationalized with variable profits alone without outside fixed savings. This happens in the equilibrium case with AT&T, and in the unilateral case with AT&T with no price changes. In contrast, none of the Sprint mergers can be justified, due to Sprint being not large enough so that the benefit to Sprint’s profit

123 Table 2.11: Dropping T-Mobile Product Line

Unilateral Unilateral Equilibrium Adjustment Adjustment Adjustment by AT&T by Sprint

No Price 5% Price No Price 5% Price No Price 5% Price Change Change Change Change No Price 5% Price

Median % AT&T 3.14 3.65 - - 2.90 6.41 Change in Sprint - - 8.75 -0.06 8.15 13.04 Base Stations T-Mobile ------Across Markets Verizon - - - - 2.46 6.13 124 AT&T+T-Mobile -47.09 -46.30 - - -47.16 -44.91 Sprint+T-Mobile - - -45.54 -50.79

AT&T 9.60 3.93 8.92 11.14 8.81 10.52 % Change in Sprint 13.71 23.23 14.77 1.63 13.89 18.25 State-Wide T-Mobile ------Proﬁt Verizon 8.01 13.77 7.87 9.77 7.51 11.11 AT&T+T-Mobile 2.25 -3.04 1.50 3.10 Sprint+T-Mobile -20.76 -29.83 -21.36 -18.35

$ Per Capita CS Change -1.46 -2.44 -1.43 -1.78 -1.35 -3.42 $ Per Capita PS Change 1.77 2.05 1.68 1.97 1.59 2.44 $ Per Capita TS Change 0.34 -0.38 0.25 0.20 0.25 -0.98 can make up for the complete loss of T-Mobile’s. Thus a merger with small brands would likely prefer to retain both brands in some form.

2.6.2 Retain Products from Purchased Firm with Separate Net-

works (**)

When the acquirer does retain T-Mobile and the two networks do not integrate at all, the outcome depends greatly on whether AT&T or T-Mobile is the acquirer. When the acquirer is large, i.e. AT&T, almost all the T-Mobile base stations are removed, since the AT&T products have a much higher mean utility, net of their network. The counterfactual then ends up resembling the dropped products case very closely, since T-Mobile is practically dropped. Without price changes, base stations for AT&T actually rise on average since the drop in signal quality of T-Mobile is so great that the strategic substitutability of base stations overcomes the internalization of the cannibalization eﬀect. When the two merging parties are more similar, as when Sprint is the acquirer, the adjustment in base stations is not so asymmetric. T-Mobile does lose a signiﬁcant amount of base stations absent price changes, but Sprint also decreases a few percent on average. The equilibrium case is therefore better for consumers than when AT&T was the acquirer,

so that consumers, holding prices fixed, only lose $0.23 a month rather than $0.98, and, increasing price by 5%, they lose only $2.30 a month rather than $3.00. Mergers are more profitable in this scenario than in ∗, since firms can profit off consumers who have such high taste shocks for T-Mobile they still buy it even when the quality degrades significantly. Again, the only case where the profit does not justify the merger for AT&T is the unilateral case where price increases. The Sprint mergers are also profitable for all but the unilateral case with the price changes, in contrast to ∗ where Sprint cannot benefit from the T-Mobile product line.

125 Table 2.12: AT&T Buys T-Mobile, Separate Networks

Unilateral Equilibrium Adjustment Adjustment by AT&T & T-Mo

No Price 5% Price No Price 5% Price Change Change Change Change

Median % AT&T 0.62 0.09 0.42 3.19 Change in Sprint - - 6.02 10.42 Base Stations T-Mobile -84.17 -80.46 -84.06 -80.26 Across Markets Verizon - - 1.99 5.75 AT&T+T-Mobile -39.20 -37.82 -39.31 -36.02

AT&T 6.73 0.94 6.15 7.46 % Change in Sprint 9.25 18.26 9.29 12.96 State-Wide T-Mobile -44.68 -39.86 -44.96 -35.77 Proﬁt Verizon 4.82 10.22 4.47 7.68 AT&T+T-Mobile 3.27 -1.79 2.72 4.57

$ Per Capita CS Change -1.06 -2.00 -0.97 -3.00 $ Per Capita PS Change 1.50 1.75 1.37 2.18 $ Per Capita TS Change 0.44 -0.24 0.39 -0.82

126 Table 2.13: Sprint Buys T-Mobile, Separate Networks

Unilateral Equilibrium Adjustment Adjustment by Sprint & T-Mo

No Price 5% Price No Price 5% Price Change Change Change Change

Median % AT&T - - 0.69 3.89 Change in Sprint -4.98 -13.29 -4.93 -1.66 Base Stations T-Mobile -27.63 -30.91 -27.74 -18.56 Across Markets Verizon - - 0.69 4.62 Sprint+T-Mobile -15.61 -22.14 -15.56 -10.23

AT&T 1.83 4.61 1.81 2.93 % Change in Sprint 2.63 -8.25 2.45 5.72 State-Wide T-Mobile -2.24 -11.00 -2.48 9.00 Proﬁt Verizon 1.54 3.96 1.48 4.46 Sprint+T-Mobile 1.12 -9.10 0.92 6.73

$ Per Capita CS Change -0.31 -0.76 -0.29 -2.30 $ Per Capita PS Change 0.50 0.90 0.48 1.24 $ Per Capita TS Change 0.19 -0.14 0.19 -1.06

127 This simulation most resembles the previous literature in merger repositioning, so I can directly compare my findings with the effects predicted and estimated therein. Qualitatively, the effects are in line with theory - as in Gandhi et al. (2008) and Mazzeo, Seim, and Varela (2014) the merging firms differentiate their products to reduce cannibalization and as a side effect, the non-merging firms benefit as at least one of the merged product lines is now much more differentiated relative to them. The simulations effects are also in line with empirical studies of the ice cream industry by Draganska, Mazzeo, and Seim (2009) and of the newspaper industry by Fan (2013). The simulations’ effects are even consistent with studies of mergers of radio in Berry and Waldfogel (2001), Sweeting (2010), and Jeziorski (2014), which are partially motivated by the observation that radio station diversity has indeed become less diverse overall since the Telecommunications of Act of 1996 allowed radio station mergers. Sweeting (2010) attempts to resolve this paradox by pointing out that in his model, mergers result in increased differentiation between merging partners, but one of the partners ends up more similar to rivals which may result in an overall decrease in diversity. Interestingly, in my simulations this does not quite happen. In the AT&T and T-Mobile merger, Sprint and AT&T become closer in overall quality, but Verizon furthers its advantage and differentiates itself even more. In the Sprint and T-Mobile merger, both Sprint and T-Mobile degrade in quality and thus further differentiate themselves from the non-merging AT&T and Verizon. These differences suggest that even if the core economics of reducing cannibalization is consistent in mergers in many industries, the exact merger effects will depend heavily on the application.

128 Table 2.14: AT&T Buys T-Mobile, Single Network, AT&T Costs

Unilateral Equilibrium Adjustment Adjustment by AT&T & T-Mo

No Price 5% Price No Price 5% Price Change Change Change Change

Median % AT&T 26.49 29.77 26.67 33.11 Change in Sprint - - -5.02 -1.73 Base Stations T-Mobile 37.84 40.86 38.03 43.93 Across Markets† Verizon - - -1.50 2.47 AT&T+T-Mobile -34.96 -33.98 -34.91 -32.03

AT&T ---- % Change in Sprint -6.91 2.26 -6.72 -3.61 State-Wide T-Mobile ---- Proﬁt†† Verizon -4.19 1.32 -3.97 -1.09 AT&T+T-Mobile 11.67 7.29 12.08 14.70

$ Per Capita CS Change 0.78 -0.26 0.72 -1.30 $ Per Capita PS Change 0.91 1.27 1.00 1.86 $ Per Capita TS Change 1.69 1.01 1.73 0.55

† For the merged firms the final base station count used for the difference is the effective count, which sum of the count of the two merging firms.

†† For the merged firms cost cannot be disentangled between the two networks so for those firms I do not report individual profits.

129 Table 2.15: AT&T Buys T-Mobile, Single Network, T-Mobile Costs

Unilateral Equilibrium Adjustment Adjustment by AT&T & T-Mo

No Price 5% Price No Price 5% Price Change Change Change Change

Median % AT&T 159.45 168.02 160.38 174.40 Change in Sprint - - -22.66 -19.90 Base Stations T-Mobile 251.21 264.24 251.43 270.55 Across Markets† Verizon - - -7.06 -3.36 AT&T+T-Mobile 43.22 49.02 43.46 51.19

AT&T ---- % Change in Sprint -26.28 -17.45 -24.24 -21.79 State-Wide T-Mobile ---- Proﬁt†† Verizon -15.84 -10.55 -15.04 -12.54 AT&T+T-Mobile 30.04 26.51 31.99 35.62

$ Per Capita CS Change 3.25 2.12 3.01 0.98 $ Per Capita PS Change 1.43 1.88 1.87 2.80 $ Per Capita TS Change 4.68 4.00 4.88 3.78

† For the merged firms the final base station count used for the difference is the effective count, which sum of the count of the two merging firms.

†† For the merged firms cost cannot be disentangled between the two networks so for those firms I do not report individual profits.

130 Table 2.16: Sprint Buys T-Mobile, Single Network, Sprint Costs

Unilateral Equilibrium Adjustment Adjustment by Sprint & T-Mo

No Price 5% Price No Price 5% Price Change Change Change Change

Median % AT&T - - -4.36 -1.94 Change in Sprint 123.60 115.82 125.22 137.30 Base Stations T-Mobile 103.55 96.91 104.70 117.22 Across Markets† Verizon - - -3.23 0.21 Sprint+T-Mobile 6.60 4.13 7.11 13.27

AT&T -11.85 -8.41 -11.65 -10.78 % Change in Sprint ---- State-Wide T-Mobile ---- Proﬁt†† Verizon -10.90 -7.86 -10.46 -7.93 Sprint+T-Mobile 75.16 62.75 77.63 88.52

$ Per Capita CS Change 2.12 1.49 1.99 -0.01 $ Per Capita PS Change -0.83 -0.32 -0.66 0.15 $ Per Capita TS Change 1.29 1.17 1.33 0.14

† For the merged firms the final base station count used for the difference is the effective count, which sum of the count of the two merging firms.

†† For the merged firms cost cannot be disentangled between the two networks so for those firms I do not report individual profits.

131 Table 2.17: Sprint Buys T-Mobile, Single Network, T-Mobile Costs

Unilateral Equilibrium Adjustment Adjustment by Sprint & T-Mo

No Price 5% Price No Price 5% Price Change Change Change Change

Median % AT&T - - -8.48 -6.76 Change in Sprint 229.24 221.89 232.46 253.73 Base Stations T-Mobile 284.76 276.12 288.63 311.64 Across Markets† Verizon - - -6.42 -2.95 Sprint+T-Mobile 69.88 65.31 71.44 82.64

AT&T -16.84 -13.20 -16.54 -15.74 % Change in Sprint ---- State-Wide T-Mobile ---- Proﬁt †† Verizon -14.10 -10.90 -13.53 -11.09 Sprint+T-Mobile 109.68 96.76 113.56 126.27

$ Per Capita CS Change 3.00 2.30 2.83 0.82 $ Per Capita PS Change -0.89 -0.32 -0.65 0.20 $ Per Capita TS Change 2.11 1.96 2.18 1.02

† For the merged firms the final base station count used for the difference is the effective count, which sum of the count of the two merging firms.

†† For the merged firms cost cannot be disentangled between the two networks so for those firms I do not report individual profits.

132 2.6.3 Retain Products from Purchased Firm with Fully Integrated

Networks (***)

Here I assume the network of the ﬁrms can be combined into a single network and are readjusted accordingly. As in Section 2.2 this means that consumers of ﬁrm k merging with

∗∗ firm h experience an effective network of size Nk = Nk + ρNh, where ρ represents the spillover. Efficiency comes from the fact that a network can be used by two products lines, so a base station can provide quality to consumers who had idiosyncratic taste preference for either brand, rather than just one. Thus more consumers are attracted by the marginal base station, so quality provision becomes effectively cheaper. As in the example, I assume ρ = 1 for simplicity and to represent the maximum amount of integration possible. The fixed cost of the new merged network now needs to be specified, since it is not clear what the fixed cost of the new network will be. Once ρ = 1, base stations from a quality standpoint are completely fungible with each other and with different costs base stations for one network will be unambiguously more or less expensive. Thus if I assume the ex post joint firm has access to both kind of base stations, then it will clearly choose only the cheapest one, and all its base stations will have the lowest of the two pre-merger costs.40 This is probably too optimistic, since much of the costs come from long-term contracts in real estate, labor and equipment that would be still valid after the merger. Much of the cost also appears to be non-pecuniary and is related to how the firm deals with delays and regulation, which might have more to do with the identity of the managing and legal staff. Due to the costs of reintegrating multiple teams, it might be that the acquirer retains its

40 If one assumes ρ < 1, then it can be the case that both networks are utilized since it may be more eﬃcient to utilize network built speciﬁcally for one group of consumers, rather than to remove one network and only allow to formers consumers of that network to experience only ρ of the remaining network. In practice, this only happens in counterfactuals when ρ is relatively low: above approximately 0.1 there are corner solutions with only one network being used. Given the strong strategic substitutability in this counterfactual, as detailed in Appendix B.2, this is not surprising.

133 staff but removes the corresponding staff of its acquiree, even if they are more capable. To cover all the possible situations, I report the counterfactuals but I assume either the acquirer cost or T-Mobile’s cost is used. In practice, T-Mobile is almost always cheaper - the case where the firm literally chooses the lower of the two markets cost is very similar to the T-Mobile case so for space concerns I decline from reporting it. The counterfactuals under these assumptions yield very different results from the two previous cases. The efficiencies increase incentives for quality improvement of the merging firms substantially. Due to the strategic substitutability of signal quality, non-merging carrier decrease their base stations in equilibrium, though the accommodating effects of their actions seem to be small relative to the efficiency gains. As a result, the unilateral cases are qualitatively quite similar to the equilibrium cases. With AT&T as the acquirer and assuming AT&T costs, the merged entity has less total base stations, but the total is still greater than AT&T or T-Mobile individually ex ante, so signal quality has improved. Assuming Sprint as the acquirer and using Sprint costs, the median total of base stations between the two firms actually increases, from between 4.14% to 13.27%. In all the cases examined with the lower T-Mobile costs, the growth in base stations is even larger: the median combined number of market base station from the merging firms grows at least by 40%. Without price changes, the non-merging firms reply with modest base station removals, but overall consumer welfare improves since the quality gains for consumers of the merging firms are large. Assuming acquirer (aquiree) costs and holding price fixed in the quality

setting equilibrium: the monthly per capita gain is $0.72 ($3.01) for the AT&T merger and $1.99 ($2.81) for the Sprint merger; the AT&T joint entity ends up with 7% (26%) in increased statewide proﬁt and the Sprint joint entity with 65% (98%). Letting price increase 5% in quality setting equilibrium and under costs of the acquirers there is a net consumer welfare loss of $1.30 per capita per month from the AT&T merger

134 but a loss of only $0.01 from the Sprint merger. Under T-Mobile costs the mergers are both net beneficial to consumers, with gains of $0.98 per capita per month from the AT&T merger but a loss of only $0.82 from the Sprint merger. Again the size of the acquirer matters, and so does the assumption of the cost change. Also key is the retention of the T-Mobile brand - without it there is no real efficiency, and it reverts to the first counterfactual type which was clearly consumer-harming. However, if the antitrust authorities think prices ex post a merger will remain under the 5% increase rule of thumb then further quality benefits can indeed lead to net consumer benefit in this industry. I therefore conclude merger authorities, in the wireless and other industries with similar network efficiencies, should take seriously claims of cost efficiencies, as the savings can be quite large, and consumer quality can be improved. However, such claims are quite contingent on the actual way the technology could be reconfigured after the merger, and under only somewhat different circumstances the end results could instead be very anti-competitive. To identify which case is before them, merger authorities should request very detailed information about the industry, the technology used by each firm, and plans on how the efficiencies will be realized. This is particularly important as firms may erroneously believe merging networks to be highly compatible, harming both the merging parties and consumers. In 2005, Sprint and Nextel famously merged with expectations of a smooth integration of networks with different technologies. Sprint uses CDMA, while Nextel used iDEN, a unique standard developed by Motorola. That integration never materialized, with the two networks coexisting until Sprint decided to completely decommission the iDEN network in 2013. The deal is now infamous for ending up with a merged entity smaller in market value than the merger purchase price.41

41See “Was Sprint Buying Nextel One Of The Worst Acquisitions Ever At $35b?” Forbes.com, 11/29/2012, http://www.forbes.com/sites/quora/2012/11/29/was-sprint-buying-nextel-one-of-the-worst- acquisitions-ever-at-35b/.

135 The merger authority therefore may need to assume the role of objective observer to check unrealistic expectations of merging parties. A final note about all the counterfactuals is the effect on the non-merger parties. Non- merging parties actually do worse when the benefits to consumers are highest, as the network quality improvements in the joint firm imply the joint firm is going to be a stronger com- petitor. This discourages the non-merging firms from investment and they produce lower quality services and settle for reduced market share. For example, when there is no integration possible and holding prices fixed, the AT&T merger allowing all firms to adjust quality would lead to 9.59% profit increases for Sprint and 4.47% for Verizon. With full integration and T-Mobile costs, that merger implies profit decreases of 24.24% and 15.80% percent for Sprint and Verizon respectively. Thus, the benefits of a merger for the rest of the industry might go in the opposite direction relative to consumers.

2.7 Conclusion

I have conducted an analysis of how market structure aﬀects the incentives for providing a particular component of product quality, signal quality in mobile phone networks. Using a unique statewide dataset, I estimate a structural model of mobile phone service demand that relates consumer value to the density of base stations in a consumer’s local market. Estimates reveal that marginal base station density is most important for Sprint and T- Mobile, even though AT&T and Verizon have more market share. This is possibly because their larger and more diverse spectrum portfolios allow them to reach levels of high signal quality and rapidly diminishing returns with fewer base stations. Own and cross elasticities of demand with respect to base stations are relatively mild, but still translate into sizable costs per base station. The demand system implies strategic substitutability of base stations, which will mitigate any change in base stations by one ﬁrm with changes in the opposite

136 direction by their rivals. Counterfactual analysis of several recently proposed mergers are consistent with findings in theory and other industries that when firms can reposition their products after a merger, the diversity of merged product lines increases to reduce cannibalization. The simulations also show that results for consumers and firms can differ greatly based on the assumption of how the two formerly separate networks and products are integrated ex post the merger. Under removal of the T-Mobile product line, consumer welfare falls greatly despite increases in signal quality by all remaining firms. Maintaining two separate networks under one company results in degradation of the smaller (T-Mobile) network, and overall welfare losses to consumers. In contrast, integration of the networks makes the effective cost of base stations much smaller, and both the merging firms and consumers benefit. In addition, these findings give credence to possible merger defenses where integration is possible, but merger authorities should be cautious since small (5%) price increases tend to erase consumer gains unless there are comparable improvements in costs elsewhere. Merger authorities should therefore take seriously claims of cost synergies in network industries, but demand sufficient information and detailed plans from merger applicants to determine the validity of those claims.

137 Chapter 3

The Costs and Clarifying Eﬀects of Regulation for Business Investment: Evidence from Cell Siting

Patrick Kainin Sun

138 3.1 Introduction

Land use regulation is regulation that controls whether something can be built on land and, if so, how. Such regulation can have significant economic impacts and such regulation is ubiquitous. A vast literature attempts to measure the effects of this kind of regulation on the housing market.1 In contrast, very little research has focused on the effects of land use regulation on business investment. Given that buildings like stores, plants, and offices are important investments for firms, land use regulations should have an important impact on firms. What limited research is generally limited to retail stores, and focuses on a single type of regulation or index of regulatory stringency.2 In general, land use regulations can act as barrier to entry - Suzuki (2013), Nishida (2014, 2015), and Griffith and Harmgart (2012); can reduce employment - Bertrand and Kramarz (2002); can decrease productivity - Haskel and Sadun (2012); and can change the nature of competition by forcing firms to locate closer together - Kuzmenko (2007), Ridley, Sloan, and Song (2011), and Datta and Sudhir (2013). Schivardi and Viviano (2011) actually found evidence for multiple effects: less stringent entry regulation appears to increase productivity and investment in information technology and appears to lower profits, prices, and labor costs. This paper contributes to this small but important literature by examining cell siting, where land use regulation may be the strongest determinant of costs. Unlike previous papers in the literature, which, except for Suzuki (2013), focus on retail, I will look at what are

1The literature is too large to summarize here; a good reference is Gyourko and Malloy (2014).

2Nishida (2014, 2015)), Kuzmenko (2007), Ridley, Sloan, and Song (2011), Datta and Sudhir (2013) study the eﬀect on an area being zoned or not on retail entry. Bertrand and Kramarz (2002), Schivardi and Viviano (2011), Haskel and Sadun (2012), and Griﬃth and Harmgart (2012) study the impacts of direct regulation of retail entry through zoning and planning commissions. Suzuki (2013) uses the Wharton Residential Land Use Regulatory Index (WRLURI) from Gyourko, Saiz, and Summers (2008) discussed later in Section 3.6 to study hotel entry in Texas.

139 essentially investments in industrial facilities, as cell sites are a kind of quality improving plant. This links the land use regulation literature to the plant location literature. Gray (1997) finds fewer plants open in states with more stringent regulation. In contrast, Bartik (1988) and Levinson (1996) find no relationship between plant openings and regulation. In addition, while previous studies used either indices or focused on a single kind of statute, I will look with greater detail into the text of the statutes, and highlight how differences in the composition of the zoning code have different effects. In particular, I test two main hypotheses about how different kinds of regulations may have different effects for investment. First, regulation that increases costs to the firm reduces the incentive to locate in that jurisdiction. Second, clearer regulations that reduce the possibility of legal disputes with regulators increase the incentive to locate in that jurisdiction. While related to the transactions costs and property rights models of regulatory behavior which emphasize the risk for regulated firms to commit asset-specific capital, this point about the statutory content appears to be novel to the literature.3 I expand the analysis in Chapter 2 where a crude measure of regulation, the share of the total zoning code devoted to telecom facilities, is used as instrument for costs and do a much more careful analysis to ensure identification of the impact of regulation on investment. I then use a tractable discrete choice model of site location over towns, defining the consideration set for each existing site as each town in a half mile radius. This exploits variation in regulation within small areas across town borders, which controls for unobserved local variation. My application is most similar to Turner, Haughwout, and van der Klaauw (2014), which looks at how land use regulation affects house prices, and Gowrisankaran and Krainer (2011) which thinks about firm location decisions across a border.4

3See Williamson (1975), Goldberg (1976), Klein, Crawford, and Alchian (1978) and Salant and Woroch (1991, 1992).

4Border discontinuity has long history as an empirical strategy. For prominent examples, see Card and

140 To measure different kinds of regulation, I conduct 1) a manual analysis of the documents and derive indicators for potential regulations of interest and 2) a Latent Dirichlet Allocation from the computational linguistics literature (David M. Blei and Andrew Y. Ng and Michael I. Jordan (2003)) to assign “topics” to text found in the zoning regulations. I identify these “topics,” essentially distributions of words, with different kinds of regulations and compare and contrast these measures and their effects with the manually coded measures. To my knowledge this study is the first use of topic-modeling to examine the effects of land use regulation in economics, and the first combination of topic-modeling with a border discontinuity approach. Topic modeling has been used in economics before, most prominently in Hansen, McMahon, and Prat (2014) to evaluate the career concerns model with respect to meetings of the Federal Open Market Committee. Different computational linguistics methods have been used in other applications. Moszoro, Spiller, and Stolorz (2015) uses computational linguistics to classify public contracts into different contractual forms. The REGDATA project by Al-Ubaydli and McLaughlin (2014) has produced a database of federal regulatory stringency measures by industry. The REGDATA database has been used in a growing number of papers. In both specifications, I largely confirm the hypotheses of qualitative effects of different regulations. More burdensome regulations - requirements with regards to landscaping, radiation, and power, and time limits on approvals - reduce the probability of locating in a town. In contrast, language that improves clarity - acknowledgment of federal and state standards and authority and greater precision about timing - increases the probability of locating in a town. This second finding with regards to federal authority is in contrast to some legal arguments, such as Tan (1997) and Niehaus (2001) that argue that the Telecommunications Act of 1996 had the negative effect of making the jurisdiction over siting decisions more

Krueger (1994) and Dube, Lester, and Reich (2010) in the minimum wage literature, and Black (1999) and Bayer, Ferreira, and McMillan (2007) in the education literature.

141 ambiguous. A counterfactual where the variation in these effect between towns are eliminated, simulating something like a uniform adoption of regulations, shows the effect of these regulations on reallocation across towns are modest, with about 7% of sites being reallocated. In addition, the relative effect of burdensome regulation is more impactful on reallocation relative to clarifying regulation. Thus the welfare distortions of regulations are likely most felt in the extensive margin of the number of sites built, which is beyond the scope the parameters estimated. Finally, this paper adds to the vast empirical literature about how regulation affects firm investment decisions. In particular, telecommunications has frequently been used as the subject for such studies. Cambini and Jiang (2009) surveys the literature specific to the broadband industry. Chang, Koski, and Majumdar (2003), Nitsche and Wiethaus (2011), and Grajek and Röller(2012) look at the impact of interconnection fee regulation and its effect on infrastructure investment.

3.2 The Impact of Land Use Regulation

Given the nature of the approval process there are two main potential impacts of a zoning code on siting costs.

3.2.1 Burdens

First, a zoning code can increases costs by imposing greater burdens on the investing ﬁrm. The burdens are often directly pecuniary: towns often require application fees. These can be signiﬁcant, in the thousands of dollars, but are often small relative to the other potential costs of the regulations. Many Connecticut towns also require a lump sum to create a municipal bond so that if a tower is abandoned by applicant at a future date the

142 removal costs are covered. What can be even more costly is simply producing the application. Site applications are long, complicated documents with many analyses by lawyers and engineers. The more required analyses by more outside experts will result in greater costs. A town can even require for applicants to pay for the town’s analysis, with experts of their choosing. Further, the requirements can lead to inferior coverage or more costly locations by re- stricting what is a feasible site. Geometrically, there is an optimal place to put a base station to cover a “‘deadspot” (i.e. the middle of the “deadspot”), but that location is generally infeasible since it is not available for development. Rather, site developers have to settle for nearby locations where space can be leased and balance the resulting coverage benefit from the costs of developing that site. Regulation requirements for the site can make this harder. For example, minimum lot sizes can greatly reduce the set of admissible locations if real estate is scarce. Thus the developed site might have higher costs or lower effective coverage than a comparable, impermissible site. Finally, the application procedures are lengthy, often lasting years. The time needed to get to actual construction can cost the firms significantly in opportunity cost. A major problem is actually town response to applications. The wireless industry association, the CTIA, had demonstrated to the FCC 760 out of 3,300 applications for site development had been pending for over one year and 180 had been pending for more than three. In response, the FCC has established deadlines for final decision by zoning boards, the so-called FCC Shot Clock.5

5The initial order was sent out in 2009 and was not decided with ﬁnality until the 2012 Supreme Court decision in City of Arlington, Texas v. FCC.

143 3.2.2 Clariﬁcation

Risk is a signiﬁcant concern for applicants because even zoning codes are often ambiguous. As Glaeser and Ward (2008) notes,

“. . . land use regulations are often astonishingly vague, which increases the likelihood that there will be disputes about implementation.”

A firm may therefore develop a proposal for a site, line up leases with landowners, conduct the required tests and reports and confer with all local parties involved, but due to the ambiguity of the statutes may accidentally fail to fulfill the application requirements as intended by the town. This leaves the firm open to litigation from the town or third parties. Given the firm has committed to the site, the town and other litigants can extract concessions from the firm, creating ex post regret for having ever applied. The risk of the regulatory process described above has the flavor of an incomplete contract from the transactions costs or property rights literatures. The transactions costs literature notes that transactions costs plague market transactions and may lead to alternative forms of economic organization.6 In particular, the literature highlights the threat of opportunistic behavior of opposing parties once one is locked into a transaction. The property rights literature focuses on how incomplete contracts may induce insufficient investment in those assets prior to contracting because worries about ex post renegotiation of the contract.7 If one considers the application as a kind of asset-specific investment, regulatory clearance as a transaction, and the zoning code as terms of the “social contract,” then the applicant has a problem with contract incompleteness in the sense that a vague code will lead to con- tingencies where the regulator may decide the firm has not fulfilled its statutory obligations

6The earliest work in this literature begins with Williamson (1975, 1979, 1985) and Klein, Crawford, and Alchian (1978).

7The earliest work in this literature begins with Grossman and Hart (1986),Hart and Moore (1990).

144 in the application, breaking the terms of the “contract.” This forces the firm back into “renegotiations,” which can be broadly construed as further demands for modifications to the application at the review stage or even litigation. Even if the firm does well in renegotiation, by perhaps winning a court case to allow its original application go through, the firm loses through the delay and legal costs. This point is not just theoretical: there are numerous cases where firms have sued towns for denying applications based on what they consider to be incorrect interpretations of their own statutes:8

• In 2012, AT&T sued the town of San Pablo, CA for denying its application to add a base station to a roof because (among other reasons) it believed it fulﬁlled the zoning requirements for such an addition.9

• In 2013, Verizon threatened to sue the town of Pinole, CA, to enforce a lease that was somehow signed by the city but that the City Council voted down. The issue hinged on whether a signed lease is suﬃcient or whether the statute requires the Council’s approval.10

• In 2014, Verizon sued the town of Cape Elizabeth to allow construction on a water

8These issues are possible with regards to federal statutes. The aforementioned case, T-Mobile South, LLC v. Roswell, started with a denied 2010 application for a cell tower that had did not include a speciﬁc reason for denial, but merely referred to the later released minutes of the meeting in which the town council decided to deny application. In the Telecom Act, reasons for denial must be provided in a “written record.” The applicant, T-Mobile South, LLC, contested that the notice of denial without speciﬁed reasons was therefore invalid. The Supreme Court ruled that since the minutes were released much later than the notice, the charge was valid, but had the minutes been released at the same time as the denial, it would have not been.

9Lau, Steven. “AT&T Sues Albany Over Council Decision on Cell Tower.” Albany Patch August 17, 2012. http://patch.com/california/albany/at-t-ﬁles-lawsuit-against-albany

10 Lochner, Tom. “Pinole council to reconsider cellphone tower in park after Verizon threat- ens lawsuit.” Contra Costa Times July 8th, 2013. http://www.mercurynews.com/ci 23620348/ pinole-council-reconsider-cell-phone-tower-park-after

145 tower since it believed it met the zoning code’s requirement of an “Alternate Tower Structure.”11

Even if a tower is approved, there can ex post litigation by third-parties claiming both the applicant and the zoning board misinterpreted the statute:

• In 2012, private citizens sued Nevada County, NV for allowing a special exception to build a tower, claiming that legally it could only be waived for a “hardship”.12

• In 2012, private citizens and the town of Greenwich, CT sued the local water company for planning to lease a water tower as an AT&T rooftop site. The plaintiﬀs contend that the site cannot go through because of a deed restriction that requires the land to be used only for water purposes and to maintain the natural beauty of the area.13

• In 2015, private citizens sued the town of Neenah, WI for allowing construction of a cell tower, claiming the tower does not escape a deed restriction that only allows the land in question to be used for municipal purposes.14

• In 2015, private citizens and the town of Warwick, NY sued the town’s ﬁre department for a deal where it uses a unique zoning exemption for the ﬁre department to allow

11 Gardner, Kate. “Cape Elizabeth asks federal judge to pull plug on Verizon Wireless cell tower lawsuit.” Bangor Daily News September 18th, 2014. http://bangordailynews.com/2014/09/18/news/portland/ cape-elizabeth-asks-federal-judge-to-pull-plug-on-verizon-wireless-cell-tower-lawsuit/

12 Rubin & Associates Public Relations. “Nevada County Residents Take On Phone Giant Verizon Wireless.” Yuba.Net January 19, 2012. http://yubanet.com/regional/ Nevada-County-Residents-Take-On-Phone-Giant-Verizon-Wireless.php

13Vigdor, Neil. “Town to join homeowners’ ﬁght against cell tower plan.” greenwichtime.com June 15, 2012. http://www.greenwichtime.com/news/article/ Town-to-join-homeowners-ﬁght-against-cell-tower-3637596.php

14Behnke, Duke. “Town of Neenah faces lawsuit over cellphone tower.” Post-Crescent Media January 27, 2015. http://www.postcrescent.com/story/news/local/2015/01/27/ town-neenah-faces-lawsuit-cellphone-tower/22415889/

146 an AT&T tower to be built. The plaintiﬀs claim the location does not qualify for the exemption.15

Thus adding regulations can actually incentivize more investment if they can reduce risk of later disputes. This can be done via language that is clearer, or establishes objective standards set by third parties like the state or federal governments. I consider these regulations to be “clarifying.”16

3.2.3 Causes of Ineﬃcient Regulation Stringency

Taking together the two above effects, the pure burden effect and increased clarity, the overall effect of these regulations should be still to increase applicant costs. Having no regulations at all, and thus no regulatory hearing or restrictions on site design, should be the least costly outcome for firms. Clearer regulations simply make the eventuality of costly disputes less likely when the costs of the regulations already exist. From a societal standpoint though, the effect of land use regulation will be more ambiguous. In their survey of the zoning literature, Pogodzinski and Sass (1990) note that land use regulations will be welfare-enhancing if they successfully cause actors to internalize externalities from their actions in development. For example, since sites reduce the aesthetic quality of their neighborhoods and thus reduce land prices, perhaps imposing greater costs that reduce the number of sites is welfare-optimal.17 In contrast, land use regulation may still be excessive relative to its benefits to society.

15Easley, Hema. “2 lawsuits challenge Pine Island cell tower.” Times Herald-Record March 9, 2015. http://www.recordonline.com/article/20150309/NEWS/150309383

16In the economics literature, one often thinks of more informative signals as “clarifying.” However, while related, whenever I use this word, I will instead mean a something that reduces the risk of ex post dispute.

17Various studies have shown cell sites reduce property values. See Bond and Beamish (2005), Bond and Wang (2005), and Bond (2007a,b).

147 In particular, the determination of the regulation may favor certain interests over others. Generally speaking, cell siting has a general public benefit by providing better network quality to the community as a whole. However, the costs of the siting are concentrated with the direct neighbors of the site in the form of radiation impacts (real or imagined), reduction in aesthetic neighborhood quality, and reduced land prices. Following the ideas of Mancur Olson, in particular Olson (1971), opposition forces to the sites might then organize more easily than their proponents and thus have their interests catered to more strongly. Perhaps they directly appeal to officials through demonstration or speeches during town meetings. Perhaps there is a lobbying process as in Grossman and Helpman (1994). As shown by Fredriksson (2000), if no one part of the town wants to house a site with public benefits but private costs, such lobbying can lead to the “Not in My Backyard” effect seen in siting landfills, waste sites and polluting plants. Proxying the potential for such collective action using voter participation rates, Hamilton (1993) argues that the effects from collective action are strong enough that firms respond to it. Even without these political effects, officials in towns may overemphasize the benefits of landowners because of public finance. Town governments depend disproportionately on land taxes for revenue, which may be substantially impacted by a new site. If officials disproportionately value fiscal spending relative to the network improvement benefits, officials may be excessively hostile to new sites regardless of political activities. In addition, towns probably do not internalize benefits sites have on other towns, which leads to NIMBY effects across towns. These facilities have benefits for multiple communities, but since that is the case as long as a facility is placed somewhere, every individual community will be more hostile to housing the facility itself than is efficient. These effects have been studied in theory by Kennedy (1994), Levinson (1999), and Hoel (1997), which find environmental standards and taxes will be set too high relative to the optimum. Empir- ically, it has proved difficult to actually determine if observed standards and taxes are too

148 high in these kinds of location problems, though the calibration of Feinerman, Finkelshtain, and Kan (2004) conclude that the political process in Israel results in approximately eﬃcient locations of sites through the lens of their NIMBY model with political lobbying. Thus in general, there is reason to suspect to that welfare may not be optimized by observed regulations in towns. The balance of burdensome and clarifying regulation may thus be ineﬃcient, resulting in either too few facilities overall or too many facilities built in particular towns.

3.3 A Model of Site Location Choice Under Regula-

tory Opportunism and NIMBYism

As qualitatively noted above, regulation of sites has an overall negative effect on firm investment. However, that may be tempered by clarifying regulation to reduce the risk of costly disputes. The levels of these regulations are selected by a potentially distorted political process which may overemphasize particular special interests and ignore externalities on other jurisdictions. I introduce a formal model to bring together all these disparate forces and to discipline empirical analysis of the data. In the following section, I first set up a general model of facility location choice amongst multiple towns and regulation determination by those towns. I then focus on the one-town case, to highlight the impact that regulatory opportunism has on investment incentives. I show why the existence of zoning codes may be beneficial to both towns and applicants, as zoning codes reduce the probability of opportunistic hold-up of the applicants through excessive demands to revisions to initial facility plans. Second, I show how more requirements in a town’s code will lower probability of application, while regulations that reduce the probability of disagreement between the regulator and applicants will raise the probability

149 of application. I then expand the scope to the general case, in which I emphasize the importance of externalities across towns. I show that relative to the case where social benefits are internalized, towns set regulatory standards that are too stringent. This emphasizes the potential welfare losses from regulation as it currently exists, and why clarifying regulations may be ex ante useful. I abstract from the exact level of distortion (if any) in town incentives from the social optimum (however defined) and its origin. As noted earlier there are numerous potential models of distortions, and I lack the data to distinguish from them or even quantify them even if I assumed one in particular. I move forward assuming the objective function of the town as exogenously determined, and simply there may be further welfare losses if this objective function is distorted from the true social optimum. I use essentially a modified version of the model Bajari and Tadelis (2001) use to study procurement contracts with regards to incentives and transactions costs. I focus on the transactions costs, and recast the buyer in that model as a firm who wants to complete a facility for the town, and recast the seller as a town regulator, who will allow construction to go forward if the regulator’s design requirements are met. The regulator is essentially a seller of access to town lands, who demands investment from applicants to reduce the town’s cost of hosting the site. The regulator does this through its zoning code, which acts as a contract. The zoning code’s incompleteness, or how “clear” it is, is modeled as a probability of ex post dispute.18 Goldberg (1976) first explicitly proposed thinking about regulation as a kind of relation-

18 Shavell (2006) and Schwartz and Watson (2013) use an alternative approach to model a contract’s clarity. There, the contract can only distinguish between discrete intervals of the state space. Thus a contract with more intervals is more indicative of the underlying state, or “clearer.” However, clearer contracts (contracts that are more contingent) are assumed to cost more to the actors, which puts a constraint on how clear a contract can be. Based on the legal regime, the optimal level of clarity varies.

150 ship between the regulator and the regulated who have opposing interests which might be best mediated through a long term contract. In the following model, I build on this approach, but I focus not on the repeated interactions but on the reduced incentives for investment. The model therefore resembles much more Salant and Woroch (1991, 1992), where a public monopoly can invest in greater capacity but can be taken advantage of by the regulator ex post investment through requirements of zero variable proﬁt pricing.

3.3.1 Setup of the General Model

Consider two types of risk-neutral economic actors: applicants and regulators. For now, consider a single applicant, but multiple regulators, each representing a single town. Index each regulator by m ∈ M. The applicant has the decision of whether to build a facility or not and where to place that facility if it so chooses. Conditional on the town m, the applicant

can gain a gross profit of Um from building the facility, which is the additional future stream of profits from new wireless service customers net the minimum required costs to build the facility. Building the site induces three additional components of profitability. First, there is a random profitability shock Fm which the applicant incurs upon committing to building in the town m. Fm has a CDF G, which is independent for the applicant between towns. Throughout this chapter, I will treat this shock as possibly positive or possibly negative,

so one could interpret this as either a sunk cost or a sunk beneﬁt. Fm eﬀectively acts as a random “taste” shock for the applicant that makes the applicant more or less predisposed to certain towns. Second, there are costly preparations made by the applicant before con-

struction, Im ≥ 0. This includes search costs for appropriate sites, and testing the site for appropriateness. Finally, there are costly design choices of the facility that are implemented during construction. I call these Pm.

151 I model Im and Pm as having no inherent benefit for the firm - they would rather escape them if possible. However, they benefit the town. Im reduces the costs of having the site located in the town via a cost function Cm(Im), which decreasing and convex in Im (i.e.

0 00 Cm(Im) < 0 and Cm(I) > 0). Pm is directly enjoyed as less intrusive or more aesthetically pleasing design, in other words, a concession to the town. For simplicity, I model this as a pure transfer between the two. In addition, I assume that there is a beneﬁt Bm from the site to the town as long as the site is built.19 This implicitly assumes that all the towns under consideration are nearby to each other, an assumption I will maintain throughout the paper and will discuss in detail more when considering its empirical application.

As explained above, the costs and benefits of the town, in the terms Cm, Pm and Bm, are functions of the underlying preferences of the town’s residents and are all potentially distorted from the social optimum (however that is defined). For example, perhaps the town overemphasizes the utility of homeowners. Pm would focus too much on land price preserving measures like camouflaging sites, and the cost Cm, capturing drops in land prices, would be too large relative to Bm. Given the lack of data that could shed light on what exactly these distortions are, I simply leave the model silent on what exactly the distortions are, but these could be first order sources of welfare losses. Further, in the following theory sections, I will discuss “socially optimal” outcomes: those will be with respect to the town utility functions because I will be ignoring potential sub-town distortions. Relative to the classic buyer-seller model, one can interpret the application for construct- ing the site as an offer to purchase a discrete good. The applicant is the buyer; the regulator is the seller. Cm is the cost of the good; the Pm is the transacted price; and Im is an amount the buyer can invest with the seller to reduce the cost of the good. Such arrangements are not uncommon, for example, Joskow (1985) shows power companies build plants next to

19 A more realistic model would alter Bm based on the location where the site is actually built, but for simplicity I assume it is constant across ﬁnal locations.

152 Figure 3.1: Regulatory Application Timing

Applicant/Buyer Applicant chooses chooses town m or nothing. investment, I0 . Contract executes: m Applicant transfers ρ 0 Date 0 Accept Date 1 Date 2 m Pm(Im) to Regulator. Reject 1 − ρ m Dispute arises: contract is renegotiated. Regulators/Sellers propose No application/sale. zoning code/contract, Pm(Im). Fm learned.

their suppliers’ coal mines.20 The applicant and one of the regulators can contract, and this occurs in relationship with the above investment and payoﬀs according to the following timing. At Date 0, every regulator proposes a take-it-or-leave-it oﬀer of contract Pm(Im) to the applicant which promises

21 a level of concessions Pm for a certain level of investment Im. The applicant learns what

its sunk shock Fm is for each town. The applicant chooses one of the regulators or rejects all of them and gets a payoﬀ of 0. If the applicant accepts some regulator, the game continues,

and at Date 1 the applicant chooses what level Im to invest. At Date 2, I model the incompleteness of the contract by allowing there to be some

sort of dispute about the terms of the contract. With probability ρm, the contract executes

normally, and Pm(Im) is paid to the agent. With probability 1 − ρm the contract must be renegotiated. With contract execution, the following proﬁt π for the applicant and beneﬁt

20 Alternatively, one could think of the regulator as the buyer of network coverage with signiﬁcant costs, and the applicant as the seller of that coverage. The seller then can improve the “quality” of his product through investment. The unusual feature then becomes that the price is paid to the buyer, but ultimately this is simply an issue of interpretation: the two models are isomorphic.

21I use take-it-or-leave-it oﬀers for tractability, but it is easy to extend this framework to Nash Bargaining given a single town m at this stage with no change in qualitative results.

153 V for the regulator are:

πContract = Um − Pm(Im) − Im + Fm (3.1)

VContract = Pm(Im) − Cm(Im) + Bm (3.2)

If renegotiation occurs, then the regulator makes a take-it-or-leave-it oﬀer for concessions

22 Pm. Investment is sunk, so if the applicant refuses, then the project does not occur: Cm,

Bm, Pm and Um are not realized, but Im is still expended and Fm was already realized. Payoﬀs are thus:

πRenegotiation,Reject = −Im + Fm (3.3)

VRenegotiation,Reject = 0 (3.4)

0 If the renegotiation is accepted, then Um, new Pm and Bm are realized:

0 πRenegotiation,Accept = Um − Pm − Im + Fm (3.5)

0 VRenegotiation,Accept = Pm − Cm(Im) + Bm (3.6)

3.3.2 The Value of Zoning Codes: The One Town Case

Before ﬁnding the equilibrium of the full model, it is informative to review what beneﬁt the applicant and the regulator obtain from zoning codes assuming 1) extreme values of ρm and

2) only a single town. For illustrative purposes, I will also let Fm be known to the regulator, which I will relax in the main model.

22It is somewhat awkward to call take-it-or-leave-it oﬀers “renegotiation,” but again, I can have Nash Bargaining with one town m at this stage with no qualitative changes in results.

154 Total surplus is

π + V = Um − Cm(Im) + Bm − Im + Fm. (3.7)

∗ Assume that total surplus is optimized at some Im > 0, so investment has a social beneﬁt.

0 ∗ As a further note, one can show this is where Cm(Im) = 1.

Assume that there is no contracting, so that the game starts at Date 1, and ρm = 0. Then renegotiation occurs every time. Under these conditions, I ﬁnd the standard result for this kind of model:

Proposition 1. Without a contract, no investment occurs.

Proof: From the regulator’s perspective at Date 2, it would like to reduce costs Cm(Im) with the highest levels of concessions Pm possible. Thus it offers Pm = Um + Fm as this will make the applicant indifferent between accepting the offer or rejecting, i.e. it is the largest level of concessions that will induce acceptance of the offer. However, at Date 1, the applicant will foresee that for any outcome of the offer at Date 2, the payoff for the applicant is −I. Thus the applicant’s profit-maximizing investment is 0, which is below the socially optimal level.

Assume that there is contracting, and it is totally complete, ρm = 1. Then renegotiation never happens. Under these conditions, I ﬁnd the other standard result for this kind of model:

Proposition 2. With a contract, the socially optimal amount of investment occurs.

Proof: At Date 2, there are no longer any decisions, so consider the investment decision

0 at Date 1. The applicant simply chooses an Im to optimize Equation (3.1). At Date 0, the regulator makes a take-it-or-leave-it oﬀer of contract Pm(Im). If rejected, the outcome is 0

155 for both parties. Under these conditions, the regulator problem in designing a contract to be accepted is:

0 0 max Pm(Im) − Cm(Im) + Bm (3.8) Pm(Im)

0 0 s.t. Um − Pm(Im) − Im + Fm ≥ 0

0 and Im = arg max Um − Pm(Im) − Im + Fm Im

and Pm(Im) ≥ 0 ∀ Im

and Im ≥ 0

From the regulator’s perspective, there is no reason to leave any rent to the agent, so the

0 0 Individual Rationality constraint of the applicant binds, or Pm(Im) + Im = Um + Fm. Thus the regulator now simply wants the total surplus to be as large as possible since it obtains all

0 ∗ the added rents. Thus it chooses Pm to incentivize the socially optimal investment Im = Im, but leaves the ﬁrm with no rent in equilibrium.

Thus I find the general result that a contract, when complete, is socially beneficial as it can induce the socially optimal investment level. For my application to cell siting, this has the insight that maintaining a well-defined legal code on site application increases preparation by the applicant, and thus benefits society by having applicants find and build sites that are less socially costly. The legal code is a kind of contract: by virtue of being law of the land, all applicants are a party to it as long as they apply. The stipulations on application requirements thus represent ex ante investments that the firm would not undergo since without the protection of the code/contract the regulator could exploit the applicant with burdensome modifications to their original plan to the site. With these insights, let us return to a model with zoning code incompleteness, with

ρm ∈ {0, 1}. I ﬁnd that

156 Proposition 3. With an incomplete contract, the socially optimal amount of investment

∗ occurs if ρmU ≥ Im − Fm but the equilibrium concession is lower than in complete case. If

∗ ρmUm < Im − Fm, then either the investment will be ineﬃcient.

Proof: Taking the results from the no contracts case and the complete contracts case together as the two possible outcomes to Date 2, at Date 1 the applicant now chooses the investment that optimizes the following problem:

0 Im = arg max ρm(Um − Pm(Im)) − Im + Fm (3.9) Im

That is, for an exogenous ρ fraction of the time, the applicant gets a complete contract and obtains its complete contract payoﬀ. However, 1 − ρ of the time, the applicant is held up by

the regulator, get Pm = Um and thus in net only has the sunk cost of investment −I + Fm. The problem at Date 0 for the regulator designing the take-it-or-leave-it oﬀer is thus now:

0 0 max ρPm(Im) + (1 − ρm)Um − Cm(Im) + Bm (3.10) Pm(Im)

0 0 s.t. ρm(Um − Pm(Im)) − Im + Fm ≥ 0

0 and Im = arg max ρ(Um − Pm(Im)) − Im + Fm Im

and Pm(Im) ≥ 0 ∀ Im

and Im ≥ 0

Again, the regulator would like the Individual Rationality constraint of the applicant to

0 0 bind at Pm(Im) = Um − (Im − Fm)/ρm. If this is possible, then a contract can be written to implement this. The difference with the complete contracts case is that at the optimum, the level of concession differs, as a lower concession must be demanded since with some probability the profit Um is appropriated by the regulator. In the complete contracts case,

157 ∗ P (I∗ ) = U − I∗ + F but in the incomplete contracts case P (I∗ ) = U − Im−Fm so the m m m m m m m m ρm equilibrium concession must be lower.

However, this may not be possible at the socially optimal level of investment since if ρm is

∗ very low, then ρmUm < Im −Fm. To satisfy the IR constraint of the applicant, then Pm must be negative in equilibrium, which violates the ex ante constraint that the concessions can only be transfers from the applicant to the regulator.23 The most investment the regulator

∗ can induce is Im = ρmUm + Fm < Im, which is ineﬃcient. If this investment is induced, the regulator optimal contract then sets Pm = 0, which leads to a gain to the regulator of

(1−ρm)Um −C(ρmUm +Fm)+Bm. This investment may be not even better for the regulator than the alternative of no investment being induced. Then the regulator simply forgoes the contract, and the gain/surplus is back to Um − Cm(0) + Fm + Bm. Whatever the case, the level of investment is ineﬃcient.

Unfortunately, there is no way to test Proposition 3 with the data at hand: with some

∗ unobserved regulator heterogeneity across towns in Cm, diﬀerent Im would be optimal, so

any observed variation in Im or Pm can be consistent with Proposition 3.

While I have assumed ρm exogenous, if there were a relatively small cost of setting it for

∗ the regulator, the regulator would at least make so that ρmUm ≥ Im −Fm. Then the optimal amount of investment could be induced from the applicant, and the regulator would collect the optimal surplus. Thus the regulator always has some incentive to set ρm > 0, which explains why zoning codes exist and why towns put some eﬀort into writing a code that is at least somewhat clear.

23This might be possible in practice using subsidies, but to my knowledge there is no example of this.

158 3.3.3 Zoning Codes in Equilibrium: Solution of the Multiple Town

Case

Returning to the multiple town case, I solve the game where towns present contracts simul-

taneously at Date 0. Unlike the previous subsection, I make sunk shock Fm unknown to the regulators. This is to allow across-town NIMBY eﬀects to occur - since I model this as essentially a market transaction, the Coase theorem of Coase (1960) will imply that under complete information, the outcome will be eﬃcient in the sense that the town with smallest (possibly distorted) cost will bear the location.

In any case, unobserved Fm enhances realism of the model as towns set zoning codes that apply to all possible applicants. Zoning codes can change, but infrequently and exceptions are not made ad hoc for particular applicants. Thus the code is written with a distribution of individual circumstances of applicants in mind. Further, I make the simplifying assumption that the codes consist of a contract with a

0 specific pair Pm and Im in mind, which results in Pm = 0 whenever the chosen Im 6= Im. Allowing the scheduling in this case, with private information, can increase efficiency by revealing the applicant’s type, but this makes the problem significantly more complicated and is also unrealistic given the structure of actual zoning codes.

Finally, I take a stand on how ρm is determined. In the previous one town case, Fm

is a known and bounded quantity so there is a minimum ρm that the town would prefer to set to make sure a zoning code satisﬁes its non-negativity constraints. However, in this

formulation, if Fm is unobserved and may be unboundedly positive (which I will assume in

the empirical exercise), then Pm can be positive in equilibrium with a low enough ρm since

some applicant of some suﬃciently high Fm will apply with some non-zero probability. In fact, it will be the case that ρm will be either be set to 0 or 1 since ρm will enter linearly into the proﬁt function without a further cost function.

159 To handle this issue, I assume a cost of not only writing a particularly clear zoning code but also a particularly unclear zoning code. This is realistic since writing a code with specific terms with minimal effort would result in a document that would still be mostly interpretable. To result in disagreement most of the time would require actually more effort

to be purposely obfuscating. So I assume a U-shaped cost function W (ρm) which is strictly non-negative, twice diﬀerentiable and convex (W 00 > 0), but at 0 and 1 approaches inﬁnity

(limρm→0W (ρm) = ∞; limρm→0W (ρm) = ∞). This will keep the optimal level of ρm interior to (0, 1). Under these assumptions, the game becomes very similar to an oligopoly with diﬀer-

entiated products. There is a population of buyers (applicants), who pay a schedule (Pm and Im) and with random taste shocks (Fm) depending on which seller (town) they choose. This builds in an inefficiency already as the competition between the towns will not be as strong as when tastes are not differentiated: since giving better terms Pm and Im does not steal all possible applicants due to the random tastes shocks, an individual town will have a decreased incentive to offer favorable terms. This effect is heightened under the presence of the positive externalities Bm: being rejected means that the town can still enjoy Bm as long as the site is built somewhere, so the town is not punished as much if it does not get the site. There is no way to extract further information about applicants via the contract schedule

Pm since the applicant heterogeneity is additive and sunk once the contract is signed. All

0 applicants choose the same level of investment Im, so the contract type specifying only one

Pm for one level Im is suﬃcient. Moreover, the opportunistic behavior of the regulator still results in the same outcome in disagreement.

Deﬁne the share of the potential applicant population who chooses m as Sm and the share of the population who reject all towns as S0. Formally the payoﬀs and probability of contracting are now:

160 πm = ρm(Um − Pm) − Im + Fm (3.11)

Sm = P r(πm = max{π1, ..., πM , 0}) (3.12)

Vm = {(1 − ρm)Um + ρmPm − Cm(Im)}Sm − W (ρM ) + Bm(1 − S0) (3.13)

Without even solving the game, this setup leads to my main testable implications. Con- ditional on I∗ and P ∗, whether the applicant accepts or rejects the contract (i.e. applying or declines to apply) is a probability. In particular:

∗ ∗ ˆ P r(Accepting {Pm,Im}) = Gm(ρ, U, P, I) (3.14)

ˆ Gm represents the probability that πm is the maximum proﬁt, given the individual distribu-

tion of the shocks Fm and for all potential towns their revenues (U), levels of clarity (ρ) and oﬀered contracts (P, I).

Taking the partial derivatives of this probability with respect to ρm, Pm and Im, one ﬁnds the immediate following proposition

Proposition 4. Given a set of contract terms, Pm,Im, and a level of completeness ρm, the

probability of an applicant applying is increasing in ρm, and decreasing in Pm and Im.

Proof: By the chain rule, the partial derivative of Equation (3.14) with respect to one

of the parameters, X, standing for either ρm, Pm and Im, is the product of derivative of the Gˆ and the partial derivative of the expected payoﬀ for the applicant while under contract ˆ but before investment, ρm(Um − Pm) − Im. Since G is increasing in ρm(Um − Pm) − Im, the sign of the derivative is determined by partial derivative of ρm(Um − Pm) − Im itself, which is positive for ρm, and negative for Pm and Im.

161 Thus, different regulations should have different effects. If a regulation decreases the risk of dispute and opportunistic renegotiation by the regulator, then this is analogous to an increase in ρm and should encourage more applications by firms. If a regulation increases the requirements for preparing the application or the requirements when the facility is opera- tional, then the regulation is analogous to increases in Im or Pm, respectively, and discourages applications by firms. The following sections of this chapter will test this implication by examining location choices made by firms in Connecticut and separating different Connecticut zoning regulations into categories that are analogous to ρm, Im, or Pm. Solving the model (and assuming an internal solution) leads to further conclusions about the impact of across-town NIMBY effects.

Proposition 5. With M > 1, positive externalities of sites, Bm, lead to higher Pm and Im for all towns relative to the welfare equivalent game with no externalities.

X Proof: Deﬁne the elasticity of choosing town m with respect to variable Xm as ηm. The ﬁrst order conditions imply analogues of the Lerner index (Lerner (1934)):

(1 − ρ )U + ρ P − C (I ) − B ∂S0 / ∂Sm m m m m m m m ∂Pm ∂Pm 1 = − P (3.15) ρmPm ηm (1 − ρ )U + ρ P − C (I ) − B ∂S0 / ∂Sm m m m m m m m ∂Im ∂Im 1 0 = − I (3.16) −Cm(Im)Im ηm

These equations immediately highlight the impact of the NIMBY effects. A town need not lose the benefit of the site if it is placed in another town, so each town has a lower incentive to offer a good deal to a prospective applicant because it free-rides off the offers made by other towns. This is represented by the last term of the numerator of the RHS.

Since ∂So and ∂So are positive and ∂Sm and ∂Sm are negative, there is an upward shift in ∂Pm ∂Im ∂Pm ∂Im the equilibrium levels of Pm and Im relative to when Bm = 0.

162 The case with Bm = 0 may have no externalities, but is not total welfare equivalent to the current case, so it is hard to make direct comparisons. A total welfare equivalent case is when there are no externalities, but if a town obtains the site they also obtain an extra

beneﬁt equivalent to all the externalities of all the towns in the externalities case, MBm. Then a town has a payoﬀ of:

Vm = {(1 − ρm)Um + ρmPm − Cm(Im) + MBm}Sm (3.17)

which translates into conditions:

(1 − ρm)Um + ρmPm − Cm(Im) + BmM 1 = − P (3.18) ρmPm ηm (1 − ρm)Um + ρmPm − Cm(Im) + BmM 1 0 = − I (3.19) −Cm(Im)Im ηm

1 > − ∂X0 / ∂Xm , since the loss of the share of m from an increase in I or P will have ∂Xm ∂Xm m m to be divided between the outside option and the other towns. So as long as M > 1, then

the regulator sets higher requirements Pm and Im with externalities than without.

Cross-town NIMBY eﬀects will lead to stricter regulations and lower investments.24 Thus,

24 The comparative statics with respect to ρm are ambiguous since it depends on the exact shape of Wm. The equivalent equations to the externalities case and the total welfare equivalent case are the following, respectively:

0 ∂S0 ∂Sm (1 − ρm)Um + ρmPm − Cm(Im) − (W (ρm) + Bm )/ m ∂ρm ∂ρm 1 = − ρ (3.20) Pm − Um ηm 0 ∂Sm (1 − ρm)Um + ρmPm − Cm(Im) − W (ρm)/ + BmM m ∂ρm 1 = − ρ (3.21) Pm − Um ηm

0 The impact of Bm interacts with Wm(ρm) and Pm −Um which may be both negative and positive depending on the levels of all the variables.

163 NIMBYism is a possible source of total welfare losses in addition to ex ante distortion in the objectives of towns. Further, the solution of the game implies lower elasticity is indicative of stronger regulation requirements. Very low elasticities would be suggestive of, though not evidence for, excessively high regulations. However, as I will note later, I am restricted by my estimation strategy which will not generate the elasticity with respect to the probability of choosing a town or not building at all, but with respect to the probability of choosing a town conditional that some town is chosen in a particular set. Moreover, the estimation strategy also prevents me from scaling implied values of utility with a clear objective standard: not building at all. Thus I have no real way to measure welfare. As a result, I will focus on establishing what the eﬀects of particular regulations are, rather than measure potential welfare losses directly.

3.4 Econometric Speciﬁcation

Empirical application of the model to the dataset requires a few modifications. First, I focus entirely on the probabilities of choosing a town, since I lack the data to proxy for town preferences. The empirical exercise will estimate how covariates impact the incentive to choose a particular town. The profit from each town varies, based on characteristics Xm, like demographics, and on a variety of regulation metrics REGm. A point that bears emphasizing is that the previous model assumed a set of towns which competed for all possible applicants and shared in the benefits of additional wireless service from a new facility. In principle, a carrier or tower company chooses amongst every location in their service area, which might be national. Some locations are thus simply out of range, and even at local ranges distance matters as coverage provision is not as strong to locations farther out from the site. That framing results in a global entry problem with local spillovers, like in Jia (2008),

164 Holmes (2011) or Seim and Waldfogel (2013). In contrast, I reframe the problem so the number of possible towns for a particular facility is small and all the possible towns are in range and have approximately the same benefit. I justify this assumption with the observation that carriers build facilities to fulfill a particular need to plug a coverage gap or decrease the dropped call rate of a particular area. Carriers thus limit themselves to relatively small “search rings” around the center of the affected area for appropriate locations. Many examples of search rings can be found in the records of the CSC application process:

“AT&T established a search ring for the target service area on December 2005. This search ring is centered just north of the proposed site along Gungy Road and has a diameter of approximately one mile. ” —DOCKET NO. 381, Findings of Fact, Connecticut Siting Council, 12/8/2009.

This is consistent with the idea of a global entry decision process with a sort of nesting structure - the firm divides the country into smaller regions - i.e. the search rings - that may or may not need additional service and decides whether to build in that area or not. Once the area is selected, the firm decides where exactly within the region to build, which is the level where I focus my analysis. This cuts down on the computation complexity of the decision process, and thus is both more consistent with anecdotal evidence on the industry decision process and easier to implement.25 I now return to the reduced form profit function of the firm from Section 3.3 but explicitly indicate how REGm, measures of different kinds of regulation, enter the equation:

πm = ρ(REGm)(Um(Xm) − P (REGm)) − I(REGm) + Fm (3.22)

25Zheng (2014) uses a version of this approach to simplify the location decision component of the model in Jia (2008).

165 ρm, Pm, and Im are all function of regulation metrics REGm. Presumably some regulations matter only for one or some of these variables. Some metrics, for example, decrease ambiguity and the likelihood of a renegotiated outcome, and thus embody ρm. Some metrics presumably regulate the design of the base station, and thus aﬀect mostly Pm. Still some metrics reﬂect increased requirements for the report or facility location, and thus increase the level of required ex ante investment, Im.

Fm again appears as the sunk cost component known to the ﬁrm but that is unobserved to others, including now the econometrician. Not building the facility induces no cost and no additional revenue relative to the status quo on average. Given that these Fm are all random variables, this means choosing a facility location is a multinomial choice problem with random utility. Some practical consideration should be considered in the actual speciﬁcation used for estimation. First, it is hard to disentangle regulations with ex ante investment requirements,

Im, or ex post design requirements, Pm. In particular, a design requirement might have to be prepared via significant ex ante investment by the applicant. Second, according to the specification, ρ interacts with both Um and Pm, and thus they should be interacted in the empirical specification. However, it is not practical to do this since there is too little data to estimate so many variables with any power.26 Thus the implemented specification will simplify and make the effect of regulation and non-regulation variables separable as a reduced form, first order approximation. First, a cost function will subsume all the effects of regulation into a single function H(REGm). Regulations that increase H I identify with ρ and regulations that decrease H I identify with I and P . Second, all non-regulation effects are subsumed into two separable functions. Specifi-

26This is even true though I use LASSO for variable reduction later. Interactions of the topics would results in more variables than observations, which even LASSO would not be able to handle.

166 cally, there is additional proﬁt from serving the town where the facility is actually located,

NEAR(Xm), and additional proﬁt from serving the other towns j 6= m ∈ M in range,

F AR(Xj). Xm thus could include demographic variables as proxies for demand for mobile phone service. In addition, NEAR(XM ) represents costs for building in that location that do not matter for regulation, like land values. Under these assumptions, the net additional reduced form proﬁt of building a facility in town m ∈ M is:

M X πm = NEAR(Xm) + F AR(Xj) + H(REGm) + Fm (3.23) j6=m and if no facility is built

π0 = 0 (3.24)

The ﬁrm chooses the option with highest beneﬁt.

3.4.1 Conditional Logit

The above is a multinomial discrete choice problem that can be estimated from outcome data using maximum likelihood. The estimation is complicated by the fact that the above cannot be directly estimated since all town characteristics contribute to the value of all options except the outside option. If one has many observed choices for every town and many overlapping consideration sets, one in principle could estimate this without problems. Since the average number of sites per town is only slightly more than 10, and the consideration sets I use will be geographically based and thus only overlap with adjacent consideration sets, I will handle the issue in a diﬀerent way.

I normalize each reduced form proﬁt function in two ways. First, I assume that Fm is

167 the diﬀerence of two type 1 extreme value shocks - one speciﬁc to m and the decision maker:

m; and one speciﬁc to the decision maker: 0. Thus,

Fm = m − 0 (3.25)

PM I then subtract j=1 F AR(Xj) − 0 from each profit. Since this normalization does not change the differences between the profits, there is no change to decision problem. The profits can thus be rewritten as

π˜m = NEAR(Xm) − F AR(Xm) + H(REGm) + m (3.26)

M X π˜0 = − F AR(Xj) + 0 (3.27) j=1

Under type 1 extreme value shocks, the applicant’s choice over all choices is a multinomial logit. As has been recognized since Luce (1959), the multinomial logit obeys the independence of irrelevant alternatives, such that adding an additional option to a choice set should not change the probability of choosing an original option conditional that something in the original choice set has been chosen. Conversely, subtracting an option should also not change the conditional probability, so parameters will be consistent if one uses data eliminating options and eliminating individual who choose that option. This has actually been the strategy of tests of IIA ﬁrst proposed by McFadden, Train, and Tye (1977); here I simply assume the IIA property since my data is not large enough to support estimation of correlation in the taste shocks between towns.27 This approach is also consistent with nested logit if the inside options are grouped together in a nest - in that model, only choices across options between

27This could be done using a multinomial probit in which the options are available for many individuals. However, the current application has each town only considered a few times.

168 nests are not IIA, so estimating logit within the nests will still be consistent. Under IIA, then the probability of choosing town m conditional on there being a facility built is

exp(NEAR(X ) − F AR(X ) + H(REG )) P r(m) = m m m (3.28) PM j=1 exp(NEAR(Xj) − F AR(Xj) − H(REGj))

In additional to avoiding estimating the beneﬁt values for the outside option, I can estimate the model using only observed facilities. Thus I also avoid the issue of not observing some decisions because the ﬁrm ultimately decided not to build a facility at all. This approach means that I will not be able to get at the rate of substitution between building a site and not building a site - I simply assume that at the level of choice I consider the site will always be built. This limitation is unfortunate, as the model shows there is connection between the elasticities of the unconditional probability of choosing a town and the level of regulation. Thus my analysis will inherently be limited to comparing the level of regulation between towns and not the welfare impacts themselves.

3.4.2 Potential Omitted Variable Bias

The employed speciﬁcation is admittedly crude in several respects which lead to potential omitted variable bias. First and foremost, I can only control for the proﬁtability of the location up to the

observed characteristics Xm and REGm. Many idiosyncratic factors in citing in a particular town will not be controlled for simply because those are not captured in available covariates.

It is therefore very important that I employ a rich set of controls in Xm. Second, in employing the NEAR and F AR functions I abstract from the exact distance from borders. As noted before as locations become farther away from the site the strength declines, so distance does matter. In particular, for a particular town m that is not chosen,

169 this will mean that depending on what town does gets chosen the F AR function for m will be diﬀerent. To illustrate this point more formally, imagine the consideration set including four square towns, {A, B, C, D}, that meet in a single corner.

Figure 3.2: Four Towns

A B

C D

For concreteness, assume facilities must be placed in the center of towns. Placing the facility in the center of A would be equally far from B and C, but farther from D. Then relative to placing in B, the gain from D would be greater since it would be adjacent, and

C would be less since it is farther away. Thus for different chosen m, F ARj can differ. Finally, another potential unobserved variable is the competitive environment. Building another facility in a locale may not be very profitable if the market is already saturated. Thus controlling for the environment seems to be important, but this is complicated given the complex nature of competition in this industry. For example, one might estimate a multiple firm entry game, but this is computationally intense, especially with a combination of vertically integrated firms, tower companies and public entities all involved as players.28

3.4.3 Border Discontinuity

These omitted variables concerns can be partially addressed via a border discontinuity strategy. I make the observation that if a facility is near a border, the town across that border

28Probably the most similar paper to attempt this in the study of regulation is Gowrisankaran and Krainer (2011), which looks at ATM rollout with respect to diﬀerent surcharge fee limits.

170 would have also been considered as a location. Because this industry explicitly uses small search areas in the planning stages, only that town, or towns very close, could be part of the consideration set. Thus I limit the consideration set of locations for each observed site to only locations in towns very nearby. To illustrate why this strategy may help with identiﬁcation, consider the case where there is an omitted variable ξlm in the reduced-form proﬁt function.

π˜m = NEAR(Xm) − F AR(Xm) + H(REGm) + m + ξlm (3.29)

The index lm represents the local variation in profitability across within-town locations l. The best location in town m for the site is l, and for some town m0, the best location is l0. These terms naturally contain town specific effects regarding demand for wireless, and so proxy for the ignored differences in the F AR functions and competitive environment.

0 0 If the location lm and l m are very far away, ξlm and ξl0m0 are likely to be very different. However, consider my assumption that sites are only chosen between adjacent towns and indeed only from locations that are right next to each other. Then the unobserved variation is likely to be almost identical, so ξlm ≈ ξl0m0 ≈ ξmm0 . Discrete choice models only estimate differences of utility between choices, so essentially ξmm0 is differenced out as a choice specific fixed effect. This is true also for components of the NEAR and F AR functions that proxy demand, so almost all the variation in profit should be in the thing that discontinuously varies across the border: costs from regulation. As an added benefit, this approach reduces the bias from ignoring the interactions between revenues and the regulations that affect the risk of disputes, since the revenues are not much different across options.29

29 This does very little to address that regulation aﬀecting the risk of disputes (i.e. ρm) suggests interactions with regulations on ex post operation and design of facilities (i.e. Pm).

171 Given the conditional logit formulation of the problem, I need only use these observed sites to achieve consistent estimates. To implement this strategy, I create my sample for estimation according to the following algorithm. Noting that 0.5 miles is a common search ring radius in the CSC regulatory record, I retain the subsample of 2013 facilities within 0.5 miles of a border between two Connecticut towns.30 I then assign to each site i ∈ N those all

towns at most 0.5 miles away and consider them the consideration set Mi for each decision to build site i. I therefore use the joint set of {N,M} for estimation. This approach is not perfect - any omitted variable bias attributable only to regulatory differences but not present in the zoning code will not be corrected by this approach. Thus even if regulation variables have significant impact, they may reflect underlying differences in attitudes towards sites or regulatory efficiency at the town level. For example, one can imagine that large towns will have much more resources than smaller towns and thus be much more sophisticated and thus easier to site in. Given that large towns tend to have shorter or even no zoning code sections specifically referring to telecom, this would bias the estimated importance of regulation upwards. This problem is worse considering that my approach will naturally sample large towns in consideration sets since they have larger borders, so there is a potential selection effect. Further, my areas are actually quite large for most of the literature. Bayer, Ferreira, and McMillan (2007) is partially about finding a technique that will work with distances to borders smaller than 0.3 miles due to concerns about the effectiveness of the strategy at these distances. Thus, there may be significant unobserved variation between the locations considered even at the 0.5 miles level. However, due to sample size concerns and knowing that the industry uses search rings of approximately this magnitude, I stick with buffers of 0.5 miles.

30There are many sites along state borders, which I cannot use unfortunately.

172 A final point about this estimation strategy is that there is a further peculiarity of the Connecticut regulatory system - not only does the town in which that site is built receive a copy of the application and requires a consultation, but also any town 2,500 feet away. This is about half a mile so every town in the constructed consideration sets will receive an application. Thus it may be the case that no matter the choice in the consideration set, one will incur the regulatory costs of all the towns. However, it appears in practice that towns are far less likely to mind when sites are not within their borders. Looking at the CSC’s Findings of Fact reports available for all site applications since 2000, one finds 29 recorded cases in which multiple towns have been informed of the application, and in only two cases did the secondary town object or have critical comments.31 Thus it is likely the applicants will not feel the full force of regulatory burden from the neighboring town when applying, and the difference in regulatory burden will still have strong effects. However, this will be another reason, along with the presence of the state level review, that these Connecticut-specific results will be likely be of lower magnitude relative to other states.

3.5 Data

Again I use the base station location data from the Connecticut Siting Council. For this exercise I limit the data to base stations from 2013 to approximate the long-run equilibrium investment. Table 3.1 categorizes facilities with regards to how many towns are within 0.5 miles and the distribution of the size of consideration sets. It also breaks down the locations by towers

31This record does not include rooftop sites and older sites that do not have their proceedings online, so the number is relatively low. There are also two applications that were withdrawn and did not reach the review stage so a Finding of Fact was not written.

173 Table 3.1: Facilities by Consideration Set

Towns Within All Towers Rooftops AT&T Verizon Sprint T-Mobile MetroPCS 0.5 Miles Sites

1 1,150 554 596 552 439 435 478 232 2 470 222 248 211 195 181 189 104 3 79 43 36 47 29 32 34 17 4 4 2 2 3 2 2 3 2

2+3+4∗ 553 267 286 261 226 215 226 123 1+2+3+4 1,703 821 882 813 665 650 704 355

* This the set of observations used for estimation. and rooftops, which might have differential effects since towers will require much more drastic construction, and by whether there is one of the five carrier’s base stations on that facility. There is also a breakdown by carrier presence on the sites, to proxy for the base station location decision by carrier and to detect potential firm variation in costs, regulatory or otherwise. About a third (553) of facilities qualifies for estimation, which is approximately the area which I am using for estimation. There are still plenty of facilities with consideration sets of two, far less with three, and very few with four. The data is split evenly between towers and rooftops, with rooftops having a slight edge. Only between a half and a third of sites are eliminated when only looking at one of four national carriers, reflecting how much collocation there is. The most sites are lost when looking at MetroPCS because it only entered in 2009 and has limited coverage for rural areas. This data is matched to a dataset of zoning codes for almost every town in Connecticut. As mentioned in Chapter 1, each zoning code was either collected from the town website or acquired from the town’s town clerk. There is a “section” about regulating telecommunications present in 139 out of the 169 towns. Cornwall has its telecom regulations in an addendum not included with the main document. Thus I do estimation on 168 towns in-

174 cluding the 30 that do not have telecom regulations at all. Note again that of the towns that do have the section, the mean length is 13,611 characters and 2,487 words. Finally, I use data from the 2009-2013 American Community Survey to construct controls for the later regressions. These are town level demographic variables that are designed to pick up any unobserved heterogeneity that is not controlled by the regression discontinuity design. As noted earlier, larger towns may be more easily sited in due to greater sophistication of more urban towns. These towns may also be more profitable markets since they are more densely populated. I include log population and log area to proxy for larger and denser towns. If, as suggested in the zoning literature, that towns disproportionately care about the concerns of incumbent homeowners or land tax revenue, then it is important to control for this via measures of wealth, home value and home ownership as these could proxy for the attitude town officials have for telecom facilities that I do not pick up in my regulation measures. I include then median household income, median home value and the fraction of owner occupied units. I also include the fraction of the population with a post-secondary education as more educated populations may be more liable to use advanced technology like cell phones. In addition, I include percent of Hispanic and black residents as potentials controls for possible differences in cell phone demand by different demographics. Political environment might also be important, as this may affect how the application is treated during the application process. To capture these concerns, I use data from the Connecticut Secretary of the State.32 A town with a mayor may be more easily negotiated relative to a town that is controlled by town hall meetings. I therefore include dummies for

32See http://www.ct.gov/sots/site/default.asp.

175 whether a town has an executive officer (mayor/manager), whether it has a town council, whether it has selectmen, which a traditional form of representative government peculiar to New England towns, and whether it has town hall meetings decide issues. Like Hamilton (1993), I also include percent of turnout for municipal elections from 2006 to 2012 to measure the potential for collective action, and the percent who voted Democratic in 2012 presidential election, as the Democratic party is friendlier to regulation than its main rival, the Republican party. Summary statistics for the town demographics used for estimation are listed in Table 3.2. In general, Connecticut towns are wealthier, more educated and denser than the rest of the U.S. population. Median home value is quite high: some towns have median home values exceeding $1 million.33 The minority population is relatively small, and there are towns with no blacks or Hispanics at all. Politically, there is significant variation in party affiliation, but the voters of the average town are split almost fifty-fifty between democrats and rival parties.34 Town government varies quite a bit too: 63 have executives, 53 have town councils, 108 towns have selectmen, and 56 have town hall meetings.

3.6 Measuring Stringency of Regulation

Measuring the extent of regulation is diﬃcult, as this generally requires the interpretation of legal statutes by the researcher. When the regulation is not or only inconsistently quantitative, translating regulation into a quantitative measure is diﬃcult. For example, often a

33This actually creates a data problem - the Census does not report the actual housing value when the home value exceed $1 million; it just makes note of it. I therefore include a dummy for when this is true in my regressions.

34This does mean that Connecticut leans Democratic since third parties, though very small, split the remaining vote with the Republicans.

176 Table 3.2: Town Characteristics

Variables Mean Median SD Min Max

Population (000s) 21.14 12.60 25.28 0.95 144.45 Area(km2) 74.21 70.97 32.67 13.09 159.46 Median Household Income (000s) 83.85 80.66 26.64 28.93 213.42 Median Home Value (000s)† 317.61 291.30 118.93 158.3 906.90 Median Age 42.83 43.00 4.52 21.30 54.00 % With Post High School Education 45.16 47.00 8.48 22.60 62.57 % Black 3.76 1.33 7.37 0.00 54.66 % Hispanic 6.34 3.40 7.52 0.00 43.05 Mean % Turnout 39.10 39.09 8.65 18.30 60.80 % Voted Democratic 53.63 51.91 9.92 34.36 93.24 Observations 168

† Four towns have median home values more than $1 million, and the Census only reports that number for them. These towns are therefore left out when calculated these summary statistics.

177 wireless telecommunications regulation will have a maximum tower height. When it does, sometimes it will be an explicit height, but sometimes it will be based on a formula, such as one half of the minimum allowed distance to the next property, the setback, of the underlying zone.35 Thus it becomes difficult to quantify this regulation as the setback will vary by parcel and require reference to other parts of the zoning regulation. One way to approach this measurement problem is through the creation of an index. The Wharton Residential Land Use Regulatory Index (WRLURI, Gyourko, Saiz, and Summers (2008) ) is an index created from responses of a survey of city managers. The index is the first factor from a factor analysis of the responses and has been used extensively in the literature.36 Creation of this kind of measure requires a large amount of data on multiple dimensions of regulations not feasible in most contexts. It also collapses information into few variables, so in the case of the WRLURI, implications of particular regulations must be inferred from the factor loadings alone. Alternatively, one could rely on indicators for presence of a regulation at all, as detecting whether a regulation exists is far less ambiguous than determining how exactly strong it is. For example, Glaeser and Ward (2008) use indicators on whether town regulations exceeded state standards regarding septic systems, wetlands and subdivisions. This approach does have several drawbacks. First, it does not allow for continuous variation in regulation, just an indicator. Second, indicators still depend extensively on researcher judgment. Glaeser and Ward (2008) mitigate this issue by allowing local officials to review their variables, but this is not always possible. The procedure also requires manual analysis of each code, which is not only time-consuming but also sometimes infeasible.

35This regulation takes this speciﬁc form to make sure if a tower falls then it is far away from the next property.

36For example, Suzuki (2013) uses it in his study of the Texas lodging industry. Diamond (2015) uses it as determinant of housing supply in her study of skill sorting by location.

178 For this study, I will use the indicator approach for an initial analysis but use a new approach in a second. To produce my indicators I manually examined all 138 wireless telecommunications facilities sections found for the 168 towns. I looked for particular kinds of regulations that seemed potentially important and created an indicator if I found them in a particular regulation. This includes:

• Consultant - The town may require review of the plan or monitoring by an independent consultant, like an engineer, a cell siting specialist, or even a medical professional. This is often at the applicant’s expense, and represents both pecuniary costs and opportunity costs in further review time.

• Glossary Present - A glossary is often included to deﬁne the unique jargon used in the section. This may increase clarity by making the meaning of terminology clearer.

• Higher Authority Acknowledged - Whether the document recognizes the authority of some state or federal body, like the CSC or FCC, or a federal law, like the National Environmental Protection Act (NEPA). This may reduce disputes since there is now an explicit, well-understood reference that overrules many objections the town could have.

• Landscaping - Towns often require some sort of landscaping, like a wall of trees to block views of the facility from the street. This serves no beneﬁt to the site’s eﬀectiveness, and so it is pure extra cost to the applicant.

• Public Hearing - Small New England towns have a long history of using town meetings in their governmental organization, so it is common for Connecticut to require a public hearing over applications. As it is more likely to ﬁnd at least some individuals very strongly opposed to cell sites in a general meeting rather than a small group of government oﬃcials, this could lead to increased opposition to the application. In addition,

179 it is another procedural delay.

• Radiation Requirements - The level of radio frequency radiation (RFR) is mentioned in some way. This could indicate hostility to cell sites as being dangerous to public health, but generally RFR is mentioned in conjunction with the FCC standards of regulation. Imposition of an objective standard lessens the risk of disagreement, so this could also have a positive impact.

• Extra Radiation Requirements - To separate out the potential clarity eﬀect from the Radiation Requirements indicator, this indicator denotes whether there are extra requirement with regards to radiation, such as annual reports on radiation levels to the town. These extra measures should therefore have an unambiguously negative eﬀect.

• Removal Bond - Towns often require ﬁrms to pay an upfront amount to establish a bond that will be used in case a ﬁrm decides to abandon a site and the town has to dismantle it itself. These can be in the thousands of dollars.

• Size Restrictions - Limits on tower height, lot sizes, and setbacks from adjacent property lines or the street are all designed to minimize impact from a collapsing tower and their visual profile. These make it hard to find appropriate properties for sites and thus would increase cost of the final property used and lengthen property searches. Further, the site might have to be smaller, reducing its effectiveness.

• Visual Impact - Towns often express a concern about maintaining the aesthetic character of neighborhoods and scenic vistas, especially in historic, New England villages. Cell sites are modern, industrial facilities so inclusion of these aesthetic concerns may signal greater hostility to cell sites.

The counts of these variables are reported in Table 3.3.

180 Table 3.3: Coded Regulation Measures

Variables Count Expected Sign

Consultant 30 (-) Glossary Present 64 (+) Higher Authority Acknowledged 123 (+) Landscaping 103 (-) Public Hearing 50 (-) Radiation Requirements Mentioned 69 (?) Extra Radiation Requirements 43 (-) Size Restrictions 127 (-) Removal Bond 72 (-) Visual Impact 124 (-) Towns with Telecom Regulations 138 Total Towns 168

181 The second approach implements “topic modeling” from the computational linguistics literature using the Latent Dirichlet Allocation (LDA) algorithm introduced by David M. Blei and Andrew Y. Ng and Michael I. Jordan (2003). Often in quantitative analysis of texts, one would like to detect “topics,” words with shared meaning, and find how frequent these topics are in the documents. In the social science literature, LDA has been used to do this for the content of the U.S. Congressional Record (Quinn et al. (2010)) and the Federal Open Market Committee (Hansen, McMahon, and Prat (2014), Fligstein, Brundage, and Schultz (2014)) . In my context, different kinds of land use regulations can be thought of as different “topics,” and increased frequency of their appearance in a zoning code can indicate different levels of stringency. Topic modeling has several advantages over the indicator approach of measuring regulation. LDA is less subjective in the sense that it can run entirely unsupervised so can generate sensible topics without researcher judgment.37 Further, since it is unsupervised, it almost always feasible to run even if the entire body of text is extremely large. Finally, it naturally outputs continuous index-like measures, which I will exploit in the empirical analysis. To develop LDA formally, I follow the presentation of the methodology, including notation, from Hansen, McMahon, and Prat (2014) below.

3.6.1 The Model of LDA

Topic modeling conceives of words, like “telecommunications” or “regulation,” as distinct members of a universal set of all words. This universal set is called the “vocabulary,” V . A collection of realized words from the vocabulary is called a “document.” Say one has

37I caution that LDA cannot be considered totally “objective” since much of its implementation is based on ad hoc subsample selection and strong parametric assumptions, and the topics must be given an ex post interpretation by the researcher. It simply has the beneﬁt that it does not rely on researcher interpretation of the text directly.

182 a collection of many documents as data. Call this collection of documents the “corpus,” and denote the number of documents in the corpus as D. Indexing each document by d ∈ D allow the documents to vary in length of words, Nd. Topic modeling posits that the documents were created by a speciﬁc data generating process. First, for all documents d, take Nd as given. Then each document d is thought of as being made of 1, 2, ..., Nd − 1,Nd “slots.” Assume now that there are K discrete distributions of the simplex of V . Call these distributions “topics,” which informally represent diﬀerent “ideas.” In my context, regulations about the visual impact of towers will often mention related words like “view,” “sight” and “aesthetics,” and thus roughly map onto a distribution that puts a lot of weight on these

V words. Formally, denote topics as βk ∈ ∆ . The documents are assumed to have been created in the following way: each document

K d has its own discrete distribution over possible topics, θd ∈ ∆ . Each slot of the document drew from that distribution and realized a topic zdn. Then a word was drawn from that

topic’s βzdn to realize a particular word. The result is the document, a set of words, wd =

(wd1, wd2, ..., wdNd−1, wdNd ). The corpus is D instantiations of this process, with given Nd. In principle, one could estimate the most likely distributions to produce the observed body of text. The direct likelihood of the observed text has the form:

N Y X P r[zn|θ]P r[wn|βzn ] (3.30) n=1 zn

This is intractable, so I will impose structure on the distributions of topics θd and the distribution of words in a topic βk. The Latent Dirichlet Allocation assumptions are a popular choice: documents are assumed independent and the latent distributions of topics and words in topics are symmetric Dirichlet with hyperparameters α and η. α governs the topic distribution, and η governs the distributions of words in topics. The new likelihood of

183 the observed text is now:

K N ! Z Z Y Y X ... P r[βk|η]P r[θ|α] P r[zn|θ]P r[wn|βzn ] dθβ1 . . . βK (3.31)

k=1 n=1 zn

This equation is an object that can be computed and is used for estimation. The LDA model, while tractable and ﬂexible, does have signiﬁcant shortcomings. In particular, this is a so-called “bag of words” model, such that a document is simply a collection of word frequencies. The order of the words is not utilized and thus meaning of the documents depending on syntax is lost. LDA also assumes independent documents, which is problematic if documents are sections from a single text or body of work, as it is here. Finally, LDA also assumes that every word is in every topic, so every topic will be present in every document. Thus if presence of a topic in a document (being non-zero) is the variable of interest, one cannot directly use the estimated distribution since it will almost always be nonzero. Some other transformation of the estimated distribution must be used instead.

3.6.2 Estimation of LDA

I estimate the LDA likelihood of the entire body of the regulation texts using the Gibbs sampler approach of Griﬃths and Steyvers (2004). This procedure is a Monte Carlo Markov Chain approach, which assumes a random starting distribution, estimates a new distribution from the outcome of the draws and the observed outcomes, and draws again from the updated distribution. Over time, the chain of draws will converge to a steady state distribution that will mirror the true marginal distribution. For estimation, I follow Hansen, McMahon, and Prat (2014) extremely closely. To start, I deﬁne the documents in question as sentences observed in the wireless telecom-

184 munications regulations sections.38 Ideally, documents should represent a complete, unified expression of ideas that are not correlated with each other. A more natural document might therefore be the entire section or a subsection. However, using the entire section is impractical as there are only 138 sections in the data, and the topics are identified via variation across documents.39 An alternative is using the subsections, which roughly matches a paragraph in prose writing. However, subsections are difficult to separate out in the data, since regulations are generally reported as nested lists with very heterogeneous formatting styles. This is made even harder by the text data being recorded inconsistently with regards to whitespace, so that a carriage return which might indicate a meaningful separation of subsections in one document might indicate where the line space ran out in another document. In contrast, sentences are easy to define using periods followed by whitespace as breaks, and this results in 14,900 sentences for the dataset overall. Of course, sentences as documents makes the independence assumption across documents even less likely, but given the data limitations sentences seem to be the best I can do for defining documents. Words are therefore defined as words in each sentence, and these form a raw vocabulary of the list of all words in all documents. However, the raw vocabulary is not ideal for estimation so is improved in several ways. The full details are described in Appendix C.1.1:

1. All text in the corpus is transformed to lowercase, since capitalization does not usually convey meaning.

2. Very common two or three word compound words or phrases which represent a single

38Since I have a proposed regulation for the town of Hartland, which was never implemented, I include in the topics estimation since the added text should include information on what topics should be like. I do not use the Hartland regulation in the later regressions.

39I did attempt to estimate the topics by section: the estimation simply did not converge.

185 idea are combined into a single word. For example, “visual impact” becomes “visualimpact.”

3. Non-alphabetic characters and almost all one or two letter words are removed since these are unlikely to convey much meaning.

4. Extremely common words across documents like “the” are removed since their ubiq- uity mean they have no usefulness in distinguishing topics. What exactly qualiﬁes a words for this kind of removal is not objective so I use a common list of these “stop words” from the literature plus some words that are overly common in my context, like “Connecticut”.40

5. Similar words are converted to roughly their grammatical “root” or “stem” using the Porter Stemmer from Python’s natural language toolkit. This controls for diﬀerences due to grammatical form, not content - for example, ‘fence” and “fencing” become “fenc”.

6. The remaining stems may not be informative since after conversion they may have turned out to be overly common across documents like stop words, or are extremely rare. Using a joint measure of these two problems common in the literature, I retain only the top 1,700 stems.

The literature calls the output of this process “tokens” since after being turned into stems they are no longer necessarily real English words. I will match this convention though it should be noted that tokens are meant to map onto the words wd presented in the formal deﬁnition of LDA from Section 3.6.1.

40 The main list the same used in Hansen, McMahon, and Prat (2014) and can be downloaded from http: //snowball.tartarus.org/algorithms/english/stop.txt. I describe in Appendix C.1.1 what the other words are.

186 The estimation is performed on this transformed dataset. I leave the exact details of the estimation to Appendix C.1.2, but roughly the procedure is as follows:

1. Assign every token-slot pair to a random topic.

2. For each token and slot pair observed in each document, zdn:

(a) Using counts of observed tokens and assigned topics to calculate a posterior distribution of topics for that slot conditional on the observed token in that slot.

(b) Draw a new topic for that slot using that distribution.

3. Repeat many times until convergence.

The estimation procedure depends on knowing what K, α and η are: these cannot be estimated from this procedure. Optimal K can be calculated under some computationally intense methods, but for simplicity I choose K = 30 as the results seem to yield a balance of parsimony and interpretability. Fewer topics are too broad and are hard to interpret, while more than thirty result in topics that are very narrow and hard to all include in the empirical exercise. In Appendix C.3, I show this with estimates of topics with K = 50, that have some topics that are very similar to K = 30 topics. Using 30 regressors for the empirical exercise is still rather awkward, so I further reduce the topics via Lasso, in which I detail in Section 3.7.

50 Given K, I follow Griffiths and Steyvers (2004) and set α = K and η = 0.025. In addition, combined words are afforded only a single topic slot for estimation, and all slots will be removed for words which were deleted. A Monte Carlo Markov Chain (MCMC) such as this does not converge to an optimum - rather it improves its fit until it begins iterating around a stable distribution. The object

187 Table 3.4: Perplexity of the Gibbs Sampler Chains

Chain 1 2 3 4 5 6 7 8 9 10 11

Mean 350.65 350.91 350.92 351.38 351.37 351.46 351.35 350.79 351.99 351.25 350.84 S.D. 0.49 0.48 0.50 0.50 0.51 0.48 0.50 0.58 0.48 0.47 0.46

Chain 1 is the chain used for estimation - it has lowest mean perplexity. Twenty-four other chains were run, and

here I list the ten other that have means closest mean perplexities (< 352)

of analysis is therefore not a set of parameters but a produced chain. The distribution estimated is taken over the draws in the chain. It is an open question about when to stop the chain, though the textual analysis literature has a convention. Fit in textual analysis is often measured via the perplexity, a function of

the estimated distribution, the counts Nd,v of token v in document d and Nd the count of tokens in document d, overall:

! PD PV N log(PD K θˆkβˆv) exp − d=1 v=1 d,v k=1 d k (3.32) PD d=1 Nd

Trial and error with many chains finds that the topics often converge to a perplexity of about 391 after about 20,000 draws. However, it is the case that the distribution can get “stuck” at about 392 or higher, and sometimes a chain can get stuck and then converge to 391. 20,000 draws takes a relatively long time (about 45 minutes for each chain), so to balance finding the optimal distribution and testing robustness of the procedure, I ran twenty-five 40,000 draw chains, with 20,000 draw burn-ins. This took about one and a half days of computing time. To reduce serial correlation amongst draws, I used every fiftieth draw to estimate the distributions and find the average perplexity. Eleven chains achieve something close to 391 in average perplexity so for estimation I use the chain with the lowest one.

188 3.6.3 Estimated Topics

Table 3.5 displays the 30 estimated topics by their five most frequent tokens. Many topics are immediately recognizable from the top five tokens as referring to previously mentioned kinds of regulation. Topic 2 and Topic 10 seem to both represent Size Restrictions, as the former seems to worry about the impact on “adjac(ent) propert(ies)” and the latter has “setback” and “propline,” i.e. “property line,” as top tokens. Topic 7 and Topic 12 seem to proxy for Visual Impact but each with a different emphasis - Topic 7 directly addresses the visual concerns, putting strong weight on how “design” might have “potential adverse visual” impacts, while Topic 12 talks mostly about concerns to “preserv(e) histor(ic) land(s)” and “area(s)” from “development.” Topic 8 might be related as well as one of its top tokens is “visibl(e).” Having the words “remov(al) bond” makes a clear connection between that kind of regulation to Topic 13. Topic 15 and Topic 28 seem to match the Consultant measure, as the former talks about “licens(ed) engin(eers)” and the latter about how “town(s) may” require “review” by “independ(ent) consult(ants).” Topic 17 looks to match Radiation Requirements, as it includes “compli(ance)” to “FCC regul(ations)” on “RFR standards.” Topic 18 may map onto regulations about tower “design” to maximize collocation so “addit(ional)” carrier may be “accomod(ated).” Topic 21 seems to match the common inclusion of preference lists for tower locations (“telecomtower”), where the preferences are given a “number(ed) order” in the interest of meeting the public “need” for coverage but trying to “reduce(e)” the number of towers as much as possible.

189 Table 3.5: Estimated Topics

Topic Five Most Frequent Tokens Share(%) Tokens Present 0 co, prefer, feasibl, altern, descript 3.16 22.58 102 1 permit, follow, siteplan, specialpermit, regul 3.86 25.23 122 2 properti, within, subject, adjac, public 2.90 16.13 85 3 color, design, materi, detail, possibl 3.91 29.40 120 4 within, plan, show, map, coverag 3.19 24.64 109 5 provid, town, adequ, demonstr, capac 2.68 14.30 64 6 oper, owner, day, date, prior 2.54 11.90 78 7 minim, design, advers, visual, potenti 4.47 36.70 116 8 elev, base, visibl, point, height 2.74 15.00 90 9 commun, commerci, district, radio, residenti 3.71 20.43 105 10 zone, residenti, setback, equal, proplin 4.29 27.51 119 11 regul, servic, person, provid, provis 3.08 14.00 68 12 area, histor, develop, land, preserv 3.11 16.33 98 13 remov, abandon, upon, month, bond 3.32 24.27 118 14 build, equip, area, accessori, contain 3.87 32.31 111 15 engin, plan, licens, document, interfer 3.42 20.40 97 16 will, report, telecommun, indic, system 3.44 23.85 99 17 standard, rfr, compli, regul, fcc 3.26 19.53 107 18 addit, may, design, inform, accommod 3.66 26.98 105 19 limit, util, gener, instal, servic 3.18 20.89 97 20 approv, time, unless, specialpermit, may 3.27 15.32 87 21 need, telecomtow, order, number, reduc 3.46 25.21 93 22 mount, build, ground, roof, provid 3.49 21.51 112 23 power, part, consid, signal, type 2.19 9.94 58 24 fcc, light, feder, ﬁle, state 3.06 16.80 116 25 construct, monopol, design, support, specif 2.97 23.37 92 26 landscap, fenc, screen, surround, tree 3.46 22.64 113 27 follow, receiv, commun, transmit, telecommun 3.24 25.59 78 28 review, may, town, consult, independ 2.68 15.51 91 29 feet, height, exceed, inch, diamet 4.39 32.03 116 “Share” is the mean share over towns, “Tokens” is the median number of estimated tokens over towns, and “Present” is the number of successes of the presence proxy.

190 Topic 26 appears to clearly proxy for Landscaping, including the word “landscap(ing)” as one of its top words. Other topics have obvious correlates in the zoning codes, but were not thought of in the initial set of Coded measures. Topic 0 appears to refer to how some towns require evidence from firms that they investigated “feasibl(e) altern(atives)” to the proposed new site, such as “co(llocating)” on another site. Topic 4 appears to be about the requirement for the site “plan” to have a “map” that “show(s) project(ed) coverag(e).” Topic 5 appears to be about the applicants “demonstr(ating)” without the new site a carrier cannot “provid(e)” the “town” with “adequat(e) capac(ity)”. Topic 19 discusses the regulation of power at sites, particularly about the “limit(s)” on “servic(e)” from “util(ities)” and “install(ation)” of “gener(ators).” Topic 24 is about “state, feder(al),” and “FCC” regulation about how sites must be “light(ed).” Other topics are much more general. Topic 1 appears to collect words meaning “permit” or “plan”. Topic 29 collects length measure words like “feet” or “inch” while Topic 6 collects times measure words like “day” or “date.” Still, these could have significant effects on regulation strength: the greater inclusion of measures words could mean clearer standards on equipment dimension or expected regulation delays. Some topics are not very obvious from the first five tokens but have clear analogues in the zoning codes looking to the first twenty or so tokens. Topic 11, associated with the words “regul, servic, person, provid, provis,” appears to be picking up a certain regulation that acknowledges federal authority. “Effect,” “consist,” “prohibit,” and “telecomact,” which is my multigram replacement for the “Telecommunications Act of 1996” are the 6th, 7th, 8th and 9th most common tokens in this topic. Taken together, Topic 6 is now reminiscent of sections like 10K.2 from the Marlborough zoning code:

“Consistency with Federal Law: These Regulations are intended to be consis-

191 tent with the Telecommunications Act of 1996 in that: a) they do not prohibit, or have the eﬀect of prohibiting, the provision of Personal Wireless Services; b) they are not intended to be used to unreasonable discriminate among providers of functionally equivalent Services on the basis of the environmental eﬀects of radio frequency emissions to the extent that the regulated Services and Facilities comply with the FCCs regulations concerning such emissions.”- My emphasis.

Thus this appears to map onto the Higher Authority coded measure quite well and should reduce the risk of regulatory dispute when present. In addition, Topic 20 appears to be a proxy for a clause putting a time limit on site approval - if the time limit passes, the applicant needs to apply again. For example, here is Section 17J of Bolton’s zoning code, with the top 20 tokens of Topic 20 highlighted:

“The approval of an application for special permit or site plan review shall be void and of no effect unless the applicant has obtained a bona fide license from the Federal Communications Commission (FCC) to provide the telecommunication services that the proposed tower is designed to support and construction of the WTS is completed within one year from the date of the approval granted by the Commission. The Commission may grant up to two six-month extensions of this period upon written request by the applicant. The Commission shall not grant an extension unless the development plan is brought into conformance with any relevant zoning regulations which have been amended subsequent to the original approval and the applicant provides adequate evidence that construction is able to be completed within the extended time period sought. This evidence shall include, but not be limited to, the acquisition of any or all required government approvals and project financing. Any appeals

192 of such special permit, site plan, inland wetlands or subdivision approvals shall extend the aforementioned one-year period the length of such appeal. The Commission may as a condition of approval of a special permit establish a time period that such special permit shall remain in eﬀect.” - My emphasis.

The imposition of a time limit should have a negative effect since this means more work in the future to maintain a site relative to town with no such requirements. Finally, there are some topics, like Topics 16, 18, and 27 that seem do not seem to correlate to meaningful regulations so may simply represent the text that the model cannot explain with other topics. To measure the extent of regulation, I create different measures from the topic distributions. First, it may be important how strongly a town is focused on a particular kind of regulation. The more words the town spends on a topic the more important it might be to that town. I therefore directly use the town-level distributions of the topics from the sample. LDA produces sentence-level topic distributions, but for this application one would like to know the share of the section focused on a kind of regulation. To create what I call the Fo- cus measure, I again follow Hansen, McMahon, and Prat (2014), who create speaker-level topic distributions from topics estimated from statements uttered by the speakers in FOMC minutes. The procedure essentially consists of holding the sentence-level topics fixed and then resampling those topics to create the topic distribution at the town level. I explain this process in more detail in Appendix C.1.2. I report in Table 3.5 the average share of each topic over all towns. Most are around one-thirtieth, but a visual impact topic (Topic 7), a setback topic (Topic 10) and the length measure word topic (Topic 29) are all above 4 percent. In contrast to share, which denotes focus, it may be that the mere presence of a kind

193 of regulation is important rather than how many words are devoted to it. A very strong regulation that forbids sites in a single sentence would have a much stronger impact than a regulation that would disallow sites under mild but numerous requirements. Further, a terse regulation may have the same impact as a more verbose one since they have the same eﬀective legal impact. A Presence measure would therefore be closer to the manually-coded approach and thus easier to compare. Since any topic is possible for any word, every section will have an estimated non-zero share of every topic in practice. To approximate indicators I denote each sentence that represents whether the share of the topic in a sentence is in the 97.5th percentile of all shares (6.74%).41 I then create an indicator that is equal to one if any of the sentences exceed the 97.5th percentile for that topic for any sentence in the document. This is a very permissive measure as a single sentence at the 97.5th percentile out of dozens would register as “present.” It also has the drawback of being based on extreme draws of the sampling procedure - unlike the other variables which are based on the mean, it is more sensitive to random high draws on a particular sentence so the variable is less stable than the other two measures. As one can see in Table 3.5, most topics are “present.” The most present variable is Topic 3 with 120 towns, which seems to contain generic words about “design” characteristics, like “color,” “detail(s),” and “material(s).” The least present is Topic 23 with 58 towns, which seems to be about the “power” of the “signal” from the site. Finally, it may be the number or complexity of regulations that are important, i.e. the Volume. In that case, correcting for size of the section via the fraction of tokens assigned to that topic (i.e. the Focus measure) would be a mistake - the causal factor would be the

4195th percentile standard turns out to be too permissive and a 99th percentile standard is too strict - the former results in too many “positives” for the indicator and the later almost none. 97.5th percentile represents the top bound for a two-sided 95% test of signiﬁcance.

194 number of words the model estimates to be taken by a particular topic for a particular town. As an estimate, I multiply the town-level shares with the number of tokens per town, and then take the log since the result is very skewed. In practice these measures turn out to be extremely correlated, so I do not use them for the analysis.42

3.7 Estimation

To estimate the data I use the following linear speciﬁcation:

πm = NEAR(Xm) − F AR(Xm) + H(REGm) + m

= Xmγ + REGmκ + m (3.33)

Revenues are simply a linear function of town demographics, Xm, and enter the profit function via coefficients γ. Regulation costs are simply a linear function of regulation variables and enter via coefficients κ. Given the earlier model, I hypothesize κ will be negative for burdensome regulations and be positive for clarifying regulations.

Town characteristics Xm now include the earlier listed government, political and Census variables. If the border discontinuity methodology works well, they should be of minimal importance as the main diﬀerence in a small local area around a border should only be the diﬀerences in regulation. Population and area have been logged to reduce skewness.

REGm consists of the regulation measures, plus an indicator for the presence of a section, and the log length of the section in characters and the log length of the entire zoning code in characters.43 The section dummy controls for the fact there may be some discontinuity

42I report the correlation matrices between topics for all these measures in Appendix C.2.

43One has been added to character-based measures so that the log for zero characters is deﬁned.

195 Table 3.6: Collocation Correlation in 2013

AT&T Verizon Sprint T-Mobile MetroPCS

AT&T 1.00 Verizon 0.27 1.00 Sprint 0.21 0.25 1.00 T-Mobile 0.14 0.12 0.12 1.00 MetroPCS 0.20 0.17 0.20 0.20 1.00

between going from no regulation at all to some regulation. The log lengths of the section and whole code will control for the possibility that the total volume of the code is important rather than the exact contents as picked up by the regulation measures. The conditional probability of choosing town m with type 1 extreme value shocks is now:

exp(X γ + REG κ) P r(m) = m m (3.34) PM j=1 exp(Xjγ + REGjκ)

I estimate the regression using several different subsamples, as regulations, such as landscaping, could have different impacts for different types of facilities or different firms. These are the same subsample as Table 3.1. One sample contains only towers and the other only rooftops. The other subsamples select only sites with a base station from a specific firm. This is to simulate the base station placement decision by each firm, which might be more or less sensitive to regulation.44 These subsamples are clearly not independent as choosing to build a rooftop rather than a tower is a decision made by the siting firm. Firms also sometimes apply for sites jointly, with one firm filing the application to be the owner of the site but reporting another firm will

44A natural subsample would be all sites owned by a particular ﬁrm. However, these subsamples are too small for feasible estimation given the high number of owners in the market.

196 be on the site once it is built. Thus the subsampling should not be seen as perfectly clean identification of different effects. Still, the correlation reported in Table 3.6 for collocation does not seem especially strong: ranging from 0.12 (T-Mobile and Verizon) and 0.27 (AT&T and Verizon) Ideally, one would estimate these regressions via maximum likelihood alone. However, including the topic measures makes the number of regressors very high and the estimation becomes unwieldy and imprecise. To reduce the data, I use a variant of LASSO (Least Ab- solute Shrinkage and Selection Operator), introduced by Tibshirani (1996), which penalizes the likelihood by a function of the absolute values of the estimated parameters. The solution to this objective implies many of the coefficients to be zero, thus reducing the data to the variables with most explanatory power. In addition to reducing covariates, LASSO’s ability to select the regressors that explain more variation allows me to make direct comparison between the coded and the topic measures. That is, in addition to estimates with just the coded measures, just the Focus measures or just the Presence measures, I will estimate the parameters including all three sets of measures, and let LASSO chose amongst them. This will shed light on whether the topic measures may be improved measures with less bias relative to the coded measures. In particular, I use the procedure in Mauerer et al. (2014) which applies LASSO to a multinomial choice problem. The procedure consists of the following steps:

1. Normalize all continuous data by centering at 0 and dividing by twice the standard error, then adding 0.5. This makes the continuous variables have roughly the same domain as the binary variable, i.e. between 0 and 1. Thus all variables are roughly comparable and equally disfavored by the penalty.

2. Estimate parameters θ = {γ, κ} speciﬁc to a particular λ:

(a) If this is the ﬁrst iteration, set λ to some λMAX . For each iteration decrease λ by

197 step τ. I set τ so that the estimation takes 100 steps, overall.

(b) Maximize the log likelihood penalized by the term

K X ˆλ λ |ωkθk |. (3.35) k

Use of these weights, is called adaptive LASSO and was introduced by Zou (2006).

3. Select the estimate θˆλ with the lowest Akaike Information Criterion (AIC) from Akaike (1974).45

This process simulates a line search over the interval [0, λMAX ] for the optimal stringency of the penalty. Like Mauerer et al. (2014), I set λMAX to be

MAX T λ = max{|Zmk(Ym − P rm(θ = 0))|} (3.36) m,k

where Zmk are the covariates, Ym is the indicator vector of observed choices, and P rm(θ = 0) are the choice probabilities with parameters set to 0.46 Mauerer et al. (2014) point out that given the convexity of the logit maximum likelihood problem, the penalized problem will

1 have its solution in this interval. I also use, like Mauerer et al. (2014), ˆML as ωk. θk The only issue with this strategy is that when I use all the measures together in the Combination speciﬁcations, there is multicollinearity when all the variables are included together. Maximum likelihood using all the parameters is thus infeasible.47 To avoid this

45Mauerer et al. (2014) choose the AIC as balance between cross validation, which tends to include more variables, and the Bayesian Information Criterion from Schwarz (1978), which includes the less variables.

46Mauerer et al. (2014) actually use the results using only variables with unpenalized parameters, but since they are all penalized in my application, I have to set them all to 0.

47Even without the weights, LASSO is unable to hand collinear variables.

198 issue, I remove some of the variables ex ante in the Combination specifications.48 In all specifications, removing the regulation presence variable is necessary and usually sufficient. This is not surprising given I use many indicators for the different regulation types. For the Rooftops, I also have to delete the length of the section, the length of the code, the political variables, the percentage of black residents, and the Presence measures of the last twelve topics. This betrays that there is strong underlying covariance between particular regulations and the political and regulatory nature of towns where companies decided to build rooftop sites. I also have to remove the executive dummy for the Sprint specification and have to remove the dummies for the executive, the selectmen, the town meeting indicator, and Presence measures for topics 5, 6, 9, 12-15, and 17-29 for the MetroPCS specification. That I have to remove many variables for MetroPCS is not surprising as most MetroPCS sites are rooftops. LASSO has no analytical form for the standard errors so Mauerer et al. (2014) use bootstrapped standard errors. I take a different approach implement Post-LASSO (Belloni and Chernozhukov (2013)), which simply runs the un-penalized estimation with the restriction that it only uses variables with non-zero coefficients in LASSO. Post-LASSO has some attractive characteristics in the case of OLS, as it performs as least as well as LASSO in convergence and has a smaller attenuation bias. The relative econometric advantages of Post- LASSO versus LASSO are unknown for the multinomial logit case, but since the literature has not settled on optimal standard errors for LASSO, Post-LASSO provides a parsimonious way to produce estimates with well-understood standard errors. In this application, Post- LASSO also has the benefit that the coefficients are more interpretable than LASSO since

48Zou (2006) actually suggests implementing ridge regression to get weights, but given that this would also require search for a tuning parameter and be somewhat time intensive, I simply found the variables that were problematic and removed them. With collinear variables, one has a choice of which ones to remove, in my case, I tried to retain topic-based variables as much as I could as the Combination exercise is to help evaluate how regulation measures perform together.

199 the ex post estimation can be done on the un-normalized data.

3.8 Results

Results are reported in Tables 3.7, 3.8 and 3.9 for the researcher coded measures and the two topic-based measures.

3.8.1 Manually Coded Results

In the manually coded results, controls are often selected by LASSO and are often statistically significant. The importance of the controls implies that some of the variation across towns is not completely absorbed by the border discontinuity. What controls matter depends according to the subsample. Regulation and political controls seem to matter for the All, Towers, Rooftops Sprint, T-Mobile and MetroPCS subsamples. Interestingly, the AT&T and Verizon subsamples never select the regulation or political controls. The log zoning code character count and Executive government trait are never selected. Since each of these variables is only selected by a few subsamples and signs are often both different and statistically significant when they are, it is hard to make a clear conclusion about each.49 What might be most easily interpreted is Average Turnout, which is selected five times in the All, Towers, Sprint, T-Mobile and MetroPCS subsamples. As would be predicted based on the theory and findings in Hamilton (1993), turnout has a negative effect on siting, though it is not statistically significant in the complete sample. The fraction of Democratic voters is not selected as much (3 times), but is always negative and thus matches with the expectation that more Democratic towns may be more open to

49For the remainder of this chapter, when I write “statistically significant” I mean “statistically significant at the 10% level, or less.” For the most part, such results will be significant at the 5% level too, but I will use this as shorthand as I will be using this phrase, a lot.

200 Table 3.7: Results: Coded Regulation Measures

(1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mobile MetroPCS Section Present -1.80 0.50 (1.22) (0.40) log(1+Section Characters) -0.06 0.34 (0.06) (0.17)** log(1+Code Characters)

Executive

Council 0.44 -0.65 0.58 0.92 (0.24)* (0.35)* (0.27)** (0.43)** Selectmen -0.50 -0.84 (0.20)** (0.47)* Town Meeting 0.41 0.54 (0.20)** (0.27)** Average Turnout (Fraction) -0.93 -2.34 -3.18 -4.76 -4.85 (1.06) (1.37)* (1.89)* (1.67)*** (2.52)* Democratic (Fraction) -2.69 -5.64 -3.37 (1.30)** (2.12)*** (2.02)* log(Population) 0.12 (0.36) log(KM2) -0.20 -0.29 -0.49 -0.34 -0.90 -0.99 (0.15) (0.22) (0.25)** (0.26) (0.28)*** (0.51)* Median HH Income -8.05 -5.58 -16.47 -1.62 -34.78 -57.65 (4.46)* (5.43) (10.08) (8.17) (12.96)*** (21.52)*** Median Home Value 1.12 1.43 10.04 14.47 (1.17) (1.17) (3.31)*** (6.33)** Median Home Value > $1M 0.96 1.01 8.06 (0.65) (0.87) (2.61)*** Owned Housing Units (Fraction) -2.31 -3.54 0.35 -1.25 -1.07 (0.82)*** (1.60)** (1.24) (1.28) (1.53) More than High School (Fraction) -2.78 -3.22 -0.43 -2.09 (2.11) (2.07) (2.37) (2.65) Median Age

Black (Fraction) -3.31 -7.23 (2.06) (2.81)*** Hispanic (Fraction)

Consultant

Glossary 0.30 (0.26) Higher Authority 1.03 -0.60 0.79 1.53 (0.42)** (0.43) (0.34)** (0.59)*** Landscaping -0.27 -1.32 -0.65 -0.45 -0.17 -0.44 (0.16)* (0.36)*** (0.25)*** (0.36) (0.27) (0.44) Public Hearing -0.54 (0.35) Radiation Requirements 0.36 0.27 0.39 0.55 0.78 (0.22) (0.24) (0.25) (0.26)** (0.39)** Radiation Extra -0.37 -0.56 -1.08 0.78 (0.21)* (0.25)** (0.40)*** (0.39)** Removal Bond 0.29 (0.22) Size Restriction -0.42 -0.35 -1.13 -0.86 -0.92 -0.19 -0.76 (0.18)** (0.33) (0.48)** (0.29)*** (0.27)*** (0.27) (0.50) Visual Impact -0.71 1.04 -0.62 (0.39)* (0.43)** (0.58) Observations 553 267 286 261 226 215 226 123 Loglikelihood 394.72 187.68 189.42 186.57 155.11 150.26 153.18 74.87 Pseudo-R2 0.090 0.166 0.190 0.048 0.151 0.093 0.139 0.452 ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 Blank entries are variable not selected by LASSO. 201 regulation of firms. The other controls are often selected but are mostly only statistically significant for the T-Mobile or MetroPCS subsamples. The log area is selected six times and always negative when it is, and which might reflect that less dense towns are more costly and less profitable to site in. However, it is only significant in the MetroPCS subsample. Median household income is selected six times, and is always negative (though only statistically significant in Towers, T-Mobile and MetroPCS). Homeowner fraction is also always negative and statistically significant two out of the five times is it chosen. These findings match the idea that richer towns may be more active in government and have more to lose in property value from new sites. However, the median home value variables, while not selected as often and significant only for T-Mobile and MetroPCS, are actually always positive. This also goes against the related notion that higher real estate values are more costly for construction. It may be that home values are correlated with some dimension of profitability for cell towers that median income does not pick up, but it is hard to say. Finally, a more educated population is always statistically insignificant when chosen, and the percent of population who is black is only selected twice. It is hard to draw conclusions for these two variables. Median age and Hispanic population share are never selected. The selection for the coded measures is relatively sparse and chooses at most six of the regulations in a subsample. The Higher Authority indicator is selected four times, and the three times it is statistically significant it is also positive, as expected of potentially clarifying regulation. Higher Authority, however, is not chosen by the full sample, which might mean the result is not robust to sites, overall. Indeed, the one time Higher Authority is not significant is for Rooftops, so perhaps the clarifying effects are only felt in Towers, where it is positive, so that in the full sample the overall effect is not very strong. Landscaping regulation, which is expected to burdensome, is always negative, selected by six subsamples, and statistically significant in four of them. Likewise, Size Restrictions, which are known

202 to be burdens in the housing supply literature, are also always negative and statistically significant in four of the seven times they are chosen. Radiation Requirements is selected five times and always positive, which matches the idea that by spelling out radiation standards it has a clarifying effect. However, it is only significant twice. Similarly, the findings for burdensome Extra Radiation regulations are mixed - the estimate is negative and statistically significant in the All, Rooftops and Sprint subsamples, but actually positive in MetroPCS. The conclusion for Visual is even weaker, as it is negative in two of the three samples that select it (Towers and MetroPCS) but positive for Rooftops. Finally, Consultant, Glossary, Public Hearing Requirements, and Removal Bonds are chosen only for few subsamples, if at all, and are never significant. Thus while coded measure of Landscaping and Size Restrictions seem to be relatively robust and match the idea of burdensome regulations, the other measures seem more questionable and may be improved upon by the topic-based measures.

3.8.2 Topic Focus Results

Controls using the Topic Focus measures are again important; though different from the Coded measures in some ways. For example, the section presence variable, the log section character count and the log zoning character count are never selected. Again, the town government variables are hard to interpret given they are selected only a moderate amount of the time. Like the Coded estimates, the two political variables, voter turnout and Democratic fraction, are always negative when selected. The former is significant four of five times, and the latter is always significant, but only selected twice. Log population is always positive when selected, which matches the idea that bigger towns would be lead to more revenue for sites, but is only done so three times and is only statistically significant for the AT&T sample. Log area is actually chosen much more

203 Table 3.8: Results: Topic Focus: Share of Tokens

(1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

Section Present

log(1+Section Characters)

log(1+Code Characters)

Executive -0.85 (0.39)** Council 0.82 -1.23 0.53 0.10 (0.34)** (0.40)*** (0.27)** (0.36) Selectmen -0.30 -1.44 (0.24) (0.49)*** Town Meeting 0.43 0.70 (0.23)* (0.30)** Average Turnout (Fraction) -2.29 -3.30 -2.28 -4.58 -5.61 (1.16)** (1.63)** (1.55) (1.64)*** (2.62)** Democratic (Fraction) -3.55 -7.40 (1.41)** (2.46)*** log(Population) 0.46 0.24 0.30 (0.21)** (0.23) (0.40) log(KM2) -0.29 -0.42 -0.39 -0.98 -0.34 -1.10 -1.13 (0.17)* (0.25)* (0.29) (0.32)*** (0.26) (0.31)*** (0.59)* Median HH Income -7.15 -9.21 -4.42 -38.46 -23.55 (4.55) (5.22)* (4.41) (12.45)***(19.94) Median Home Value 8.66 (3.37)*** Median Home Value > $1M 7.28 -24.98 (2.57)*** (1.58)*** Owned Housing Units (Fraction) -3.48 -6.99 -2.09 (0.99)*** (1.99)*** (1.31) More than High School (Fraction) 3.23 (2.44) Median Age 0.05 (0.05) Black (Fraction)

Hispanic (Fraction) -6.37 (3.18)**

0 co, prefer, feasibl, altern, descript -13.32 -9.49 -13.17 -13.59 -10.14 -5.16 -14.12 (5.12)*** (6.60) (10.64) (6.01)** (6.64) (7.63) (12.16) 1 permit, follow, siteplan, specialpermit, regul -10.30 -3.24 (5.45)* (6.50) 2 properti, within, subject, adjac, public 9.13 17.96 (6.40) (14.32) 3 color, design, materi, detail, possibl 6.97 23.99 8.68 (3.71)* (7.58)*** (4.27)** 4 within, plan, show, map, coverag -12.86 -41.81 -8.37 19.70 (5.93)** (11.75)*** (7.12) (12.41) 5 provid, town, adequ, demonstr, capac 15.80 16.45 (6.01)*** (7.81)** 6 oper, owner, day, date, prior 20.28 22.04 19.44 12.36 17.46 22.10 (6.15)*** (10.24)** (7.02)*** (6.72)* (7.53)** (10.80)** 7 minim, design, advers, visual, potenti 10.50 (6.99) 8 elev, base, visibl, point, height 12.68 (7.92) 9 commun, commerci, district, radio, residenti 7.85 4.14 (3.96)** (3.23) 10 zone, residenti, setback, equal, proplin

11 regul, servic, person, provid, provis 13.74 22.50 12.10 18.34 9.47 9.79 23.59 (4.06)*** (6.18)*** (5.00)** (6.06)*** (5.04)* (6.72) (10.80)** 12 area, histor, develop, land, preserv 11.82 -8.94 -28.79 (8.97) (5.78) (9.51)*** 13 remov, abandon, upon, month, bond 11.95 -24.18 (5.36)** (9.68)** 14 build, equip, area, accessori, contain

15 engin, plan, licens, document, interfer 12.26 3.23 4.97 (7.91) (5.48) (5.26) 16 will, report, telecommun, indic, system

17 standard, rfr, compli, regul, fcc

18 addit, may, design, inform, accommod 11.77 (7.73) 19 limit, util, gener, instal, servic -17.10 -9.29 -22.65 (6.03)*** (6.91) (14.08) 20 approv, time, unless, specialpermit, may 7.07 12.30 10.97 Continued on next page

204 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

(3.32)** (5.48)** (5.56)** 21 need, telecomtow, order, number, reduc

22 mount, build, ground, roof, provid -19.69 (8.85)** 23 power, part, consid, signal, type -17.16 -16.98 -24.75 (6.35)*** (6.93)** (7.13)*** 24 fcc, light, feder, ﬁle, state -15.86 -3.92 -27.21 -11.64 (5.34)*** (7.58) (8.80)*** (6.01)* 25 construct, monopol, design, support, specif 7.56 17.28 (5.57) (12.46) 26 landscap, fenc, screen, surround, tree -12.24 -9.16 -13.37 -16.63 -19.49 -12.75 -5.02 -11.14 (4.26)*** (5.39)* (6.96)* (5.23)*** (6.38)*** (5.86)** (6.02) (9.36) 27 follow, receiv, commun, transmit, telecommun 9.00 8.61 10.19 10.04 13.47 (3.69)** (4.46)* (4.08)** (4.71)** (6.99)* 28 review, may, town, consult, independ -11.57 -6.29 -6.71 -11.57 -24.17 (5.55)** (6.60) (6.88) (8.19) (11.26)** 29 feet, height, exceed, inch, diamet -9.83 5.96 -14.29 (4.59)** (5.71) (7.18)**

Observations 553 267 286 261 226 215 226 123 Loglikelihood 380.45 179.97 177.53 180.39 144.97 150.55 148.82 71.66 Pseudo-R2 0.090 0.117 0.171 0.107 0.146 0.079 0.137 0.234

∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 Blank entries are variable not selected by LASSO. often, seven times, and like the Coded estimates is always negative. Also like the Coded estimates, median income is always negative, and home value variables are positive. However, median income is only selected five times, and only statistically significant for the AT&T and T-Mobile subsamples. Median home values are only ever selected for the T-Mobile and MetroPCS subsamples, and neither median income or home values are selected by the full sample. Fraction of owned housing units is always negative when selected, matching the expected sign, though only selected three times and statistically significant in two of them (All and Rooftops). The population share with more than a high school education, median age and black population share are at most selected once and never significant. Hispanic population is selected once by MetroPCS and is negatively statistically significant. The most consistent result, like the Coded estimates, is that landscaping requirements (Topic 26) are burdensome. Topic 26 is always selected, is always negative and is not significant in only T-Mobile and MetroPCS, which are relatively small subsamples. Topic 11, which is a good proxy for the recognition of higher authority, also has a strong, robust effect, as it is not selected for Rooftops and only statistically insignificant for T-Mobile.

205 Focus in Topic 11 might thus be a better measure of Higher Authority than the analogous coded measure. Topic 0, which appears to proxy for town preference and requirements for alternative site location in applications, is also selected seven times, and universally negative. This might suggest that such preferences indicate hostility to siting and such requirements are burdensome, though the topic is only significant in the full sample and for AT&T. Another potentially important topic is Topic 6, which collects time words that might have a clarifying effect. It is selected six times, including in the full sample and is always statistically significant. Topic 27 would also seem to be an important topic as it selected five times and statistically significant every time. However, recall this is one of the topics that has no clear interpretation, so I am unable to attach any meaning to Topic 27 always being positive when selected. What may be more meaningful is how Topic 28, which seems to proxy for consultant requirements, is selected five times, always negative and significant for the full sample and MetroPCS. This result is consistent with the intuition that consultant requirements should be burdensome. It is less clear if the other topics are important given how infrequently they are selected. The estimates for Topic 4 are reasonably straightforward. Topic 4 is always negative when selected, is selected four times, and is statistically significant in the case of the full sample and the AT&T subsample. The sign is as expected, as it refers to map requirements which might be burdensome. So is Topic 23 which seems to proxy for the Extra Radiation regulations: while only selected three times, it is always negative with statistical significance and is selected by the full sample. In contrast, Topic 3 is selected three times and is always positive and statistically significant. This topic seems to map on to design requirement of sites, which one might have assumed to be burdensome and thus have a negative effect. However, maybe a greater focus of these words represents a clarifying effect - either way, it is hard to say. This result is espe-

206 cially curious in light of Topic 24, which seems similar to Topic 3 in that it refers to design requirements, i.e. FCC light standards. Topic 24 is always negative when chosen, chosen by four subsamples, and is statistically significant in three of them including the full sample. Relative to Topic 6, one might think that the Topic 24 might be the one with positive effect, since its requirements are at the federal level and thus there might be a clarifying effect by bringing in the higher authority of the FCC. Topic 12 seems related to the concerns about the Visual Impact on historical areas and thus a candidate to proxy for that kind of type of regulation. However, Topic 12 is only selected by three subsamples, is significant in only one (MetroPCS), and is actually positive for Rooftops. Likewise, Topic 15 seems another proxy for the need for consultants but is never significant for the three times it is chosen. The remaining topics are only retained at most in two subsamples and thus are hard to interpret.

3.8.3 Topic Presence Results

The results for the Topic Presence measures have quite diﬀerent estimates for the controls. Likely because the Presence indicators have only limited variation relative to continuous measures, the controls are selected far more frequently than when using the Focus measures. However, the inclusion of controls is increased even relative to the Coded measures, which are also indicators. This might suggest that Coded measures might explain more variation than the Presence measures, but it might also be the case that the Presence measures explain more variation than the Coded measures, and in such a way that the controls can explain the remainder. For now, I put this question aside and leave it to the Combination regressions in the next subsection to decide which regressors really should be included. Presence of a section is selected four times and is signiﬁcant in three of them. Interest-

207 Table 3.9: Results: Topic Presence: 97.5th Percentile

(1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

Section Present 2.46 4.27 4.07 11.02 (1.52) (2.29)* (2.15)* (5.68)* log(1+Section Characters) -0.37 -0.70 -0.06 -0.04 -0.47 0.02 -2.62 (0.22)* (0.32)** (0.03)** (0.09) (0.29) (0.06) (0.99)*** log(1+Code Characters) 0.67 (0.30)** Executive -0.56 -4.34 (0.42) (1.51)*** Council 0.86 -2.26 0.64 7.43 (0.37)** (0.60)*** (0.27)** (1.86)*** Selectmen -0.43 -3.12 -0.75 2.35 (0.24)* (0.82)*** (0.40)* (1.77) Town Meeting 0.45 0.47 -1.09 0.71 4.34 (0.25)* (0.33) (0.43)** (0.36)** (1.32)*** Average Turnout (Fraction) -2.68 -5.69 4.25 -6.28 -26.46 (1.22)** (1.76)*** (2.20)* (1.80)*** (6.73)*** Democratic (Fraction) -3.92 -1.50 -4.09 15.74 (1.46)*** (2.01) (1.99)** (8.52)* log(Population) 0.24 0.06 0.33 4.35 (0.18) (0.14) (0.29) (1.43)*** log(KM2) -0.39 -0.45 -0.84 -0.74 -1.07 -4.11 (0.20)* (0.27)* (0.31)*** (0.36)** (0.30)*** (1.22)*** Median HH Income -8.76 11.56 -1.61 (4.27)** (13.04) (7.34) Median Home Value -2.95 6.95 1.94 2.85 (2.14) (2.75)** (1.11)* (2.30) Median Home Value > $1M -2.38 5.34 0.27 2.16 -32.24 (1.42)* (2.02)*** (1.06) (1.50) (2.66)*** Owned Housing Units (Fraction) -2.26 -2.17 -3.11 -3.28 -9.67 (1.03)** (1.41) (2.05) (1.37)** (4.86)** More than High School (Fraction) -8.60 -3.53 -2.82 (3.49)** (3.37) (3.06) Median Age 0.09 0.10 -0.04 0.19 (0.05)* (0.05)* (0.04) (0.15) Black (Fraction) -18.60 (9.02)** Hispanic (Fraction) -4.22 -30.85 (2.56)* (9.14)***

0 co, prefer, feasibl, altern, descript -1.95 1.09 (0.58)*** (1.53) 1 permit, follow, siteplan, specialpermit, regul 0.83 (0.45)* 2 properti, within, subject, adjac, public 0.51 0.64 -3.37 (0.26)** (0.37)* (1.61)** 3 color, design, materi, detail, possibl -0.42 0.47 (0.39) (1.18) 4 within, plan, show, map, coverag 0.56 1.66 0.33 3.74 (0.27)** (0.48)*** (0.34) (1.75)** 5 provid, town, adequ, demonstr, capac 0.45 0.50 1.33 0.71 1.11 0.44 2.77 (0.20)** (0.30) (0.37)*** (0.32)** (0.32)*** (0.29) (1.16)** 6 oper, owner, day, date, prior 0.27 0.37 0.84 1.61 (0.22) (0.32) (0.37)** (0.97)* 7 minim, design, advers, visual, potenti -0.28 -0.85 -0.84 -0.56 5.17 (0.28) (0.50)* (0.49)* (0.44) (1.81)*** 8 elev, base, visibl, point, height -0.64 -1.11 0.60 -4.92 (0.30)** (0.44)** (0.43) (1.58)*** 9 commun, commerci, district, radio, residenti -0.53 -0.63 3.74 (0.29)* (0.29)** (1.37)*** 10 zone, residenti, setback, equal, proplin 0.94 (0.54)* 11 regul, servic, person, provid, provis 0.42 0.94 0.61 0.48 0.81 4.72 (0.19)** (0.29)*** (0.23)*** (0.29)* (0.31)*** (1.45)*** 12 area, histor, develop, land, preserv 0.48 0.25 2.30 0.51 -0.59 (0.25)* (0.33) (0.55)*** (0.37) (1.17) 13 remov, abandon, upon, month, bond -0.55 -0.76 -1.00 -0.20 -3.59 (0.32)* (0.49) (0.64) (0.42) (1.61)** 14 build, equip, area, accessori, contain -0.17 0.73 3.67 (0.24) (0.39)* (1.41)*** 15 engin, plan, licens, document, interfer -0.88 -4.96 (0.45)* (1.82)*** 16 will, report, telecommun, indic, system -0.60 0.46 -2.84 (0.35)* (0.43) (1.06)*** 17 standard, rfr, compli, regul, fcc 0.44 0.64 0.27 -0.56 0.58 (0.20)** (0.33)* (0.34) (0.35) (0.39) 18 addit, may, design, inform, accommod 0.51 1.17 0.70 0.08 0.45 3.87 (0.29)* (0.45)*** (0.60) (0.34) (0.42) (1.27)*** 19 limit, util, gener, instal, servic -1.02 -0.76 -1.59 (0.59)* (0.37)** (1.13) 20 approv, time, unless, specialpermit, may -0.57 -0.26 -0.79 4.94 Continued on next page

208 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

(0.22)*** (0.30) (0.34)** (1.57)*** 21 need, telecomtow, order, number, reduc -0.37 -3.51 (0.35) (1.24)*** 22 mount, build, ground, roof, provid 0.00 3.09 (0.39) (1.49)** 23 power, part, consid, signal, type 0.24 0.59 (0.19) (0.37) 24 fcc, light, feder, ﬁle, state 1.40 5.29 (0.61)** (1.98)*** 25 construct, monopol, design, support, specif 0.54 -0.39 -0.47 5.02 (0.35) (0.35) (0.32) (1.68)*** 26 landscap, fenc, screen, surround, tree 0.60 -2.01 -0.38 -0.94 (0.41) (0.57)*** (0.37) (0.41)** 27 follow, receiv, commun, transmit, telecommun -0.21 -1.76 2.05 (0.21) (0.42)*** (0.83)** 28 review, may, town, consult, independ -0.56 1.58 -0.41 -6.05 (0.40) (0.52)*** (0.33) (1.61)*** 29 feet, height, exceed, inch, diamet 0.47 -3.76 (0.41) (1.56)** Observations 553 267 286 261 226 215 226 123 Loglikelihood 380.36 170.03 173.57 192.31 144.10 148.27 148.60 51.28 Pseudo-R2 0.090 0.166 0.190 0.048 0.151 0.093 0.139 0.452

∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 Blank entries are variable not selected by LASSO. ingly, it is always positive, which would suggest adding regulations would incentivize more investment. However, log character count of the section also is selected in these same samples and has a negative effect. Moreover, when there is continuous variation provided in the Focus estimates, none of these variables are selected. Overall this suggests that the Focus measures explain the negative impact of longer zoning codes, and when replaced by the Presence measures the remaining variation covaries with the length of the entire zoning code. This remaining variation has a somewhat nonlinear relationship with section length, thus results in a higher “intercept” of the reduced form profit in towns with codes relative to town without them. The town government traits are selected more, even compared to the Coded and Focus regressions. However, they are now even harder to interpret given 1) that there are statistically significant results of both signs for different subsamples (Council, Selectmen, Town Meeting) or 2) that there are very few subsamples in which the trait is selected (Executive). Average turnout and Democratic voting fraction is mostly consistent with previous results, with mostly being negative statistically significant when selected. I say mostly since Aver- age turnout has a statistically significant positive coefficient in the Rooftops subsample, and

209 Democratic voting fraction has a statistically significant positive coefficient in the MetroPCS subsample. Consistent with previous results when selected, log population is positive though it is only selected four times and only significant for MetroPCS. Log area is again negative, and is actually statistically significant in every subsample it is chosen in - it is only removed in the AT&T and Sprint subsamples. Unlike previous results, median income does not have uniformly negative impact, and median home value does not have a universally positive effect. Median income has a positive impact on siting for Verizon, and the home value variables have a negative impact for Towers. None of these effects are statistically significant, and in general these variables are selected less than they were in the previous results. Home owner fraction is now uniformly negative, in line with expectations, and selected in five subsamples, though only significant for the full sample and the T-Mobile and MetroPCS subsamples. Matching the Coded results, a more educated populace has a negative impact in the three subsamples that select it, though it is only statistically significant in the AT&T sample. Median age is selected four times and when statistically significant, it has a positive effect, opposite to the expectation that a young community may be more receptive to high technology and thus increases revenues. Finally, black and Hispanic population fractions have statistically significant negative coefficients, though these are hard interpret since they only appear in a few subsamples as regressors. The impact of topics is also different as the Presence measure gives importance to different topics than Focus. Topic 11, the Higher Authority proxy, is still important as it is selected six times, and always statistically significant and positive. The other very robust topic in the Focus results, Topic 26, is in contrast only selected four times, not selected by the full sample, and statistically significant only twice. Topic 0 and Topic 6 are also important in the Focus results but the former is only selected twice and the latter only four times. Topic 27,

210 the uninterpretable yet frequently selected topic in the Focus estimates, is only selected three times here and actually has both negative and positive statistically significant estimates for different subsamples. Likewise, the Consultant proxy Topic 28 is selected only four times, is statistically significant in only two, and these estimates are of different signs. Instead, the next most important topic seems to be Topic 5, which is selected seven times, always positive, and statistically significant five times, and Topic 18, which is selected six times, always positive, and statistically significant three times. Topic 5 appears to proxy for requirements for evidence that the location will generate sufficient coverage, so one might expect presence of Topic 5 to indicate a burden for the applicant and thus have a negative coefficient. It seems likewise for Topic 18, which appears to map onto requirements site design accommodates collocation. However, it might be possible that these Topic indicators signify that the town actually cares about coverage as the requirement are about improving coverage (Topic 5) and more efficiently using sites (Topic 18). Thus these towns may be internalizing the benefits to its cell phone-using populace more than other towns. Still, it is not ex ante obvious, even if this were true, that the effect of the town’s underlying pro-phone preference would overcome the extra burden imposed by the regulation. This reasoning might also hold for Topic 4, which is about map requirements for application and is selected by four subsamples, always positive, and statistically significant in three. Perhaps there is a positive effect because this shows that the town cares about coverage. This would be especially interesting because the Focus version of this topic has negative effects. This would not be totally contradictory: maybe presence of the requirement means the town values coverage, but if the town focuses a lot on the map in the text the required map may become especially complicated and burdensome. Topic 7, which would seem a good proxy for the Visual Impact type of regulation, seems more important here than in the Focus estimates in which it was only selected once. As a Presence measure it is selected five times and is statistically significant three times. However,

211 it is not statistically significant for the full sample and actually has a large, statistically significant positive effect for the MetroPCS sample, which is opposite what one might expect for the Visual Concerns type of regulation. In contrast, Topic 8 might do better as a proxy for Visual Concerns. It is selected only four times but is statistically significant in three of them with the expected negative sign. Finally, Topic 12, which focuses on visual concerns for historic areas, is selected more than it was as a Focus measure, but now has the wrong sign the two times it is statistically significant. Topic 13, which seems to proxy well for the Removal Bond regulation, is negative all five times it is selected, which is what one would expect. However, it is only statistically significant twice, and one of those is at the 10% level (the Full sample) so the estimated may not be that precise. While the Focus version of the Topic 17, a potential proxy for clarifying radiation requirements, was never even chosen, the Presence version is chosen five times. However, it is only statistically significant twice, so it is hard to say Topic 17 is a precise proxy. This is even if these two statistically significant effects are positive, which one expects from a clarifying regulation. Topic 20 appears to map onto regulations imposing a time limit for how long an approval to build a site will last. This is selected four times and is statistically significant thrice, including the full sample. As one might expect from such a restriction, the effect is negative in the full sample and for Rooftops, but oddly enough has a positive effect on MetroPCS. The other topics under the Presence measure are selected with less frequency and are less often statistically significant, so are harder to interpret.

212 Table 3.10: Results: All Regulation Measures

(1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

Section Present ------log(1+Section Characters) - 0.30 0.19 7.03 - (0.30) (0.23) (3.05)** log(1+Code Characters) 0.33 - 3.46 (0.16)** - (1.36)** Executive -1.00 -19.51 -1.66 -0.48 -2.80 - - (0.42)** (4.09)*** (1.39) (0.44) (0.91)*** - - Council 4.76 - 1.27 -1.87 -0.19 (1.32)*** - (0.73)* (0.71)*** (0.48) Selectmen -1.30 -14.03 - -1.91 -0.61 - (0.44)*** (3.50)*** - (1.15)* (0.62) - Town Meeting -3.33 - -1.37 - (1.36)** - (0.74)* - Average Turnout (Fraction) - -4.42 -6.23 21.76 - (2.76) (2.69)** (18.80) Democratic (Fraction) -20.53 - (6.67)*** - log(Population) -2.39 0.87 7.77 (0.94)** (0.32)*** (2.52)*** log(KM2) -0.22 1.56 -6.98 -1.95 -1.44 -0.42 -1.62 -5.04 (0.17) (1.03) (1.81)*** (0.57)*** (0.46)*** (0.55) (0.44)*** (1.67)*** Median HH Income -598.46 58.74 -35.33 -495.42 (158.35)*** (28.94)** (15.99)** (272.01)* Median Home Value -25.06 116.60 -6.47 -20.88 10.39 30.20 (5.37)*** (28.95)*** (3.57)* (7.72)*** (4.76)** (35.35) Median Home Value > $1M -17.93 93.71 -3.57 -19.55 9.22 - (4.38)*** (23.11)*** (2.09)* (6.23)*** (3.71)** - Owned Housing Units -1.28 -23.80 58.38 -4.84 43.74 (Fraction) (0.48)*** (6.53)*** (20.36)*** (1.46)*** (25.30)* More than High School -32.20 -5.50 -7.62 (Fraction) (14.55)** (4.34) (5.39) Median Age -0.34 -2.11 -0.53 (0.15)** (0.56)*** (0.36) Black (Fraction) - -6.27 -28.68 - (2.80)** (13.89)** Hispanic (Fraction) -12.34 -54.13 -7.33 -34.64 (5.27)** (14.49)*** (3.15)** (16.13)**

Consultant -6.62 -54.41 -1.37 -0.84 -3.03 19.03 (1.50)*** (11.44)*** (0.54)** (0.53) (1.50)** (5.97)*** Glossary 1.77 -19.32 7.18 0.13 2.80 (1.22) (5.12)*** (2.24)*** (0.67) (2.01) Higher Authority -9.64 24.65 0.36 -0.92 0.96 1.09 22.23 (3.17)*** (5.39)*** (0.90) (1.00) (1.61) (0.68) (6.37)*** Landscaping -5.44 -19.89 -1.09 4.22 -9.74 (1.27)*** (6.25)*** (0.59)* (1.63)*** (3.39)*** Public Hearing -0.02 -9.44 0.81 -6.26 4.35 (1.11) (3.61)*** (0.62) (1.99)*** (4.90) Radiation Requirements -9.73 -1.43 (2.07)*** (0.90) Radiation Extra 5.73 -0.35 6.96 (1.90)*** (0.39) (2.31)*** Removal Bond 6.03 22.49 1.44 -0.12 2.36 -7.12 (1.58)*** (5.13)*** (0.46)*** (0.44) (1.17)** (2.67)*** Size Restriction 19.79 -1.78 8.98 -0.94 (5.25)*** (0.70)** (2.74)*** (0.76) Visual Impact -2.36 9.97 -1.04 -9.84 (0.91)*** (2.83)*** (0.61)* (4.30)**

0 co, prefer, feasibl, Focus -68.47 -1421.62 -13.67 -36.26 -34.68 altern, descript (34.18)** (300.22)*** (16.64) (14.89)** (31.40) Pres. 35.75 -2.38 1.20 -1.61 -5.75 (7.03)*** (0.83)*** (1.59) (0.79)** (1.58)*** 1 permit, follow, siteplan, Focus -229.73 87.66 -55.68 -29.65 -71.94 -22.04 -92.54 specialpermit, regul (51.20)*** (61.36) (14.06)*** (15.46)* (33.32)** (13.36)* (80.57) Pres. 16.95 -13.65 2.24 -0.08 6.12 0.28 (4.14)*** (3.69)*** (0.87)*** (1.00) (2.44)** (0.66) 2 properti, within, subject, Focus 200.13 258.01 80.38 88.41 14.83 -123.79 adjac, public (45.11)*** (83.92)*** (16.27)*** (39.96)** (15.93) (44.48)*** Pres. 3.15 -7.56 0.07 -8.87 (1.27)** (3.36)** (0.51) (2.71)*** 3 color, design, materi, Focus 8.47 -160.09 35.24 -15.94 -20.34 209.88 detail, possibl (4.30)** (43.24)*** (10.78)*** (13.91) (11.31)* (146.22) Pres. 2.97 23.96 10.00 0.70 -10.39 (2.80) (5.84)*** (3.17)*** (0.73) (9.13) 4 within, plan, show, Focus -8.72 580.24 -41.21 -93.33 -202.51 map, coverag (5.82) (135.45)*** (19.18)** (34.62)*** (58.93)*** Pres. 19.85 -32.45 3.32 -1.26 2.05 -11.36 (4.18)*** (7.40)*** (1.05)*** (1.64) (0.90)** (8.04) 5 provid, town, adequ, Focus -230.22 982.77 -96.09 86.93 -144.43 Continued on next page

213 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

demonstr, capac (61.52)*** (268.12)*** (22.61)*** (64.95) (40.19)*** Pres. 0.50 15.86 -13.93 2.76 1.91 3.17 - (0.22)** (2.99)*** (4.42)*** (0.65)*** (0.64)*** (1.14)*** - 6 oper, owner, day, Focus 140.23 405.38 64.93 6.69 -278.21 date, prior (32.18)*** (98.54)*** (30.91)** (10.75) (99.86)*** Pres. 0.52 3.09 1.58 1.82 11.09 - (0.25)** (1.17)*** (3.47) (0.62)*** (2.81)*** - 7 minim, design, advers, Focus 29.25 -6.86 90.86 -60.93 visual, potenti (66.25) (12.84) (35.79)** (58.79) Pres. 14.91 -37.99 -0.58 -4.43 (3.67)*** (8.57)*** (1.04) (2.50)* 8 elev, base, visibl, Focus -442.93 387.08 -29.74 -16.22 -117.66 point, height (92.66)*** (98.33)*** (17.09)* (37.91) (40.89)*** Pres. 2.81 0.38 2.53 -4.29 (1.86) (0.58) (0.79)*** (1.42)*** 9 commun, commerci, district, Focus 184.07 -236.94 17.69 15.58 -3.66 -193.99 radio, residenti (37.61)*** (73.29)*** (8.06)** (10.37) (16.24) (60.08)*** Pres. -4.60 76.98 -1.60 - (1.28)*** (14.92)*** (0.62)*** - 10 zone, residenti, setback, Focus -339.48 -40.56 8.43 -160.47 equal, proplin (82.42)*** (18.50)** (7.84) (72.19)** Pres. -2.58 48.38 -1.66 -1.87 3.09 (1.35)* (10.31)*** (0.61)*** (1.23) (5.02) 11 regul, servic, person, Focus 10.78 220.18 555.73 11.06 19.96 113.02 25.33 -306.42 provid, provis (3.98)*** (37.95)*** (123.17)*** (10.76) (12.74) (36.53)*** (11.70)** (119.71)** Pres. 2.31 -34.56 1.96 0.85 2.77 3.23 (1.27)* (7.81)*** (0.51)*** (0.70) (1.38)** (2.95) 12 area, histor, develop, Focus -138.92 523.56 -112.34 -41.70 -209.86 land, preserv (43.21)*** (129.65)*** (37.44)*** (12.70)*** (62.07)*** Pres. 0.59 4.82 -48.31 0.85 8.17 2.47 - (0.22)*** (1.91)** (10.67)*** (0.52)* (2.33)*** (0.75)*** - 13 remov, abandon, upon, Focus 135.37 -968.13 23.38 -1.96 month, bond (35.91)*** (190.93)*** (14.47) (36.05) Pres. -0.64 -16.23 -1.04 -8.45 -2.03 - (0.27)** (3.67)*** (1.02) (3.10)*** (0.70)*** - 14 build, equip, area, Focus 89.62 351.60 -2.85 -38.98 -192.29 accessori, contain (26.12)*** (83.39)*** (12.67) (22.67)* (80.10)** Pres. -12.93 -1.19 -1.28 - (3.02)*** (0.79) (0.74)* - 15 engin, plan, licens, Focus -122.57 44.39 -10.10 28.96 24.01 document, interfer (31.00)*** (61.91) (12.57) (21.67) (27.92) Pres. -10.68 -39.00 -1.07 -0.87 -3.63 -0.85 - (2.27)*** (7.78)*** (0.58)* (0.75) (1.44)** (0.64) - 16 will, report, telecommun, Focus 103.44 -705.28 34.97 101.92 169.17 indic, system (43.21)** (146.71)*** (16.88)** (36.08)*** (58.48)*** Pres. -10.37 27.86 -1.75 -0.11 -5.55 (2.41)*** (5.01)*** (0.66)*** (0.71) (2.11)*** 17 standard, rfr, compli, Focus 226.72 -1288.95 -38.25 regul, fcc (59.08)*** (253.80)*** (78.74) Pres. 0.35 9.83 24.71 0.68 1.03 -2.43 1.55 - (0.21)* (2.01)*** (5.22)*** (0.59) (0.66) (1.24)** (0.53)*** - 18 addit, may, design, Focus 127.85 269.71 -1.34 -35.35 -63.15 inform, accommod (35.98)*** (107.42)** (9.45) (39.00) (23.28)*** Pres. - 0.97 -1.33 - - (1.75) (0.66)** - 19 limit, util, gener, Focus -16.46 -219.64 838.71 -9.85 -35.66 -86.27 -22.16 -409.79 instal, servic (6.53)** (61.87)*** (171.79)*** (19.57) (14.26)** (42.24)** (14.24) (189.41)** Pres. 3.98 - 1.29 - (1.21)*** - (0.76)* - 20 approv, time, unless, Focus 5.20 13.88 14.47 15.29 -33.13 190.06 specialpermit, may (3.43) (21.17) (9.82) (11.56) (25.72) (70.72)*** Pres. -0.56 -0.68 - -2.01 -4.21 -0.34 - (0.20)*** (1.25) - (0.55)*** (1.48)*** (0.44) - 21 need, telecomtow, order, Focus -150.44 197.61 31.53 31.70 129.52 64.52 number, reduc (41.55)*** (64.20)*** (12.57)** (18.29)* (50.88)** (48.00) Pres. -10.03 - -1.09 -5.28 - (2.18)*** - (0.63)* (2.07)** - 22 mount, build, ground, Focus 60.51 -233.08 -54.08 -17.96 37.61 roof, provid (30.83)** (62.69)*** (34.31) (11.40) (36.77) Pres. 4.71 - -0.65 0.37 - (1.68)*** - (0.61) (0.58) - 23 power, part, consid, Focus -13.89 -251.50 -914.82 83.57 -78.40 -229.96 43.19 signal, type (5.57)** (62.36)*** (224.24)*** (22.89)*** (20.85)*** (85.47)*** (43.85) Pres. -2.38 - -1.72 1.56 4.15 - (1.03)** - (0.55)*** (0.60)*** (1.42)*** - 24 fcc, light, feder, Focus -176.52 -81.41 107.19 ﬁle, state (45.36)*** (52.34) (66.63) Pres. -0.33 9.86 - 0.77 -0.02 -0.81 1.61 - (0.25) (2.45)*** - (0.66) (0.66) (1.21) (0.82)** - 25 construct, monopol, design, Focus 40.03 -1467.85 24.15 128.97 243.81 support, specif (33.45) (292.70)*** (17.47) (53.07)** (83.40)*** Pres. -1.46 - -0.74 -1.44 -14.45 -0.35 - Continued on next page

214 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

(1.50) - (0.55) (0.85)* (3.67)*** (0.60) - 26 landscap, fenc, screen, Focus -11.65 -119.60 -407.44 -51.05 -29.21 -266.59 -27.02 -85.42 surround, tree (4.10)*** (29.25)*** (87.24)*** (14.66)*** (14.26)** (69.77)*** (10.90)** (39.59)** Pres. 3.68 - 0.89 1.13 - (1.50)** - (0.82) (0.61)* - 27 follow, receiv, commun, Focus 8.22 323.30 315.85 41.45 39.95 38.67 23.80 transmit, telecommun (3.88)** (60.59)*** (75.70)*** (10.34)*** (11.52)*** (24.58) (10.24)** Pres. -9.84 - -1.86 -1.48 -2.04 -2.21 - (2.34)*** - (0.72)*** (0.76)* (1.38) (0.79)*** - 28 review, may, town, Focus 721.20 consult, independ (178.08)*** Pres. -2.04 - -3.12 -0.52 - (1.26) - (1.78)* (0.59) - 29 feet, height, exceed, Focus 3.88 -112.78 -14.81 93.62 -141.17 inch, diamet (3.28) (54.43)** (8.61)* (32.12)*** (79.11)* Pres. -6.99 - 1.73 1.25 - (1.61)*** - (0.89)* (0.84) - Observations 553 267 286 261 226 215 226 123 Loglikelihood 379.20 114.32 127.74 142.43 117.44 109.74 132.74 41.36 Pseudo-R2 0.093 0.439 0.404 0.295 0.308 0.328 0.231 0.558

∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 Blank entries are variable not selected by LASSO. Entries with “−” were removed ex ante due to multicollinearity.

3.8.4 Combination of Regulation Measures Results

As promised, I take advantage of LASSO’s ability to select over many regressors to run an estimation using all the regulation measures. This allows me to see if some measures can be replaced or complemented with other measures. This analysis seems especially pressing since all the measures seem to be somewhat imprecise. There are many cases where the signs of the coefficient are both opposite to what is expected and statistically significant. Moreover, results are quite different, even between the two topic-based measures, posing the question of what is exactly driving this difference. I present the results of this estimation in Table 3.10. What is immediately most strik- ing is that LASSO does not remove that many regressors from the proper subsamples, i.e. subsamples besides the full sample. For example, while the full sample only retains twenty regressors, Rooftops removes only eight.50 Thus I may have something of an overfitting problem with the subsamples - given they are so small, the patterns of variation in choices of towns is more easily explained by additional covariates that are not actually related. The problem is mitigated by a larger sample size, since that introduces variation that is hard

50As noted earlier, Rooftops has to remove many regressors due to collinearity issues.

215 to explain by random chance. LASSO does mitigate this issue, but with the small sample size and a large number of covariates this seems to be harder to accomplish. In fact, this suggests I may have been running into this issue in the other regressions with the smallest subsample of MetroPCS: it often has statistically significant results that other samples did not and with unexpected signs. Thus for this part of the analysis I will focus entirely on the Full sample as this seems to be the most robust and most easily interpreted part of the analysis. Three statistically significant controls are retained in the full sample: the Executive government trait, the Se- lectmen government trait, and the fraction of home ownership. That both Executive headed governments and governments with Selectmen both have statistically significant negative impacts implies that it may be incorrect to assume a more modern form of town organization is preferable to an applicant. Old-fashioned government run by Selectmen are still clearly problematic, but it seems having a strong Executive may be harder to deal with than a group of people who may be less unified in opposition. The statistically significant impact of home ownership fraction is also expected, as it suggests that the town government may put more weight on potential impacts on housing values since the voting base is more likely to be hurt. Somewhat remarkably, none of the coded variables are retained. While they are com- monly retained in the other subsamples, the fact that LASSO chooses to remove them all from the largest sample is evidence that the topics do a good job of proxying for the different kinds of regulation. Another interesting point is that none of the topic measures retained have both a statistically significant Focus version and a statistically significant Presence version. Only Topic 20 retains both versions, and there are both Focus measures and Presence measures retained. This coexistence of the two difference kinds of topic measures for different topics indicates that instead of one measure being a superior proxy relative to the other, some regulations

216 act primarily through how much focus the towns code gives them, while other acts though mere presence. Thus some of the differences between the estimates of the two measures are simply random variation. Of the retained topics with statistically significant coefficients, the robust Focus measures of Topic 11 and Topic 26 are of note as these have been the most robust findings across all the specifications. The negative coefficients on Focus in topics 19 and 23 are both consistent with burdensome regulations, as they appear to proxy for regulations for electric power and signal strength, respectively. Likewise, negative coefficients on the presence of topics 13 and 20 are also indicative of burdens, as the former proxies for removal bond requirements, and the later proxies for time limits on building approvals. In contrast, the positive coefficients of the presence of topics 6 and 17 are consistent with the hypothesis that increased clarity increases investment. The former indicates the use of precise time words, and the latter establishes town recognition of FCC standards of regulation. Not surprisingly, the estimates are not 100% in line with the hypotheses - Focus in Topic 3 about design characteristics should have a negative impact as a burden but has a positive impact instead. So should Presence of Topic 12, about visual concerns of historical areas, but it too has a positive impact. The positive impact of the presence of Topic 5, which requires that the new site will provide adequate capacity to the town, might be in line with hypothesis that towns that care about capacity must put more weight on the public benefits of the site, but is also clearly an added cost. Without more information it is hard to say how the effects would balance out. Finally, Focus in Topic 27, the uninterpretable topic, is still retained with a statistically significant positive effect.

217 3.8.5 Fifty Topics Results

Fifty is the number of topics used in Hansen, McMahon, and Prat (2014), though they have several times as many data points as I do. As a result, I have scaled down the number of topics in my analysis to 30. As a robustness check, I estimate the results using topics estimated with K = 50 instead of 30 and report the results in Appendix C.3. Similar topics are generated. For example, Topic 26, representing landscaping requirements in the 30 topic estimates, has four of the same top five words as Topic 33 in the new 50 topic estimates. The increased number of topics results in many of the topics being split up so there are now several topics for the same basic idea. For example, Topic 26’s fifth most common word, “tree,” now appears in Topic 3 as its most common word. Comparing the 30 topic and the 50 topic regressions shows there are differences between how the Focus and the Presence measures change when the larger number of topics are added. Results for the Focus variables are similar in that similar topics are selected and found to impact the facilities location decision. More noticeably, Focus in landscaping and the language establishing higher authority (represented now by Topic 27) are both selected and statistically significant in most subsamples, as they were in the main specifications. However, the value of the new topics is questionable for the Focus regressions. For the full sample regression of the Focus measures, the same number of topics are selected as the 30 topics specification (15), but more controls are selected. This implies that the new topic measures are now somewhat less effective at explaining variation when penalized and more weight must be given to the controls. In addition, the selected Presence measures seem quite different and seem to contradict the expected signs much more often. For example, Topic 11 in the 30 topics specifications and Topic 10 in the 50 topics specifications seem to be very similar and proxy for Visual Impacts regulation. However, while Topic 11 has often a negative coefficient and is not statistically

218 significant when it is selected, in the new regressions when Topic 10 is selected it has a statistically significant positive coefficient. More often though, the Presence measures with 50 topics seem to choose completely different topics than the original presence measures with 30. This might imply that the presence proxy becomes less informative as the number of topics goes up, which might be expected given how the presence proxy is constructed with a percentile cutoff.51 Also, the inclusion of so many new regressors means that multicollinearity is even a problem when not combining all the measures (Coded, Focus, and Presence) together. I had to remove many variables ex ante from the MetroPCS regressions and the Section Present dummy from all the Focus regressions. Moreover, the inclusion of more regressors means that the overfitting issue seen in the Combination regressions now even appears in the Focus regressions, as proper subsamples tend to eliminate far fewer regressors in LASSO than before. Given these drawbacks, I continue to use the 30 topics results as the main specifications.

3.9 Discussion

In general, the estimates do not have as much precision as one would have hoped. In particular, while estimates differ between subsamples the precision of the estimates is too low to say with any confidence that they are consistently different. Moreover, there appears to some underlying difference between towers and rooftop sites that results in them often have statistically significant effects of difference signs for the same variable. In particular, there may be strong selection of rooftop sites to particular towns with many buildings, which helps to cause the multicollinearity issues in the Combination regressions. Thus there may

51See Appendix C.3 for a full discussion of this point.

219 be some underlying process of Tower vs. Rooftop determination that future work should explore. Overall though, the topic-based measures perform well: they manage to proxy for what the Coded measures attempt to measure, sometimes very directly. Moreover, they often pick up on things that this researcher did not, such as the language associated with Topic 11 establishing federal authority, the time words of Topic 6, and the time limits of Topic 20. Given the results, it does seem that the statistically significant effects have (mostly) the expected signs across all topic specifications, and are roughly consistent with coded measures. Depending on the regression, landscaping requirements, extra radiation requirements, electric power requirements, and expiration dates for approvals are burdensome and decrease the attractiveness to locate in a certain town. Across all regressions, the negative effect of landscaping is the most consistently estimated, which might reflect how landscaping seems to be very clearly outside the core competency of carriers and tower companies and thus unusually hard for them to satisfy. In contrast, increasing clarity has a positive effect on location probability, via explicit acknowledgment of higher authorities and more precise time language. This has the overall implication that aside from the burdens stringent regulation imposes, regulations can help by improving clarity that reduces risk from applications. Thus a clearer regulation can mitigate potential disincentive effects of other parts of a regulation even while being longer. Of especially important policy implication is the positive effect from deferring to federal regulations as embodied by Topic 11 and Topic 17. By partially preempting town regulation of sites and propagating universal emissions standards, the Telecom Act seems to have made the threat of hold-up by towns less probable. The Telecom Act may therefore have had positive effects on investment. This result counters arguments put forth by Tan (1997) and Niehaus (2001) that the Telecom Act introduced greater ambiguity to siting law, increasing

220 the risk of legal disputes. More generally, this evidence contributes to the ongoing legal literature on benefits and drawbacks of the jurisdictional regime imposed by the Telecom Act and the role of the Federal government in local business regulation.52 Also of interest is what might give rise to more or less stringent regulations. For example, perhaps towns dominated by homeowners might make regulations more stringent to protect home values. To that end, I regress measures of the regulation stringency generated by the topics on the controls used in the regressions. Since the LASSO using all the regulation measures and all the data seems to be the most robust specification, I use as independent variables both the measures selected in that estimation with statistically significant coefficients, and also indices equal to the contribution of those variables to the expected reduced form profit, i.e. REGmκ. Since some topics seem to make more or less sense as proxies, I create two indices. First, there is a Core Index with only variables with the expected signs. Second, there is a Full Index that also uses variables that have unexpected signs but are still statistically significant. Figure 3.3, which plots histograms of the two indices, suggests that the Core Index does in fact more strongly represent the effect of regulation. Almost all the mass of the Core Index is left of zero, as one would expect from an indicator of regulation stringency. In contrast, the Full Index has a lot of mass to the right of zero, which would suggest a lot of towns have regulations that have a net positive impact on siting. Since intuition would suggest that regulation should almost always have a net negative impact instead, this suggests that the Full Index is including variables that are not really representative of regulation stringency. I report the results in the Table 3.11. The indices can be naturally compared to the case of no regulations as that would correspond to an index of zero. However, there is no clear way of comparing directly the topic measures to cases with no regulation at all, so in

52 For the legal literature on the impacts of the Telecom Act, see Martin (1997), Dichiara (1998), Lentz (1998), Tinari (2000), Berger (2004), and Ostrow (2011).

221 Figure 3.3: Histograms of the Core Index and the Full Index

222 Table 3.11: Measures of Regulation Stringency on Town Characteristics

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Core Topic 5 Topic 6 Topic 11 Topic 13 Topic 17 Topic 19 Topic 20 Topic 23 Topic 24 Topic 26 Full Topic 3 Topic 12 Topic 27 Index Presence Presence Focus Presence Presence Focus Presence Focus Presence Focus Index Focus Presence Focus

Executive 0.25 0.21 0.15 0.00 -0.01 -0.04 0.00 -0.00 -0.02 -0.05 0.00 0.57** -0.00 0.14 0.02* (0.36) (0.27) (0.24) (0.01) (0.08) (0.12) (0.00) (0.21) (0.01) (0.08) (0.01) (0.26) (0.01) (0.18) (0.01) Council 0.15 0.03 -0.03 0.00 0.09 0.08 -0.01* 0.07 0.00 0.39** 0.01 -0.14 0.01 0.09 -0.03** (0.26) (0.17) (0.21) (0.01) (0.17) (0.18) (0.00) (0.20) (0.01) (0.17) (0.01) (0.19) (0.01) (0.19) (0.01) Selectmen 0.48 0.59* 0.22 -0.01 0.30 0.31 0.00 0.17 -0.02 0.41** 0.02 0.32 0.00 0.26 -0.02 (0.50) (0.32) (0.35) (0.01) (0.20) (0.24) (0.01) (0.30) (0.01) (0.20) (0.01) (0.35) (0.01) (0.28) (0.02) Town Meeting -0.31 0.11 -0.10 -0.00 -0.00 -0.05 -0.00 0.13 -0.01* -0.09 -0.00 0.11 -0.00 0.20 0.01 (0.28) (0.16) (0.18) (0.01) (0.15) (0.19) (0.00) (0.18) (0.00) (0.11) (0.01) (0.19) (0.01) (0.16) (0.01) Average Turnout (Fraction) -1.18 -0.68 0.42 0.02 0.17 -0.46 0.02 0.05 0.02 -0.43 0.04* -0.89 -0.05* 0.20 0.01 (0.83) (0.53) (0.53) (0.03) (0.35) (0.42) (0.02) (0.60) (0.02) (0.33) (0.02) (0.69) (0.03) (0.46) (0.03) Democratic (Fraction) -0.69 -0.11 -0.35 -0.00 -0.04 -0.29 0.05** -1.19 -0.05 -1.34*** 0.03 0.47 -0.02 -0.10 0.08* (1.04) (0.85) (0.77) (0.04) (0.39) (0.65) (0.02) (0.81) (0.03) (0.50) (0.03) (0.83) (0.03) (0.65) (0.05) log(Population) -0.07 0.06 -0.01 -0.00 0.02 -0.05 -0.00 0.11 -0.01** -0.09** 0.00 -0.01 -0.00 0.03 0.00 223 (0.09) (0.07) (0.08) (0.00) (0.06) (0.07) (0.00) (0.07) (0.00) (0.04) (0.00) (0.08) (0.00) (0.07) (0.00) log(KM2) 0.02 -0.07 0.03 0.01** 0.02 0.17** -0.01* -0.03 0.01 0.06 -0.00 0.10 0.00 0.11 -0.01*** (0.15) (0.09) (0.10) (0.01) (0.08) (0.08) (0.00) (0.10) (0.00) (0.07) (0.00) (0.12) (0.00) (0.09) (0.00) Median HH Income 1.65 -12.15** 0.86 -0.36 -1.94 1.42 0.03 -4.06 -0.19 3.71 -0.05 -2.94 0.20 -2.51 0.10 (6.18) (5.61) (4.81) (0.27) (3.52) (4.31) (0.16) (5.36) (0.22) (3.44) (0.18) (5.13) (0.24) (4.79) (0.27) Median Home Value 1.47 1.26* 1.12* 0.14*** 0.05 0.36 -0.00 -0.03 0.08* 0.34 0.03 1.55* -0.06** 0.96* 0.05 (0.99) (0.69) (0.63) (0.03) (0.44) (0.51) (0.02) (0.79) (0.04) (0.48) (0.03) (0.90) (0.03) (0.58) (0.03) Median Home Value > $1M 0.03 1.25** 0.36 0.09*** 0.11 0.03 -0.00 0.19 0.04 -0.45 0.02 0.95 -0.04* 0.50 0.04 (0.75) (0.56) (0.50) (0.02) (0.34) (0.44) (0.02) (0.62) (0.03) (0.37) (0.02) (0.67) (0.02) (0.54) (0.03) Owned Housing Units -0.67 1.45* -0.17 0.01 0.24 0.06 0.04 1.06 -0.03 -1.29*** 0.01 0.04 -0.02 -0.09 0.03 (Fraction) (1.02) (0.79) (0.82) (0.04) (0.60) (0.63) (0.02) (0.84) (0.03) (0.48) (0.03) (0.87) (0.04) (0.77) (0.03) More than High School (Fraction) -2.26 1.97* -0.02 -0.13 0.44 -1.60* -0.02 0.16 0.01 -0.78 0.00 -2.27 -0.05 -1.36 -0.07 (1.83) (1.07) (1.15) (0.09) (0.73) (0.91) (0.03) (1.10) (0.03) (0.81) (0.04) (1.38) (0.06) (1.08) (0.05) Median Age 0.03 -0.03 0.01 0.00 -0.02 0.01 -0.00 -0.01 0.00 -0.01 -0.00 0.03 0.00 0.01 0.00 (0.03) (0.02) (0.02) (0.00) (0.01) (0.01) (0.00) (0.02) (0.00) (0.01) (0.00) (0.02) (0.00) (0.02) (0.00) Black (Fraction) -0.50 -0.36 0.23 -0.00 -0.24 0.14 -0.01 0.15 0.04* 0.43 -0.00 -1.35* -0.03 -0.43 -0.06 (1.28) (0.70) (0.90) (0.03) (0.86) (0.95) (0.02) (0.89) (0.02) (0.79) (0.04) (0.76) (0.03) (0.77) (0.05) Hispanic (Fraction) 0.52 -0.03 0.52 -0.07 -0.37 0.24 0.01 1.26 -0.02 -0.22 0.08* -0.48 -0.01 0.21 -0.02 (1.56) (0.85) (1.08) (0.05) (1.11) (1.13) (0.03) (1.27) (0.03) (0.76) (0.04) (1.10) (0.04) (0.95) (0.06)

Observations 168 136 136 136 136 136 136 136 136 136 136 168 136 136 136 R-squared 0.08 0.23 0.11 0.21 0.13 0.20 0.13 0.10 0.29 0.23 0.15 0.14 0.11 0.12 0.20 ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 the regression using the topic measures directly I remove towns without regulations. In the table, I list the Core Index first, then the measures that make it up, and then the Full Index, and then the extra measures that are included in that index. The major finding is that the controls do not strongly predict the Core Index of regulation. None of the coefficients in that regression are statistically significant. There are statistically significant variables in the individual regressions, but there is no consistent pattern and many of signs are unexpected. For example, the LASSO estimates suggest Topic 11 has a strong clarifying effect. Thus it would be in the interest of towns with high house values to focus less on it in the zoning codes to prevent more sites from being built. But the regressions here suggest higher housing values increase Focus in Topic 11. The Executive government trait, median house values, and a higher percentage of black residents all have statistically significant impacts on the Full Index, but again there is reason to doubt the Full Index is truly an appropriate proxy for regulatory stringency. Thus, there is not a strong relationship between regulatory stringency and town characteristics in the current sample. It may be the case that the regulations are more or less exogenously determined by historical accidents, though the current sample is too small to make a strong conclusion. It is therefore up to future work with richer datasets to make this relationship more explicit. Finally, I attempt to quantify the relative magnitude of the impact zoning regulations have. As explained before, the estimates do not provide any sort of welfare numbers since the reduced form profits have no concrete reference point with either costs or revenues in dollar terms. In addition, the estimates have no bearing on substitution between building a site and not building a site, since I only use the decision to build in one town rather than another for estimation. However, the estimates do imply substitution patterns between towns, conditional on sites being built. Since inefficient allocation across towns between can be a major welfare loss due to distorted towns incentives or NIMBYism, knowing how strong

224 regulations affect this conditional choice will be informative. Thus I conduct a simple counterfactual to estimate how many sites would shift between towns under a fairly drastic policy change. First, assume that all sites are built no matter how the regulatory stringency changes in the counterfactual. This would imply that whatever changes I implement they would be not large enough to induce more or less sites. Since the change I implement will be fairly drastic, this is unlikely to hold, but since I cannot model the choice not to build I will have to assume this by necessity. Second, assume that the parameters estimated give reasonable estimates of the reduced form profit functions for the decision of choosing between towns. This is because I need to directly use the model to generate new probabilities of town locations. If the estimates of the reduced form profits is inaccurate, then the results from a simulated drastic change away from the locally observed data will be misleading, to say the least. I then calculate the predicted probabilities of siting in each town using the Combination specification for the full sample of 553 sites along borders (as the other sites would theoretically not change towns given my policy change). I then calculate the expected number of sites in each location. I compare these numbers of this “baseline” model to the predicted numbers of sites in each town after subtracting from the baseline expected reduced form profit one of the following three objects: 1) the Core Index, 2) the negative components of the Core Index, and 3) the positive components of the Core Index. Respectively, 1) simulates imposing uniform telecommunications site regulations across the towns , 2) simulates imposing uniform application burdens across all towns, and 3) simulates imposing the same level of clarity across towns. These counterfactuals do not simulate the perfect allocation of the given set of sites to towns, since that allocation is unknown. Rather, the counterfactuals can be thought as if all towns were to adopt identical standards, perhaps imposed by the federal government. Given each town is different in important ways, this is likely not the welfare-optimal allocation, but

225 it at least gives one the sense of how large the impacts of such a policy could be. Moreover, since the total number of sites will be assumed constant this also does not get at the welfare losses from discouraging siting at all. Summing up the magnitude of all the differences between the towns, and then dividing by two to avoid double counting, results in 38.59 sites moved or about 6.97% of the sites total for 1); 37.90 sites or 6.85% for 2); and 20.82 sites or 3.76% for 3). What is most interesting about this result is almost all of the reallocation from changing the Core Index can be recreated by changing only burdensome regulations - clarity seems to be relatively less important. The reassignments are highly unequal - half of the predicted changes would occur in only the top sixth of the towns in the sample.53 While the predicted number of sites is not an insubstantial number, these results suggest that only a very large, targeted change in regulation would have strong reallocative effect across towns. In addition, doing a similar calculation comparing the baseline model’s predicted sites count for each town relative to actual observed numbers shows that the counterfactual difference may be swamped by the measurement error in the estimates. 84.93 sites are “shifted” between the baseline model and the observations, or 15.36% of the total. This is over twice the counterfactual change for 1) or 2) and over four times the change in 3). Thus the major welfare implications of land use regulation reform are likely to depend primarily on the extensive margin of building the site, since that might be more drastically impacted by proposed changes in the land use laws.

53These towns also have a disproportionate number of sites together: for example, for 1) towns with half of changes have slightly less than one third of the towers. So part of the inequality in shifting is simply due to the original inequality in the site distribution.

226 3.10 Conclusion

To add to the limited evidence on the effect of regulation on business investment, I study cell siting decisions using a unique dataset from Connecticut. Combining a border discontinuity approach for identification and cutting-edge computational linguistics, I find that more burdensome regulations consistently make towns less attractive for construction while improved clarity in the statutes makes towns more attractive. The latter point is, to my knowledge, novel. Moreover, the zoning codes realize this improved clarity though deference to federal standards or rules, which has implications for understanding the optimal balance between local and federal control. Looking at the actual size of the regulatory impacts, it seems burdensome regulations may have the stronger effect on the allocation of sites over towns, though the difference between the two in welfare implications and the impact on the extensive margin of construction of sites is unknown. Methodologically, the topic-base measures perform well and show that topic-based regulation measures might be a powerful tool for future research. These measures do appear to be somewhat imprecise, so would benefit from larger datasets than the one used in this application. Different regulations appear to have different mechanisms for their effects - some might be impacted by merely regulation presence while others become stronger the more the statutes focus on that type of regulation. This might suggest the best specification might include flexible nonlinear functions for each topic, which is too taxing for the data at hand even with the use of LASSO. In addition, the formal model implies interactions between different regulations, which are awkward to test with my number of observations. To gain a deeper understanding of mechanisms of regulatory impact on investment, this again suggests future work should attempt analysis of richer datasets. Finally, I limit my analysis to variation over very small regions using an estimation

227 strategy that is limited to between-town differences in regulation. Moreover, the regulatory setting mitigates the negative effects of local regulation by having a powerful state regulator. Therefore, my results should be interpreted as reduced form estimates of very local effects in a setting where they are likely to be smaller than the rest of the country. By the nature of the estimation strategy, I cannot measure welfare implications from regulatory distortions of investment nor possible negative externalities of cell sites that might justify those distortions. I thus encourage future work to explore different approaches to measure the potential welfare implications of these kinds of regulations.

228 Bibliography

Akaike, H. (1974). A new look at the statistical model identiﬁcation. Automatic Control, IEEE Transactions on 19 (6), 716–723.

Aker, J. C. and I. M. Mbiti (2010). Mobile Phones and Economic Development in Africa. Journal of Economic Perspectives 24 (3), 207–232.

Al-Ubaydli, O. and P. A. McLaughlin (2014). RegData: A Numerical Database on Industry- Speciﬁc Regulations for All US Industries and Federal Regulations, 1997-2012. Mercatus Working Paper.

An, M. Y. (1998). Logconcavity versus Logconvexity: A Complete Characteriziation. Journal of Economic Theory 80, 350–369.

Angrist, J. D. and J.-S. Pischke (2009). Mostly Harmless Econometrics. Princeton: Princeton University Press.

Bajari, P. and S. Tadelis (2001). Incentives versus transaction costs: A theory of procurement contracts. RAND Journal of Economics 32 (3), 387–407.

Bartik, T. J. (1988). The eﬀects of environmental regulation on business location in the United States. Growth and Change 19 (3), 22–44.

229 Bayer, P., F. Ferreira, and R. McMillan (2007). A Uniﬁed Framework for Measuring Prefer- ences for Schools and Neighborhoods. Journal of Political Economy 115 (4), 588–638.

Belloni, A. and V. Chernozhukov (2013). Least squares after model selection in high- dimensional sparse models. Bernoulli 19 (2), 521–547.

Berger, J. A. (2004). Eﬃcient Wireless Tower Siting: An Alternative to Section 332(C)(7) of the Telecommunications Act of 1996. Temple Environmental Law and Technology Jour- nal 23, 83–120.

Berry, S. T. (1994). Estimating Discrete-Choice Models of Product Diﬀerentiation. The RAND Journal of Economics 25 (2), 242–262.

Berry, S. T., J. Levinsohn, and A. Pakes (1995). Automobile Prices in Market Equilibrium. Econometrica 63 (4), 841–890.

Berry, S. T., J. Levinsohn, and A. Pakes (2004). Diﬀerentiated Products Demand Systems from a Combination of Micro and Macro Data: The New Car Market. Journal of Political Economy 112 (1), 68–105.

Berry, S. T. and J. Waldfogel (2001). Do mergers increase product variety? Evidence from radio broadcasting. Quarterly Journal of Economics 116 (3), 1009–1025.

Bertrand, M. and F. Kramarz (2002). Does Entry Regulation Hinder Job Creation? Evidence from the French Retail Industry. Quarterly Journal of Economics 117 (4), 1369–1413.

Bj¨orkegren, D. (2013). The Adoption of Network Goods: Evidence from the Spread of Mobile Phones in Rwanda. Working Paper.

Black, S. E. (1999). Do Better Schools Matter? Parental Valuation of Elementary Education. Quarterly Journal of Economics 114 (2), 577–599.

230 Bond, S. (2007a). Cell Phone Tower Proximity Impacts on House Prices: A New Zealand Cast Study. Paciﬁc Rim Property Research Journal 13 (1), 63–91.

Bond, S. (2007b). The Eﬀect of Distance to Cell Phone Towers on House Prices in Florida. The Appraisal Journal Fall, 362–370.

Bond, S. and K. Beamish (2005). Cellular Phone Towers: Perceived Impact on Residents and Property Values. Paciﬁc Rim Property Research Journal 11 (2), 158–177.

Bond, S. and K.-K. Wang (2005). The Impact of Cell Phone Towers on House Prices in Residential Neighborhoods. The Appraisal Journal Summer, 256–277.

B¨orsch-Supan, A. (1990). On the Compatibility of Nested Logit Models with Utility Maxi- mization. Journal of Econometrics 43, 373–388.

Bulow, J. I., J. D. Geanakoplos, and P. D. Klemperer (1985). Multimarket Oligopoly: Strategic Substitutes and Complements. Journal of Political Economy 93 (3), 488–411.

Busse, M. R. (2000). Multimarket Contact and Price Coordination in the Cellular Telephone Industry. Journal of Economics & Management Strategy 9 (3), 287–320.

Cambini, C. and Y. Jiang (2009). Broadband investment and regulation: A literature review. Telecommunications Policy 33 (10), 559–574.

Card, D. and A. B. Krueger (1994). Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania. American Economic Review 84 (4), 772–793.

Cardell, N. S. (1997). Variance Components Structure for the Extreme-Value and Logistic Distributions with Applications to Model of Heterogeneity. Econometric Theory 13 (2), 185–233.

231 Chang, H., H. Koski, and S. K. Majumdar (2003). Regulation and investment behaviour in the telecommunications sector: Policies and patterns in US and Europe. Telecommunica- tions Policy 27 (10), 677–699.

Chen, J. and J. Shao (2000). Nearest Neighbor Imputation for Survey Data. Journal of Oﬃcial Statistics 16 (2), 113–131.

Chen, J. and J. Shao (2010). Jackknife Variance Estimation for Nearest Neighbor Imputation. Journal of the American Statistical Association 96 (453), 260–269.

Chu, C. S. (2010). The Eﬀect of Satellite Entry on Cable Television Prices and Product Quality. The RAND Journal of Economics 41 (4), 730–764.

Claussen, H., L. T. W. Ho, and L. G. Samuel (2008). An Overview of the Femtocell Concept. Bell Labs Technical Journal 13 (1), 221–245.

Coase, R. H. (1960). The problem of social cost. Journal of Law and Economics 3, 1–44.

Datta, S. and K. Sudhir (2013). Does reducing spatial differentiation increase produce differentiation? effects of zoning on retail entry and format variety. Quantitative Marketing and Economics 11, 83–116.

David M. Blei and Andrew Y. Ng and Michael I. Jordan (2003). Latent dirichlet allocation. The Journal of Machine Learning Research 3, 993–1022.

Deneckere, R. and C. Davidson (1985). Incentives to Form Coalitions with Bertrand Com- petition. The RAND Journal of Economics 16 (4), 473–486.

Diamond, R. A. (2015). The Determinants and Welfare Implications of US Workers Diverging Location Choices by Skill: 1980-2000. Working Paper.

232 Dichiara, L. (1998). Wireless Communication Facilities: Siting for Sore Eyes. Buﬀalo Envi- ronmental Law Journal 6, 1–6.

Draganska, M., M. Mazzeo, and K. Seim (2009). Beyond Plain Vanilla: Modeling Joint Product Assortment and Pricing Decisions. Quantitative Marketing and Economics 7 (2), 105–146.

Dube, A., T. W. Lester, and M. Reich (2010). Minimum Wage Eﬀects Across State Borders: Estimates Using Contiguous Counties. Review of Economics and Statistics 92 (4), 945–964.

European Commission (2004). Guidelines on the assessment of horizontal mergers under the Council Regulation on the control of concentrations between undertakings. Oﬃcial Journal of the European Union.

Fan, Y. (2013). Ownership Consolidation and Product Characteristics: A Study of the U.S. Daily Newspaper Market. American Economic Review 103 (5).

Farrell, J. and C. Shapiro (1990). Horizontal Mergers: An Equilibrium Analysis. American Economic Review 80 (1), 107–126.

Federal Communications Commission (2014). Annual Report and Analysis of Competitive Market Conditions With Respect to Mobile Wireless, Including Commercial Mobile Ser- vices: Seventeenth Report.

Feinerman, E., I. Finkelshtain, and I. Kan (2004). On a political solution to the NIMBY conﬂict. American Economic Review 94 (1), 369–381.

Ferrari, S., F. Verboven, and H. Degryse (2010). Investment and Usage of New Technologies: Evidence from a Shared ATM Network. American Economic Review 100 (3), 1046–1079.

233 Fligstein, N., J. S. Brundage, and M. Schultz (2014). Why the Federal Reserve Failed to See the Financial Crisis of 2008: The Role of Macroeconomics as Sense-Making and Cultural Frame. Working Paper.

Fredriksson, P. G. (2000). The siting of hazardous waste facilities in federal systems: the political economy of NIMBY. Environmental and resource Economics 15 (1), 75–87.

Gandhi, A., L. Froeb, S. Tschantz, and G. J. Werden (2008). Post-Merger Product Reposi- tioning. The Journal of Industrial Economics 56 (1), 49–67.

Gentzkow, M. (2007). Valuing New Goods in a Model with Complementarity: Online News- paper. American Economic Review 96 (3), 713–744.

Glaeser, E. L. and B. A. Ward (2008). The causes and consequences of land use regulation: Evidence from Greater Boston. Journal of Urban Economics 65, 265–278.

Goldberg, V. P. (1976). Regulation and Administered Contracts. Bell Journal of Eco- nomics 7 (2), 426–448.

Goolsbee, A. and A. Petrin (2004). The Consumer Gains from Direct Broadcast Satellites and the Competition with Cable TV. Econometrica 72 (2), 351–381.

Gowrisankaran, G. and J. Krainer (2011). Entry and pricing in a diﬀerentiated products industry: evidence from the ATM market. The RAND Journal of Economics 42 (1), 1–22.

Grajek, M. and L.-H. R¨oller(2012). Regulation and investment in network industries: Evi- dence from European telecoms. Journal of Law and Economics 55 (1), 189–216.

Gray, W. B. (1997). Manufacturing Plant Location: Does State Pollution Regulation Matter? NBER Working Paper 5880.

234 Griﬃth, R. and H. Harmgart (2012). Supermarkets competition in england and planning regulation. The International Review of Retail, Distribution and Consumer Research 22 (1), 1–25.

Griﬃths, T. L. and M. Steyvers (2004). Finding scientiﬁc topics. Proceedings of the National Academy of Sciences 101 (Supplement 1), 5228–5235.

Grossman, G. M. and E. Helpman (1994). Protection for Sale. American Economic Re- view 84 (4), 833–850.

Grossman, S. J. and O. D. Hart (1986). The Costs and Beneﬁts of Ownership: A Theory of Vertical and Lateral Integration. Journal of Political Economy 94 (4), 691–719.

Grubb, M. D. and M. Osborne (2015). Cellular Service Demand: Biased Beliefs, Learning and Bill Shock. American Economic Review 105 (1), 234–271.

Gyourko, J. and R. Malloy (2014). Regulation and Housing Supply. NBER Working Paper 20536.

Gyourko, J., A. Saiz, and A. Summers (2008). A New Measure of the Local Regulatory Environment for Housing Markets: The Wharton Residential Land Use Regulatory Index. Urban Studies 45 (3).

Hamilton, J. T. (1993). Politics and social costs: estimating the impact of collective action on hazardous waste facilities. The RAND journal of economics 24 (1), 101–125.

Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators. Econometrica: Journal of the Econometric Society 50 (4), 1029–1054.

Hansen, S., M. McMahon, and A. Prat (2014). Transparency and Deliberation within the FOMC: A Computational Linguistics Approach. Centre for Economic Policy Research Discussion Paper No. 9994.

235 Hart, O. D. and J. Moore (1990). Property Rights and the Nature of the Firm. Journal of Political Economy 98 (6), 1119–58.

Haskel, J. and R. Sadun (2012). Regulation and UK retailing productivity: evidence from microdata. Economica 79 (315), 425–448.

Hausman, J. (1996). Valuation of New Goods under Perfect and Imperfect Competition. In T. F. Bresnahan and R. J. Gordon (Eds.), The Economics of New Goods. University of Chicago Press.

Hausman, J. (1999). Cellular Telephone, New Products, and the CPI. Journal of Business and Economic Statistics 17 (2), 188–194.

Hausman, J., G. Leonard, and J. D. Zona (1994). Competitive Analysis with Diﬀerenciated Products. Annales d’Economie´ et de Statistique (34), 159–180.

Hausman, J. and D. McFadden (1984). Speciﬁcation Tests for the Multinomial Logit Model. Econometrica 52 (5), 1219–1240.

Hoel, M. (1997). Environmental policy with endogenous plant locations. The Scandinavian Journal of Economics 99 (2), 241–259.

Holmes, T. J. (2011). The Diﬀusion of Wal-Mart and Economies of Density. Economet- rica 79 (1), 253–302.

Houde, J.-F. (2012). Spatial Diﬀerentiation and Vertical Mergers in Retail Markets for Gasoline. American Economic Review 105 (5), 2147–2182.

Ishii, J. (2007). Compatibility, Competition, and Investment in Network Industries: ATM Networks in the Banking Industry. Working Paper.

236 Jeziorski, P. (2014). Eﬀects of mergers in two-sided markets: The US radio industry. Amer- ican Economic Journal: Microeconomics 6 (4), 35–73.

Jia, P. (2008). What Happens When Wal-Mart Comes to Town: An Empirical Analysis of the Discount Retailing Industry. Econometrica 76 (6), 1263–1316.

Jiang, L. (2013). The Welfare Eﬀects of “Bill Shock” Regulation in Mobile Telecommunica- tion Markets. Working Paper.

Joskow, P. L. (1985). Vertical integration and long-term contracts: The case of coal-burning electric generating plants. Journal of Law, Economics, & Organization 1 (1), 33–80.

Kennedy, P. W. (1994). Equilibrium pollution taxes in open economies with imperfect competition. Journal of environmental economics and management 27 (1), 49–63.

Klein, B., R. G. Crawford, and A. A. Alchian (1978). Vertical Integration, Appropriable Rents and the Competitive Contracting Process. Journal of Law and Economics 21 (2), 297–326.

Kolesar, P. and E. H. Blum (1973). Square Root Laws for Fire Engine Response Distances. Management Science 19 (12), 1368–1378.

Kuzmenko, T. (2007). An Empirical model of Firm Entry with Zoning. Working Paper.

Lentz, R. G. (1998). Corporations vs. communities: Evolution of wireless services in the US and the devolution of local control. Telecommunications Policy 22 (10), 791795.

Lerner, A. P. (1934). The Concept of Monopoly and the Measurement of Monopoly Power. Review of Economic Studies 1 (3), 157–175.

Levinson, A. (1996). Environmental regulation and manufacturers’ location choices: Evi- dence from the Census of Manufactures. Journal of Public Economics 61 (1-2), 5–29.

237 Levinson, A. (1999). NIMBY taxes matter: the case of state hazardous waste disposal taxes. Journal of Public Economics 74 (1), 31–51.

Luce, R. D. (1959). Individual Choice Behavior: A Theoretical Analysis. New York: Wiley.

Luo, Y. (2011). Nonlinear Pricing with Product Customization in Mobile Service Industry. Working Paper.

Luo, Y. (2012). Bundling and Nonlinear Pricing in Telecommunications. Working Paper.

Luo, Y., I. Perrigne, and Q. Vuong (2011). Multiproduct Nonlinear Pricing: Mobile Voice Service and SMS. Working Paper.

Macdonald, V. H. (1979). Advanced Mobile Phone Service: The Cellular Concept. The Bell System Technical Journal 58 (1), 15–41.

Macher, J. T., J. W. Mayo, O. Ukhaneva, and G. Woroch (2013). Demand in a Portfolio- Choice Environment: The Evolution of Telecommunications. Working Paper.

Martin, S. L. (1997). Communications Tower Sitings: The Telecommunications Act of 1996 and the Battle for Community Control. Berkeley Technology Law Journal 12, 483–502.

Mauerer, I., W. P¨oßnecker, P. W. Thurner, and G. Tutz (2014). Modeling electoral choices in multiparty systems with high-dimensional data: A regularized selection of parameters using the lasso approach.

Mazzeo, M., K. Seim, and M. Varela (2014). The Welfare Consequences of Mergers with Endogenous Product Choice. Working Paper.

McFadden, D. (1978). Modelling the Choice of Residential Location. In A. Karlqvist, F. Snickars, and J. Weibull (Eds.), Spatial Interaction Theory and Planning Models. Am- sterdam: North Holland.

238 McFadden, D., K. Train, and W. B. Tye (1977). An application of diagnostic tests for the independence from irrelevant alternatives property of the multinomial logit model. University of California: Institute of Transportation Studies.

Milgrom, P. and C. Shannon (1994). Monotone Comparative Statics. Econometrica 62 (1), 157–180.

Miravete, E. and L. H. R¨oller(2004). Competitive Nonlinear Pricing in Duopoly Equilibrium: The Early U.S. Cellular Telephone Industry. Working Paper.

Morrow, W. R. and S. J. Skerlos (2011). Fixed-Point Approaches to Computing Bertrand- Nash Equilibrium Prices Under Mixed-Logit Demand. Operations Research 59 (2), 328– 345.

Moszoro, M. W., P. T. Spiller, and S. Stolorz (2015). Rigidity of Public Contracts. NBER Working Paper No. 21186 .

Nevo, A. (2000). Mergers with Diﬀerentiated Products: The Case of the Ready-to-Eat Cereal Industry. The RAND Journal of Economics 31 (3), 395–421.

Niehaus, S. E. (2001). Bridging the Signiﬁcant Gap: To What Extent does the Telecom- munications Act of 1996 Contemplate Seamless Service. Notre Dame Law Review 77, 641–672.

Nishida, M. (2014). The costs of Zoning regulations in retail chains: the case of the City Planning Act of 1968 in Japan. Journal of Regulatory Economics 45, 304–328.

Nishida, M. (2015). Estimating a Model of Strategic Network Choice: The Convenience-Store Industry in Okinawa. Marketing Science 34 (1), 20–38.

239 Nitsche, R. and L. Wiethaus (2011). Access regulation and investment in next generation networksA ranking of regulatory regimes. International Journal of Industrial Organiza- tion 29 (2), 263–272.

Olson, M. (1971). The logic of collective action. Harvard University Press.

Ostrow, A. P. (2011). Process Preemption in Federal Siting Regimes. Harvard Journal on Legislation 48, 289–341.

Perry, M. K. and R. H. Porter (1985). Oligopoly and the Incentive for Horizontal Merger. American Economic Review 75 (1), 219–227.

Petrin, A. and K. Train (2010). A Control Function Approach to Endogeneity in Consumer Choice Models. Journal of Marketing Research 47 (1), 3–13.

Pogodzinski, J. M. and T. R. Sass (1990). The economic theory of zoning: a critical review. Land Economics 28 (4), 294–314.

Quinn, K. M., B. L. Monroe, M. Colaresi, M. H. Crespin, and D. R. Radev (2010). How to Analyze Political Attention with Minimal Assumptions and Costs. American Journal of Political Science 54 (1), 209228.

Ridley, D. B., F. A. Sloan, and Y. Song (2011). Retail Zoning and Competition. Working Paper.

R¨oller,L.-H. and L. Waverman (2001). Telecommunications Infrastructure and Economic Development: A Simultaneous Approach. American Economic Review 91 (4), 909–923.

Salant, D. J. and G. A. Woroch (1991). Crossing Dupuit’s Bridge Again: A Trigger Policy for Eﬃcient Investment in Infrastructure. Contemporary Economic Policy 9 (2), 101–114.

240 Salant, D. J. and G. A. Woroch (1992). Trigger price regulation. The RAND Journal of Economics 23 (1), 29–51.

Salant, S. W., S. Switzer, and R. J. Reynolds (1983). Losses from horizontal merger: the effects of an exogenous change in industry structure on Cournot-Nash equilibrium. Quarterly Journal of Economics 98 (2), 185–199.

Schivardi, F. and E. Viviano (2011). Entry barriers in retail trade. The Economic Jour- nal 121 (551), 145–170.

Schwartz, A. and J. Watson (2013). Conceptualizing contractual interpretation. The Journal of Legal Studies 42 (1), 1–34.

Schwarz, G. (1978). Estimating the dimension of a model. The annals of statistics 6 (2), 461–464.

Seim, K. and J. Waldfogel (2013). Public Monopoly and Economic Eﬃciency: Evidence from the Pennsylvania Liquor Control Board’s Entry Decisions. The American Economic Review 103 (2), 831–862.

Shaked, A. and J. Sutton (1982). Relaxing Price Competition Through Product Diﬀerenti- ation. Review of Economic Studies 49, 3–13.

Shavell, S. (2006). On the Writing and the Interpretation of Contracts. Journal of Law, Economics, and Organization 22 (2), 289–314.

Sinkinson, M. (2014). Pricing and Entry Incentives with Exclusive Contracts: Evidence from Smartphones. Working Paper.

Small, K. A. and H. S. Rosen (1981). Applied Welfare Economics with Discrete Choice Models. Econometrica 49 (1), 105–130.

241 Stock, J. H., J. H. Wright, and M. Yogo (2002). A Survey of Weak Instruments and Weak Identiﬁcation in Generalized Method of Moments. Journal of Business & Economic Statis- tics 20 (4), 518–529.

Sutton, J. (1991). Sunk Costs and Market Structure: Price Competition, Advertising and the Evolution of Concentration. MIT.

Sutton, J. (2001). Technology and Market Structure: Theory and History. MIT.

Suzuki, J. (2013). Land Use Regulation as a Barrier to Entry: Evidence from the Texas Lodging Industry. International Economic Review 54 (2), 495523.

Sweeting, A. (2010). The eﬀects of mergers on product positioning: evidence from the music radio industry. The RAND Journal of Economics 41 (2), 372–397.

Tan, G. (1997). Wading Through the Rhetoric of the Telecommunications Act of 1996: Uncertainty of Local Zoning Authority over Wireless Telecommunications Tower Siting. Vermont Law Review 22, 461–491.

Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological) 58 (1), 267–288.

Tinari, N. (2000). Cell Phone Towers in Residential Areas: Did Congress Let the Pig in the Parlor with the Telecommunications Act of 1996? Temple Law Review 73, 269–294.

Topkis, D. M. (1978). Minimizing a Submodular Function on a Lattice. Operations Re- search 26 (2), 305–321.

Town, R. and G. Vistnes (2001). Hospital competition in HMO networks. Journal of Health Economics 20 (5), 733753.

242 Turner, M. A., A. Haughwout, and W. van der Klaauw (2014). Land Use Regulation and Welfare. Econometrica 82 (4), 13411403.

U.S. Department of Justice and Federal Trade Commission (2010). Horizontal Merger Guide- lines.

Waldfogel, J. (2008). The Median Voter and the Median Consumer: Local Private Goods and Population Composition. Journal of Urban Economics 63, 567–582.

Werden, G. J. and L. M. Froeb (1994). The Eﬀects of Mergers in Diﬀerentiated Products Industries: Logit Demand and Merger Policy. Journal of Law, Economics, & Organiza- tion 10 (2), 407–426.

Wikle, T. A. (2002). Cellular Tower Proliferation in the United States. The Geographic Review 92 (1), 45–62.

Williamson, O. E. (1975). Markets and Hierarchies : Antitrust Implications. New York: The Free Press.

Williamson, O. E. (1979). Transaction Cost Economics: The Governance of Contractual Relations. Journal of Law and Economics 22 (2), 233–262.

Williamson, O. E. (1985). The economic institutions of capitalism. New York: The Free Press.

Yao, S., C. F. Mela, J. Chiang, and Y. Chen (2012). Determining Consumers’ Discount Rates with Field Studies. Journal of Marketing Research 49 (6), 822–841.

Zellner, A. and H. Theil (1962). Three-stage least squares: simultaneous estimation of simultaneous equations. Econometrica 30 (1), 54–78.

Zheng, F. (2014). Spatial Competition and Preemptive Entry in the Discount Retail Industry.

243 Zhu, T., H. Liu, and P. K. Chintagunta (2011). Wireless Carriers’ Exclusive Handset Ar- rangements: An Empirical Look at the iPhone. Working Paper.

Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American statistical association 101 (476), 1418–1429.

244 Appendices

245 Appendix A

Appendix for Chapter 1

A.1 A Simple Model of Quality and Base Station Den-

sity

Mobile telephony is called “cellular” in the United States due to the practice of dividing space up into discrete “cells” served by separate base stations. Each grouping or “cluster” of base stations has access to all the firms’ licensed frequencies. If consumers move out of the range of a cluster’s cell into a new area, they are simply transferred to the cell that covers that area and its assigned frequency. In this way, a firm can reuse a limited amount of frequency, and this innovation made mass adoption of mobile phones possible. Given uniform distribution of users over space and completely flat terrain, the most efficient base station deployment distribution has base stations at the centers of identical regular hexagons that tile the space completely. Within each hexagon, the base station at the center is the closest base station, so determining the average distance between consumers and their nearest base station is simply a matter of finding the average distance between the points in a hexagon and its center. I can further tile the hexagon into 12 similar right

246 Figure A.1: From Macdonald (1979) - The Paper that Proposed the Cellular Phone Concept

Figure A.2: Each regular hexagon can be divided into 12 right triangles. a

c = 2a b = p(3)a

247 √ 3 triangles with sides of length a, b = 2 a and c = 2a, so the exercise reduces to ﬁnding the average distance between vertex bc and all the points in triangle abc. Assume what consumers care about is simply the average power of the call which determines the dropped call rate, which is inversely proportional to distance from the base station, d. To make sure the utility is deﬁned at all points assume that it takes the form:

1 U(d) = (A.1) C2 + d2

where C is some positive constant. C ensures that if an individual is right next to a base station (d = 0) their utility does not go to inﬁnity. Under the assumption of uniformly distributed consumers over the entire space, the consumer is also uniformly distributed along the line segment from vertex bc to some point of side a. Call the length of the line segment L. Index the line segment by its angle from side b, θ. The average utility from a call along this segment θ is:

Z L(θ,a) U(d) tan−1(L(θ, a)/C) E[U(d)|θ ∈ X] = = (A.2) 0 L(θ, a) CL(θ, a)

One can show that

√ 6 L(θ, a) = a(p(3) + (1 − 3) θ) (A.3) π

The average over the entire right triangle, and thus the entire hexagon and the whole

248 space is then found by simply integrating over θ:

π −1 Z 6 tan (L(θ, a)/C) 6 E[U(d)|a] = dθ (A.4) 0 CL(θ, a) π

This integral does not have a closed form solution, but numerical evaluations show that it

1 is a non-decreasing concave function in a as long as a and C are positive. If the N hexagons A are apportioned to all the area in a market, A, then each hexagon gets the area N . Thus as N → +∞:

r √ A 1 √ N 3 3a2 = ⇐⇒ = 3 3 (A.5) N a A

1 N Since a is a concave function of density, then E[U(d)| A ] is also a concave function in density. Geography and location availability cause base stations to be deployed in non-regular patterns, violating the uniformity assumption, but relaxing the assumptions are likely to make the density function even more concave as worse locations would be used later by optimizing ﬁrms.

249 Appendix B

Appendix for Chapter 2

B.1 Comparative Statics of the Example with More

than 2 Firms

Following the setup of Section 2.2, the probability of i choosing a carrier k can be written as the joint CDF of the diﬀerences of the shocks. Without loss of generality, index this carrier

by 1, the diﬀerences in mean utilities by ∆1k and the diﬀerences in the errors by Ek1:

P r(Ui1 =max{Uik ∀ k ∈ K}) (B.1)

=P r(∆12 ≥ E12, ∆13 ≥ E31, ..., ∆1K ≥ EK1) (B.2)

=G(∆12, ..., ∆1K ) (B.3)

where G is the joint CDF for all pairwise diﬀerences with shocks k. For greater clarity, I abuse notation by referring to the joint distribution of subsets of the shocks by G as well, appropriately reducing the dimension as needed. Further denote the marginal of these distributions by g.

250 Denote an arbitrary carrier by 2 without loss of generality. The cross partial of the profit of firm 1 with respect to firm 2 is now:

2 ∂ G 0 = −g (∆12)G(∆13, ...∆1K |∆12) ∂δ1∂δ2 | {z } Direct eﬀect on substitution between 1 and 2 ! ∂G(∆13, ...∆1K |∆12) X −g(∆12) − g(∆1k)G(∆13, ..., ∆1,k−1, ∆1,k+1, ..., ∆1K |∆1k∆12) ∂∆12 k6=1,2 | {z } Indirect eﬀect on substitution between 1 and all other goods (B.4)

There are two parts to this equation, the first part which represents the direct effect on substitution between 1 and 2, and the second part which represents the indirect effect on substitution between 1 and every other good. In the two good case, G(∆13, ...∆1K |∆12) is

0 completely degenerate, so the ﬁrst part is g (∆12) and the second part does not exist. Thus the sign is the negative sign of the slope of the PDF in the two goods case. That aspect of

the equation is still expressed somewhat in the general equation since G(∆13, ...∆1K |∆12) is

0 always positive so g (∆12) will have the same sign. However, G(∆13, ...∆1K |∆12) < 1 so the eﬀect is smaller, and the second part is always negative. The general case is therefore more predisposed to strategic substitutes. If the number of goods is very numerous, the sum of the conditional marginals in the second part will clearly dominate. Under a joint distribution with shrinking thin tails, this implies strategic

complements if ∆1k is large for all k. Then the first term will be positive, and all the conditional marginals will be very small, so the second term overall will be small. In general, without further restrictions, whether the first part and thus the same comparative statics of the two goods case dominate depends on whether the joint distribution makes it so that the second part is always relatively small compared to the first part. I conjecture that log concavity of the joint shock distribution is sufficient for this, as it implies

251 unimodality and shrinking tails for the shock diﬀerence distributions.

B.2 General Comparative Statics of the Merger Sce-

narios

Assume the setup of the static Nash stage game in quality from Section 2.2. For full generality, consider the proﬁt function πk which is equal to total revenue Rk minus total cost function φ:

πk(Q) = Rk(Q) − φ(Qk) (B.5)

The necessary condition for a Pure Nash equilibrium is

∂π ∂R (Q) ∂φ(Q ) k = k − k (B.6) ∂Qk ∂Qk ∂Qk and the cross partial in the quality of ﬁrm h is

∂2π ∂2R (Q) k = k (B.7) ∂Qk∂Qh ∂Qk∂Qh

Under constant absolute markups, any derivative of R will simply be 1) a sum over each plan type and 2) within each plan type the product of the markup, market population and the share function. Any condition assumed about the derivatives of R will therefore actually be conditions on the derivatives of the share functions. Consider when ﬁrms k and h merge. Assuming the cross partial is negative locally, quality is a strategic substitute. In Scenario ∗, when h is dropped, nothing changes about the form of the above equations. Interpreting discontinuation as an inﬁnite decrease in quality, the

252 remaining firm will increase quality. Consider Scenario ∗∗, where joint firm ∗∗ of h and k internalizes the cannibalization effect of quality. In effect, this adds an additional term to the first order condition for k, relative to the equation found in *:

∂π∗∗ ∂π ∂π k = k + h (B.8) ∂Qk ∂Qk ∂Qk

The last term represents lost revenue for good h from the quality of k. The cannibalization eﬀect is thus negative and reduces the incentive to provide quality of both k and h. Again, the overall results will be ambiguous assuming strategic substitutes. Furthermore, the cross partials for the insiders is now diﬀerent:

∂2π∗∗ ∂2π ∂2π = k + h (B.9) ∂Qk∂Qh ∂Qk∂Qh ∂Qh∂Qk

The cross partial now essentially includes the cross partial for the other product h. Assuming that both of these terms are still negative at the new equilibrium, the cross partial is even more negative than it was before. In the case where the insiders are very asymmetric in costs or exogenous quality, the joint firm has a large incentive to differentiate their products. Consider next the case where carriers k and h are still placed under the joint management of firm, the products are retained. Denote the counterfactual and this joint firm as ∗ ∗ ∗. Posit that there are network spillovers in the sense now callers on one network can use some of a rival’s network. Call the fraction of each other’s network that can be used ρ ≤ 1. Effective quality of k is

∗∗∗ Qk = Qk + ρQh (B.10)

which enters into utility of k customers instead of Qk alone.

253 It turns out that the ﬁrst order condition for k (and analogously h) can be expressed in the following way:

∗∗∗ ∗∗ ∗∗ ∗∗∗ ∂π ∂πk ∂πh ∂φ(Qh ) = ∗∗∗ + ρ ∗∗∗ − ∗∗∗ = 0 (B.11) ∂Qk ∂Qk ∂Qh ∂Qh

∗∗∗ That is, take the FOC for k for Scenario ∗∗, replace the Qk for Qk , and then add the same for the other firm h but subtracting the cost component and multiplying by ρ. This second term is the spillover, which has only the benefit of quality but not the cost. The firm

∗∗∗ now provides effective quality Qk at the cost of the quality specific to that network Qk in Scenario ∗∗. The cross partials also include extra terms representing spillovers when only one of the firms in question is one of the merging firms. Under strategic substitutes these are negative and so would induce stronger strategic substitutes. In particular, the cross partial for network 1 of the merged firm is:

2 ∗∗∗ ∗∗2 ∗∗2 ∗∗2 ∂ π 2 ∂ π1 ∂ πk ∂ πh = (1 + ρ ) ∗∗∗ ∗∗∗ + ρ ∗∗∗ ∗∗∗ + ∗∗∗ ∗∗∗ (B.12) ∂Qk∂Qh ∂Qk ∂Qh ∂Qk ∂Qk ∂Qk ∂Qk

The first product is simply the second order condition of the merger with joint control but separate networks multiplied by a factor of 1 + ρ2. The term inside the parenthesis represents the concavity of the problem for the firms, so they must be negative. Thus the whole term is negative. The results for the above scenario are far less ambiguous if quality is a strategic complement. The discontinued product in ∗ and the internalization in ∗∗ would induce drops in quality by all firms so the effect for consumers would be clearly negative. The efficiencies from ∗∗∗ would also spur the non-merging firms to increase quality. However, the incentives of the firms inside the merger are ambiguous since in this case, since Equation (B.12) is now

254 not necessarily negative.

B.3 Comparative Statics Under Multi-Product Logit

Demand Model

As noted in Section 2.2, strategic substitutability depends entirely on the cross partial derivative of the demand function. In the multiple plan-type case this is the sum of cross partials for each plan-type. For each plan type, and suppressing the market and time subscripts, this term is:

2 Z P P ∂ Djk(N) Sijk(N) l∈J Silk0 (N) 1 − 2 l∈J Silk(N) = − γik di (B.13) ∂Nk∂Nn NkNh

γik is allowed to vary by consumer to admit the possibility of random coeﬃcients. For each consumer the cross partial is a product, so the sign of whole product can be

deduced from the signs of its components. γik is assumed always positive. Shares are always positive, while the base station counts in the denominator are always positive. Thus there P is only one term that can be negative, 1 − 2 l∈J Silk(N), and that sign is contingent on whether the predicted probability is less or more than 1/2. If consumers are identical, then the total market share of all the firm’s products is pivotal since the integration does nothing. If share is less than 1/2 then the whole term is negative and quality is a strategic substitute; if it is more the whole term is positive and then a strategic complement. If consumers differ, either because of consumer heterogeneity or random coefficients, then it is ambiguous. For example, assume there are rich and poor consumers: k has almost a pure monopoly on rich consumers but sells almost nothing to poor consumers, and h has almost a pure monopoly on poor consumers but sells almost nothing to rich consumers. If h increases its quality, then it steals both rich and poor con-

255 sumers from k, but the eﬀect almost purely from the rich consumers since k’s shares of rich consumers is bigger. At a near monopoly of rich consumers, the marginal rich consumer must have a low taste shock for k, since rich consumers with only the lowest taste shocks have decided not to choose k. Lowering the relatively quality slightly then must change the marginal rich consumer to those with a more moderate taste shock, which is more numerous in logit. Then there is an increase in the number of marginal consumers from k’s quality increase, since the eﬀect for rich consumers dominates. The rival increase in quality of h thus increases incentives for quality improvement for k, so quality is a strategic substitute. This is regardless of the actual total share that k has, as long as k has a near monopoly in a particular subgroup and low shares in other groups.

256 B.4 Additional Tables and Figures

Table B.1: Hausman-McFadden Tests of Independence of Irrelevant Alternatives for the Demand Model

Carrier Removed T-Stat Deg. Free. 5% Crit. Val.

AT&T -118.90 836 904.37 Sprint 7.43 891 961.55 T-Mobile -12.69 837 905.42 Verizon 22.19 836 904.37 Other -95.09 844 912.70

Reports Chi-Square test of diﬀerence between estimated pa-

rameters and parameters estimated using data without 1) the

products from the removed ﬁrm and 2) all individuals who

choose the removed products.

257 Table B.2: Postpaid-Carrier-Year Fixed Eﬀects from Pure Logit

η AT&T Sprint T-Mobile Verizon

2008 -0.085 -2.15*** -2.96*** 0.93 (0.30) (0.36) (0.33) (0.67) 2009 -0.27 -2.23*** -2.67*** 1.21 (0.32) (0.38) (0.33) (0.67) 2010 0.13 -2.23*** -3.20*** 0.88 (0.33) (0.37) (0.38) (0.68) 2011 -0.03 -2.29*** -3.28*** 0.77 (0.32) (0.39) (0.37) (0.71) 2012 0.00 -2.23*** -2.97*** 0.76 (0.33) (0.39) (0.36) (0.72)

***, **, * indicate 1%, 5% and 10% signiﬁcance,

respectively.

The above represents the mean utility of post-

paid plans net of signal quality, for women between

the ages of 35 and 64, in multiple-person house-

holds(families) that earn from $50-75 thousand annually.

258 Table B.3: Correlation in Mean Utility Attributable to Only Demographics and Years

AT&T Sprint T-Mobile Verizon Other Pre Post Post Pre Post Pre Post Pre Post

AT&T Pre 1.00 ------Post 0.72 1.00 ------Sprint Pre 0.58 0.91 1.00 ------T-Mobile Pre 0.45 0.66 0.69 1.00 - - - - - Post 0.53 0.81 0.87 0.64 1.00 - - - - Verizon Pre 0.63 0.69 0.59 0.41 0.58 1.00 - - - Post 0.69 0.96 0.88 0.61 0.76 0.62 1.00 - - Other Pre -0.17 -0.06 0.01 -0.02 0.01 -0.14 0.10 1.00 - Post -0.03 0.02 0.07 0.01 -0.03 -0.23 0.02 0.58 1.00

This table displays the correlation of the diﬀerence between the estimated product utility for a consumer and the brand-market utility, δijkt − ξkmt, for the Nielsen sample.

259 Appendix C

Appendix for Chapter 3

C.1 LDA Algorithm Details

C.1.1 Text Transformation

The text is transformed in several ways to ease estimation. First, very common three or two word compound words or phrases which are meaningful are treated as single words. This helps conserve some of the syntactical meaning loss by the “bag of words” nature of LDA. I report the list of these “multigrams” and the replacements less spaces in Table C.1. The list was constructed by finding two or three word phrases with at least twenty-five appearances in the entire body of text and then selecting ones which were meaningful. Since some phrases are quite similar but are different by slight word form, i.e. “wireless telecommunication facilities” versus ‘wireless telecommunications facilities,” I replace some phrases with a common compound word. Second, I remove words which ex post might not be very informative when included in a topic. All words containing non-alphabetic characters are removed, which removes all

260 Table C.1: Multigram Replacement List

Multigram Replacement Multigram Replacement Multigram Replacement national register of historic places nrhp permitted uses permitteduse telecommunications facility telecomfacilit national historic preservation act nhpa telephone number phonenumber telecommunication facility telecomfacilit telecommunications act of 1996 telecomact property lines propline telecommunications facilities telecomfacilit telecommunications act telecomact property line propline telecommunication facilities telecomfacilit national environmental policy act nepa propagation modeling propmodel telecommunication sites telecomfacilit access road accessroad propagation model propmodel telecommunication site telecomfacilit access roads accessroad property values propval communications facilities telecomfacilit adequate coverage adequatecoverage property value propval communications facility telecomfacilit amateur radio amateurradio public health publichealth communications towers telecomtower american national standards americannationalstandards public hearing publichearing communication towers telecomtower application fee applicationfee public road publicroad telecommunications tower telecomtower base station basestation public safety publicsafety telecommunication tower telecomtower connecticut siting council csc radial plots radialplot tower sites telecomtower siting council csc radial plot radialplot tower site telecomtower dish antennas dishantennas electromagnetic waves rfr facility sites telecomfacilit dish antenna dishantennas electromagnetic wave rfr facility site telecomfacilit emergency services emergencyservice electromagnetic emissions rfr tower failure towerfailure emergency service emergencyservice electromagnetic emission rfr utility poles utilitypole environmental protection agency epa frequency emissions rfr evergreen screen vegscreen equipment cabinets equipmentcabinets radiofrequency emissions rfr vegetative screening vegscreen equipment cabinet equipmentcabinets radiofrequency emission rfr view shed analysis viewshedanalysis equipment shelters equipshelters radio frequency emissions rfr visual impact visualimpact equipment shelter equipshelters radiofrequency radiation rfr visual impacts visualimpact federal aviation administration faa radio frequency emission rfr visual eﬀects visualimpact facility owner facilityowner frequency emission rfr water towers watertower fall zones fallzone frequency emissions rfr water tower watertower fall zone fallzone radio frequencies rfr whip antennas whipantenna federal communications commission fcc radio frequency rfr zoning commission zoningcomm good faith goodfaith ridge lines ridgeline zoning districts zoningdistricts ground elevation groundelev ridge line ridgeline zoning district zoningdistricts ground level groundlevel roof mounted roofmount zoning regulations zoningregs ground mounted groundmounted satellite dish satellitedish zoning regulations zoningreg guy wires guywire security barrier securitybarrier special uses permits special use guy wire guywire service area servicearea special use permits special use historic district historicdistrict service provider serviceproviders special uses permit special use historic districts historicdistrict service providers serviceproviders special use permit special use industrial zones industrialzone sight line sightline special uses special use industrial zone industrialzone site plan siteplan special use special use licensed carrier licensedcarrier special exception specialexcept tower sharing towersharing maximum height maxheight special permit specialpermit general statutes generalstatutes microwave dish microwavedish square feet squarefeet supporting structures supportstructures minimum distance mindist square foot squarefeet supporting structure supportstructures minimum height minheight square meter squaremeter support structures supportstructures tower height towerht square meters squaremeter support structure supportstructures minimum lot size minlotsize wireless communications services wcs permit applications permit application minimum lot area minlotsize wireless telecommunications services wcs permit application permit application lot size lotsize wireless communication services wcs specialized mobilized radio mobileradio lot area lotsize wireless communications service wcs mobile radio mobileradio minimum property lines minpropline wireless telecommunications service wcs town plan townplan minimum property line minpropline wireless telecommunication services wcs structural integrity structinteg minimum separation distances minsepdist wireless communication service wcs minimum separating distances minsepdist wireless telecommunication service wcs separation distances sepdist telecommunication services telecomservices separating distances sepdist telecommunication service telecomservices environmental policy act nepa communication services telecomservices noise regulations noiseregulation communication service telecomservices panel antennas panelantenna telecommunications facility sites telecomfacilit panel antenna panelantenna telecommunications facility site telecomfacilit personal communication services pcs telecommunication facility sites telecomfacilit permitted use permitteduse telecommunication facility site telecomfacilit

261 numbers and punctuation. Almost all very short words of one or two letters are removed.1 I then remove words that are mostly present for grammatical reasons and thus carry little information by themselves and are extremely common across documents. These are called “stop words” in the literature. There is no universal list of stop words, so I use the list used by Hansen, McMahon, and Prat (2014).2 This list includes all English pronouns, auxiliary verbs, and articles. In addition to the conventional stop words, I remove words which are common due to the setting but have very little regulatory meaning. This includes

• All written out versions of numbers “one” to “twenty,” and then every tenth number until a “hundred,” and the numbers “thousand” and “million.”

• All town names, “Connecticut,” and the cardinal directions (which show up in town names).

• Extremely common words that are likely to be used in wireless telecommunications facilities sections of zoning codes: “section,” “wireless,” and “cell”.

Next, I use the default Porter Stemmer from Python’s natural language toolkit to reduce the remaining words into “stems,” which can be thought of as approximating linguistic roots. Stems can be Natural English words like when “effects” and “effecting” both become “effect”; but more usually results in a stems that are not words, such as well “fence” and “fencing” become “fenc”. Still the result is desirable, as the stems join vocabulary that is essentially the same in meaning but may be different simply due to word form. Many of these stems will still not be ex post informative either because they appear only infrequently in the corpus overall or appear in a very large percentage of documents like the stop words. Given a widely

1“Radiofrequency radiation” is usually shortened as “RFR,” but sometimes as “RF.” In practice, I replace radiofrequency radiation, RF and many related words with RFR. This prevents RF from being removed in this step. I report this substitution in C.1.

2This list can be downloaded at http://snowball.tartarus.org/algorithms/english/stop.txt .

262 Figure C.1: Value of tf-idf of Stem Over its Descending Ranking

5 0 500 1000 1500 2000 2500 3000 3500

used measure for informativeness, I find a natural cutoff and drop all stems above this cutoff in the measure.3 After this processing, I am left with 1,700 of what I shall call “tokens,”

which will serve as the word objects, wd, in the LDA model.

3The measure of informativeness is term-frequency, inverse-document-frequency (tf-idf) score:

D log(1 + Nv)log( ) (C.1) Dv

where Nv is the number of times stem v appears in the entire body of text, D is the total number of documents and Dv are the total number of documents stem v appears in. The ﬁrst term, log(1 + Nv), rewards stems that are common while the second, log( D ), penalizes stems that appears into too many documents and Dv thus are not informative when they appear. In this application the ﬁrst term is the most meaningful: given the use of sentences as documents which are short and numerous, stems that are not stop words are unlikely to appear in almost all sentences.

263 C.1.2 Estimating Algorithm

Once the text is converted to tokens, I perform the Gibbs Sampling procedure to generate the topic distributions. This can be done since one can calculate the conditional probability that a particular observed word is in a particular topic taking K, α, η, the distribution of words in all documents w and the other topic assignments for all other words as given. Let several variables, denoted by m, represent diﬀerent counts:

d • mk - Count of words in document d allocated to k.

k • mv - Count of times token v is allocated to topic k.

d • mk,−n - Excluding token n, the count of words in d allocated to k.

k • mv,−(d,n) - Excluding token n in document d, the count of unique token v allocated to k.

Then an estimate of the conditional probability of a particular “token” being in a topic is:

mk + η vd,n,−(d,n) d P r[zd,n|z , w] ∝ (m + α) (C.2) −(d,n) PV k k,−n v=1(mv,−(d,n) + η)

The estimating procedure continues in the following steps:

1. Assign tokens to topics randomly to start.

d k 2. Calculate mk and mv.

3. For each token in each documents

d k (a) Calculate mk,−n and mv,−(d,n).

(b) Assign a new topic for wd,n according probabilities calculated via (C.2).

d k (c) Calculate mk and mv with wd,n in its new topic.

264 (d) Repeat for the next token until all words have new topics.

4. Repeat many times until convergence.

For any particular draw for all the tokens, the estimated probability that a token of the vocabulary is in a topic is

mk + η βˆv = v (C.3) k PV k v=1(mb + η) and the estimated probability of topic being in a document is

md + η θˆv = k (C.4) k PK d k=1(mk + η)

C.1.3 Town-Level Topic Distributions

Formally, denote each town by i ∈ I. Deﬁne θi as the topic distribution of the town’s section,

i wi,n as the nth word in the section, zi,n its topic assignment, vi,n its token index, mk the

i number of section i’s words assigned to topic k, and mk,−n that number excluding the current word n. Also deﬁne z−(i,n) as the vector of topic assignments in section i excluding word n and wi as the vector of words in section i. For each draw that is used to estimate the distribution {20050, 20100, ..., 40000}:

i 1. Calculate mk.

2. For each word wi,n:

i (a) Calculate mk,−n.

(b) Assign new topic to wi,n according to

ˆvi,n d P r[zi.n = k|z−(i,n), wi] =∝ βk (mk,−n + α) (C.5)

265 3. Repeat 30 times.

The chain-levels estimate of the town level distributions is now

mi + α θˆk = k (C.6) i PK i k=1 mk + α

The ﬁnal estimate is the average of this measure over all the chains.

266 C.2 Additional Tables

Table C.2: Correlation Matrix of Coded Regulation Variables

Consultant Glossary Higher Landscaping Hearing Radiation Radiation Size Re- Visual Authority Extra strictions Impact

Consultant 1.00 Glossary 0.27 1.00 Higher Authority 0.28 0.42 1.00 Landscaping 0.27 0.32 0.65 1.00 Hearing 0.24 0.37 0.33 0.28 1.00 Radiation 0.21 0.42 0.40 0.22 0.28 1.00 Radiation Extra 0.15 0.33 0.29 0.27 0.36 0.70 1.00 Size Restrictions 0.27 0.42 0.74 0.53 0.26 0.43 0.31 1.00 Visual Impact 0.26 0.42 0.81 0.63 0.31 0.39 0.30 0.70 1.00

Correlations across the 168 towns besides Cornwall.

267 Table C.3: Correlation Matrix of Topic Focus

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

1 1.00 2 0.49 1.00 3 0.40 0.38 1.00 4 0.44 0.24 0.09 1.00 5 0.42 0.25 0.02 0.65 1.00 6 0.31 0.19 0.20 0.14 0.16 1.00 7 0.43 0.43 0.52 0.27 0.21 0.42 1.00 8 0.22 0.09 0.18 0.37 0.46 0.17 0.31 1.00 9 0.28 0.39 0.39 0.31 0.31 0.25 0.49 0.30 1.00 10 0.23 0.38 0.40 0.03 -0.04 0.23 0.41 0.11 0.51 1.00 11 0.45 0.40 0.05 0.67 0.65 0.12 0.40 0.42 0.27 0.07 1.00 12 0.34 0.33 0.36 0.25 0.30 0.12 0.25 0.33 0.29 0.17 0.31 1.00 13 0.33 0.32 0.31 0.31 0.29 0.18 0.42 0.32 0.42 0.41 0.27 0.27 1.00 14 0.30 0.34 0.34 0.39 0.37 0.30 0.31 0.41 0.50 0.33 0.33 0.18 0.33 1.00 15 0.19 0.23 0.27 0.32 0.35 0.06 0.22 0.44 0.39 0.34 0.29 0.33 0.47 0.37 1.00 16 0.18 0.37 0.43 0.24 0.21 0.27 0.38 0.20 0.58 0.36 0.21 0.25 0.36 0.63 0.44 1.00 17 0.25 0.39 0.33 0.08 0.04 0.29 0.41 0.15 0.46 0.70 0.12 0.22 0.44 0.18 0.36 0.26 1.00

Correlations across the 168 towns besides Cornwall.

268 Table C.4: Correlation Matrix of Topic Volume

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

1 1.00 2 0.92 1.00 3 0.87 0.88 1.00 4 0.92 0.88 0.83 1.00 5 0.91 0.86 0.80 0.95 1.00 6 0.87 0.87 0.84 0.84 0.82 1.00 7 0.91 0.91 0.92 0.89 0.87 0.90 1.00 8 0.86 0.83 0.84 0.90 0.90 0.84 0.89 1.00 9 0.89 0.92 0.91 0.89 0.88 0.89 0.93 0.89 1.00 10 0.84 0.88 0.89 0.79 0.76 0.85 0.89 0.82 0.91 1.00 11 0.93 0.90 0.82 0.96 0.95 0.83 0.90 0.90 0.89 0.81 1.00 12 0.85 0.84 0.88 0.85 0.84 0.81 0.86 0.85 0.87 0.82 0.86 1.00 13 0.89 0.89 0.89 0.89 0.87 0.86 0.93 0.89 0.92 0.89 0.89 0.87 1.00 14 0.90 0.90 0.90 0.91 0.90 0.89 0.91 0.91 0.94 0.88 0.90 0.85 0.91 1.00 15 0.85 0.85 0.87 0.88 0.87 0.81 0.87 0.89 0.92 0.88 0.88 0.87 0.91 0.91 1.00 16 0.87 0.90 0.91 0.89 0.87 0.87 0.92 0.87 0.94 0.89 0.88 0.86 0.91 0.94 0.92 1.00 17 0.85 0.90 0.85 0.80 0.79 0.86 0.89 0.82 0.90 0.91 0.82 0.82 0.89 0.85 0.86 0.86 1.00

Correlations across the 168 towns besides Cornwall.

269 Table C.5: Correlation Matrix of Topic Presence

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

1 1.00 2 0.71 1.00 3 0.56 0.54 1.00 4 0.64 0.49 0.48 1.00 5 0.70 0.61 0.52 0.65 1.00 6 0.37 0.42 0.38 0.42 0.32 1.00 7 0.72 0.60 0.64 0.61 0.65 0.40 1.00 8 0.60 0.51 0.50 0.57 0.59 0.44 0.56 1.00 9 0.47 0.50 0.47 0.50 0.44 0.33 0.42 0.60 1.00 10 0.47 0.52 0.50 0.39 0.33 0.35 0.51 0.39 0.37 1.00 11 0.69 0.57 0.62 0.74 0.62 0.36 0.68 0.63 0.55 0.47 1.00 12 0.39 0.32 0.59 0.47 0.41 0.30 0.51 0.36 0.37 0.50 0.47 1.00 13 0.75 0.71 0.69 0.66 0.70 0.40 0.78 0.63 0.58 0.55 0.72 0.48 1.00 14 0.59 0.48 0.57 0.62 0.60 0.47 0.56 0.68 0.47 0.43 0.66 0.43 0.65 1.00 15 0.46 0.37 0.54 0.54 0.46 0.36 0.54 0.55 0.55 0.47 0.54 0.49 0.54 0.53 1.00 16 0.64 0.52 0.69 0.54 0.56 0.35 0.72 0.60 0.58 0.48 0.69 0.48 0.65 0.51 0.63 1.00 17 0.57 0.66 0.56 0.50 0.51 0.54 0.62 0.52 0.48 0.54 0.58 0.38 0.68 0.47 0.40 0.63 1.00

Correlations across the 168 towns besides Cornwall.

270 C.3 Results with 50 Topics

Table C.6: Perplexity of the Gibbs Sampler Chains - 50 Topics

Chain 1 2 3 4 5 6 7 8 9 10 11

Mean 305.34 305.43 305.88 305.92 306.17 306.18 306.18 306.25 306.30 306.31 306.36 S.D. 0.49 0.54 0.53 0.50 0.53 0.50 0.49 0.50 0.66 0.53 0.50

Chain 1 is the chain used for estimation - it has lowest mean perplexity. Twenty-four other chains were run, and here I list the ten other that have means closest mean perplexities.

As a comparison with the main specification with 30 topics, I compare with a specification using 50. Fifty is the number used by Hansen, McMahon, and Prat (2014) in their analysis of topics in FOMC minutes, but since my sample is much smaller than theirs (14,900 sentences vs. 46,502 interjections), I use 30 topics for the main specifications and check the 50 topics specifications as a point of comparison. Like the main specifications, I estimate twenty-five chains and choose the one with the lowest perplexity. I report the perplexity of the chosen chain and the 10 closest chains in Table C.6. As expected, the perplexity using 50 chains is much lower, with the chosen chain in the main specification having a perplexity of 350.65 while the 50 topics chain has a perplexity of only 305.34. This means the 50 chain topics are a better fit, though as will become apparent there are several reason why the 30 chain version is preferred for the empirical exercise. Table C.7 reports the estimated 50 topics. There are numerous parallels between topics here and the topics of K = 30. For example, Topic 26 for K = 30 clearly represents landscaping requirements, with its top five tokens being “landscap,” “fenc,” “screen,” “surround,” and “tree”. When K = 50, Topic 33 is very similar with “fenc,” “landscap,” “screen,” “surround,” and “veget.” In contrast, increasing the number of topics essentially “splits” some of the previous topics. For example, there is another topic now that seems to be about

271 Table C.7: Estimated Topics - K = 50

Topic Five Most Frequent Tokens Share(%) Tokens Present 0 may, determin, inform, submit, addit 1.98 13.18 106 1 remov, abandon, upon, month, appropri 2.28 18.91 114 2 review, consult, independ, monitor, result 1.60 8.53 100 3 tree, part, plant, may, camouflag 1.75 11.57 95 4 fcc, environment, assess, file, submit 1.51 7.77 92 5 will, design, upon, avail, provid 1.74 12.18 90 6 height, feet, point, distanc, highest 2.36 14.13 117 7 rfr, measur, regul, meet, nois 1.55 6.15 58 8 provid, adequ, capac, demonstr, adequatecoverag 1.41 6.63 55 9 condit, bond, town, may, remov 1.66 8.40 94 10 minim, advers, potenti, visual, protect 3.10 24.61 121 11 util, land, construct, access, instal 1.83 11.95 109 12 addit, user, commun, design, accommod 2.14 15.88 114 13 residenti, zone, commerci, district, industri 2.48 17.38 116 14 owner, complianc, within, mainten, maintain 1.44 5.32 74 15 properti, subject, adjac, owner, name 1.74 9.21 107 16 elev, grade, foot, draw, mean 1.40 7.63 87 17 telecomtow, need, number, futur, accommod 2.41 16.42 98 18 regul, commun, purpos, safeti, gener 1.88 8.99 86 19 power, height, type, chang, frequenc 1.14 3.35 49 20 mount, roof, build, ground, project 2.27 14.20 109 21 prefer, altern, technolog, order, feasibl 2.07 14.82 111 22 feet, height, exceed, inch, diamet 2.94 18.96 125 23 regul, state, feder, csc, local 1.81 8.55 97 24 unless, light, faa, otherwis, paint 2.37 16.12 121 25 build, provid, follow, watertow, standard 2.21 11.98 106 26 standard, compli, specif, amend, gener 2.25 14.46 111 27 servic, person, regul, effect, provis 2.16 7.83 77 28 setback, proplin, greater, lot, zone 2.78 16.12 126 29 develop, histor, area, preserv, scenic 1.97 10.52 103 30 approv, specialpermit, construct, time, grant 1.92 6.97 78 31 servic, commun, limit, transmiss, gener 2.13 11.18 104 32 plan, show, map, coverag, scale 1.88 13.60 110 33 fenc, landscap, screen, surround, veget 2.29 14.68 122 34 date, time, day, prior, balloon 1.44 6.14 91 35 support, monopol, design, lattic, exampl 1.65 9.65 83 36 will, telecommun, report, indic, system 2.44 15.91 110 37 area, base, within, fallzon, wetland 1.65 8.80 74 38 engin, plan, licens, equip, document 2.34 12.65 105 39 within, public, radiu, mile, owner 1.74 11.83 107 40 co, can, feasibl, share, licensedcarri 1.75 10.74 102 41 build, equip, accessori, box, cabinet 2.50 19.40 119 42 siteplan, follow, regul, specialpermit, specialexcept 2.37 16.18 115 43 color, materi, detail, build, possibl 2.55 20.58 120 44 visibl, design, visualimpact, visual, simul 1.84 11.85 108 45 oper, provid, provis, insur, effect 1.49 8.02 80 46 receiv, transmit, rfr, equip, signal 2.26 14.64 84 47 area, contain, signal, analysi, squarefeet 1.61 9.59 88 48 town, provid, within, written, abut 1.74 9.58 89 49 permit, may, allow, except, sign 2.16 13.68 122 “Share” is the mean share over towns, “Tokens” is the median number of estimated tokens over towns, and “Present” is the number of successes of the presence proxy.

272 landscaping, as Topic 3’s top words are “tree,” “part,” “plant,” “may,” and “camouflag.” The Focus measures are now scaled down relative to the main specifications, since each share must now be smaller. The variance in share across topics seems to have increased, at least proportionally. While also smaller, the numbers of tokens remains relatively similar across topics. Finally, the presence proxies are similar to the 30 topics specifications as they are still mostly successes. Table C.8 reports the result of the Focus regressions with 50 topics. In comparison with the 30 topics regressions, there are many cases where similar topics are selected frequently with the same statistically significant signs: feasible alternatives (before Topic 0 in the 30 topics regressions and now Topic 21 in the 50 Topics regressions), signal strength (before Topic 23 and now Topic 19), higher authority (before Topic 11 and now Topic 27), landscaping (before Topic 26 and now Topic 33), time words (before Topic 6 and now Topic 34), and material design (before Topic 3 and Topic 43). However, because of the added number of variables causing multicollinearity, I had to ex ante remove the Section Present dummy as a possible covariate for all the Focus regressions. In addition, many other variables also had to be ex ante removed from the MetroPCS regression. Even so, the subsamples tend to retain more regressors than with 30 topics, implying that with the increased number of topics, overfitting my small samples is already an issue. This is especially an issue with the Rooftops and MetroPCS subsamples that tend to include the most variables and have implausibly high parameter estimates. Further, the value of the added regressors in even the full sample regression is suspect, as the same number of topic measures are chosen by the 30 topics regressions (15). There are more variables chosen overall, but these are the controls that were all also available in the 30 topics regressions. This implies that by splitting the measures into multiple versions of themselves in the new topic measures, the new measures have less explanatory power in

273 Table C.8: Results: Topic Focus: Share of Tokens - 50 Topics

(1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

Section Present ------log(1+Section Characters) -0.01 -0.27 -1.81 0.06 -0.24 -0.53 -0.90 (0.06) (0.24) (0.78)** (0.16) (0.14)* (0.24)** (0.55) log(1+Code Characters) -0.17 - (0.16) - Executive -1.15 -2.18 -10.28 -0.25 -3.42 -0.02 -4.03 -2.21 (0.51)** (0.67)*** (2.80)*** (0.57) (0.99)*** (0.35) (1.19)*** (2.31) Council 1.65 -4.76 1.10 1.83 - (0.55)*** (1.61)*** (0.61)* (0.73)** - Selectmen -1.05 -15.07 -0.45 -1.68 -2.74 - (0.52)** (3.39)*** (0.63) (1.03) (1.12)** - Town Meeting -1.35 0.70 1.47 - (1.62) (0.37)* (0.71)** - Average Turnout (Fraction) -3.43 -6.16 -20.36 -7.02 -3.47 -6.24 - (1.18)*** (2.15)*** (7.86)*** (2.50)*** (1.93)* (3.28)* - Democratic (Fraction) -2.54 -4.96 -25.35 -6.32 -0.24 -4.41 -39.50 (1.43)* (3.92) (10.08)** (3.03)** (1.45) (4.21) (53.35) log(Population) 0.31 0.46 1.90 (0.25) (0.38) (0.58)*** log(KM2) -0.39 -1.43 -1.38 -1.53 -0.67 -3.13 0.50 (0.18)** (0.40)*** (0.92) (0.47)*** (0.33)** (0.76)*** (2.29) Median HH Income -3.18 -9.00 38.88 -76.86 296.61 (5.94) (15.25) (40.48) (23.93)*** (319.61) Median Home Value 2.33 11.00 -99.40 (8.71) (5.35)** (100.77) Median Home Value > $1M -0.64 -1.27 -3.78 -0.60 10.16 -105.66 (0.55) (1.40) (7.12) (0.89) (4.33)** (81.83) Owned Housing Units (Fraction) -2.20 -5.53 -20.04 -2.26 -7.97 -1.90 -14.68 (1.05)** (2.68)** (11.59)* (1.38) (2.31)*** (3.99) (24.88) More than High School (Fraction) 26.99 -2.87 17.12 -28.07 (12.73)** (2.14) (4.84)*** (19.29) Median Age -0.24 0.06 0.15 (0.24) (0.04) (0.06)** Black (Fraction) 8.63 -1.19 24.10 (3.13)*** (4.61) (39.03) Hispanic (Fraction) -7.64 17.68 - (3.21)** (9.40)* - 0 may, determin, inform, 72.59 102.73 20.66 77.31 submit, addit (18.88)*** (96.34) (15.97) (27.21)*** 1 remov, abandon, upon, -14.57 -140.71 164.16 month, appropri (13.64) (59.94)** (97.38)* 2 review, consult, independ, -46.92 -176.14 -38.37 -381.27 monitor, result (20.35)** (63.74)*** (24.51) (141.34)*** 3 tree, part, plant, 292.75 18.28 -1.30 17.57 509.63 may, camouﬂag (164.23)* (11.45) (20.89) (18.24) (446.70) 4 fcc, environment, assess, 48.22 27.28 482.29 ﬁle, submit (30.17) (27.27) (246.25)* 5 will, design, upon, 33.29 22.68 -10.67 564.29 avail, provid (62.33) (11.89)* (19.28) (422.08) 6 height, feet, point, -70.32 -214.20 -5.93 2.64 16.30 -101.27 150.50 distanc, highest (17.58)*** (77.12)*** (8.89) (13.53) (10.77) (24.71)*** (214.71) 7 rfr, measur, regul, 14.18 51.04 -312.15 50.15 17.68 96.18 653.62 meet, nois (7.88)* (20.69)** (85.38)*** (22.86)** (13.43) (29.77)*** (406.13) 8 provid, adequ, capac, -26.48 101.53 -40.67 -91.04 455.37 demonstr,adequatecoverag (17.22) (82.17) (15.48)*** (22.61)*** (292.35) 9 condit, bond, town, 30.52 -78.74 26.52 may, remov (13.87)** (42.32)* (14.30)* 10 minim, advers, potenti, 9.61 73.23 11.57 8.97 56.78 195.89 visual, protect (9.25) (46.07) (11.99) (8.26) (15.92)*** (84.97)** 11 util, land, construct, -39.65 119.23 -30.32 48.04 -1008.16 access, instal (15.70)** (63.14)* (15.42)** (23.22)** (745.25) 12 addit, user, commun, 28.84 237.14 12.20 29.37 -99.62 design, accommod (13.13)** (62.26)*** (7.45) (16.21)* (59.01)* 13 residenti, zone, commerci, 51.78 38.82 34.56 68.57 116.18 district, industri (12.46)*** (42.92) (12.06)*** (17.59)*** (76.36) 14 owner, complianc, within, 19.34 302.64 17.54 11.52 36.35 1335.89 mainten, maintain (8.86)** (95.08)*** (11.61) (21.95) (14.79)** (919.20) 15 properti, subject, adjac, 457.78 -13.71 38.49 263.31 owner, name (125.69)*** (17.65) (20.10)* (154.33)* 16 elev, grade, foot, 269.27 -99.52 9.87 -2597.71 draw, mean (121.78)** (33.42)*** (23.78) (1722.81) 17 telecomtow, need, number, 46.83 18.87 -737.58 futur, accommod (45.06) (13.67) (420.56)* 18 regul, commun, purpos, 47.79 -111.95 50.35 65.06 -147.18 safeti, gener (18.17)*** (47.37)** (17.17)*** (24.27)*** (110.88) 19 power, height, type, -41.13 -80.49 -142.99 -15.63 -60.61 181.13 chang, frequenc (10.98)*** (27.71)*** (98.81) (12.93) (22.31)*** (231.16) 20 mount, roof, build, 27.41 599.77 ground, project (67.45) (373.88) Continued on next page

274 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

21 prefer, altern, technolog, -14.13 -81.91 -205.66 -20.43 -61.54 -24.37 -115.60 -1245.00 order, feasibl (8.60) (21.69)*** (113.66)* (9.19)** (17.65)*** (14.73)* (30.97)*** (946.83) 22 feet, height, exceed, -11.69 80.41 -12.78 -307.44 inch, diamet (7.07)* (26.13)*** (7.15)* (124.12)** 23 regul, state, feder, -14.69 -38.61 -302.57 csc, local (7.89)* (13.18)*** (70.99)*** 24 unless, light, faa, 44.63 102.70 4.33 64.80 332.09 otherwis, paint (14.49)*** (45.68)** (7.56) (20.83)*** (153.27)** 25 build, provid, follow, 23.60 45.88 18.50 5.83 32.36 569.17 watertow, standard (12.85)* (41.56) (7.26)** (8.40) (12.69)** (273.88)** 26 standard, compli, specif, -11.28 230.23 -73.82 511.45 amend, gener (12.57) (81.34)*** (27.20)*** (370.51) 27 servic, person, regul, 18.31 60.70 226.72 16.97 46.66 16.80 37.08 211.30 eﬀect, provis (4.33)*** (13.73)*** (59.42)*** (6.98)** (13.78)*** (8.08)** (12.09)*** (89.51)** 28 setback, proplin, greater, -88.14 -19.82 50.99 lot, zone (21.20)*** (9.00)** (36.64) 29 develop, histor, area, 24.51 84.60 -205.44 preserv, scenic (13.46)* (54.65) (202.71) 30 approv, specialpermit, 102.97 9.26 -301.04 construct, time, grant (31.88)*** (9.78) (177.65)* 31 servic, commun, limit, 12.29 15.27 33.46 -542.42 transmiss, gener (25.25) (9.23)* (10.31)*** (313.07)* 32 plan, show, map, -22.06 -476.84 -32.81 507.55 coverag, scale (9.20)** (116.26)*** (13.24)** (303.03)* 33 fenc, landscap, screen, -24.44 -80.54 -272.86 -36.76 -69.27 -22.74 -41.74 -533.49 surround, veget (6.79)*** (17.23)*** (66.21)*** (10.45)*** (17.44)*** (11.38)** (15.53)*** (349.01) 34 date, time, day, 13.89 52.62 75.63 -241.33 prior, balloon (7.74)* (15.57)*** (44.83)* (139.49)* 35 support, monopol, design, 17.17 74.72 -137.17 16.66 15.86 50.80 653.36 lattic, exampl (10.11)* (22.63)*** (90.22) (13.27) (20.92) (31.72) (507.95) 36 will, telecommun, report, -19.98 -232.82 -44.88 -24.37 -191.98 indic, system (16.95) (110.38)** (19.76)** (22.39) (92.80)** 37 area, base, within, 20.98 134.01 329.67 fallzon, wetland (14.54) (70.08)* (214.35) 38 engin, plan, licens, 3.27 13.54 161.41 10.04 11.76 55.37 161.66 equip, document (4.51) (9.72) (75.72)** (10.14) (9.16) (16.96)*** (121.78) 39 within, public, radiu, 34.57 -296.07 53.58 -713.95 mile, owner (18.67)* (87.87)*** (29.08)* (607.78) 40 co, can, feasibl, -31.35 184.65 -11.74 -209.04 share, licensedcarri (19.05)* (109.96)* (18.50) (162.14) 41 build, equip, accessori, -12.02 -20.42 -155.36 -17.02 -26.00 18.84 -583.04 box, cabinet (5.85)** (12.99) (83.42)* (7.96)** (12.27)** (13.01) (433.90) 42 siteplan, follow, regul, -34.85 59.33 11.67 -59.71 487.05 specialpermit, specialexcept (13.64)** (58.10) (9.44) (16.14)*** (347.06) 43 color, materi, detail, 21.21 23.36 389.28 25.08 38.98 10.73 360.37 build, possibl (5.18)*** (10.11)** (97.29)*** (7.51)*** (10.90)*** (8.32) (267.50) 44 visibl, design, visualimpact, -82.28 44.98 25.82 -17.44 1082.70 visual, simul (135.31) (19.00)** (14.51)* (21.86) (722.07) 45 oper, provid, provis, 15.11 40.07 336.69 23.50 -997.42 insur, eﬀect (10.88) (22.45)* (130.11)*** (14.56) (850.66) 46 receiv, transmit, rfr, 6.01 17.90 55.87 9.35 15.02 25.28 16.45 329.00 equip, signal (4.59) (8.82)** (29.35)* (6.04) (11.09) (8.35)*** (14.45) (197.47)* 47 area, contain, signal, 58.34 - analysi, squarefeet (33.88)* - 48 town, provid, within, 65.56 - written, abut (91.48) - 49 permit, may, allow, -21.16 -81.41 -21.61 -66.07 -68.06 - except, sign (19.31) (61.71) (11.97)* (24.16)*** (24.65)*** - Observations 553 267 286 261 226 215 226 123 Loglikelihood 375.04 148.66 143.18 172.72 120.65 138.77 123.01 44.56 Pseudo-R2 0.103 0.271 0.332 0.145 0.289 0.151 0.287 0.524

∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 Blank entries are variable not selected by LASSO. Entries with “−” were removed ex ante due to multicollinearity.

Table C.9: Results: Topic Presence: 97.5th Percentile - 50 Topics

(1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

Section Present 0.28 -0.78 26.47 10.62 -4.32 1.97 -206.75 (0.52) (3.49) (10.82)** (4.15)** (6.94) (5.58) (29.22)*** log(1+Section Characters) 0.03 -4.50 -1.81 -0.08 1.06 -0.23 26.09 (0.53) (2.00)** (0.64)*** (0.11) (1.10) (0.94) (4.07)*** log(1+Code Characters) 0.50 0.40 - (0.41) (0.30) - Executive -0.95 -0.30 -2.03 1.25 -2.35 -84.86 (0.52)* (1.20) (0.86)** (1.05) (0.94)** (9.29)*** Council 0.73 1.30 -0.81 0.77 - Continued on next page

275 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

(0.58) (0.60)** (0.74) (0.53) - Selectmen -0.12 -0.06 -2.31 0.50 - (0.20) (0.79) (0.86)*** (0.71) - Town Meeting -4.54 1.12 1.15 - (1.36)*** (0.52)** (0.51)** - Average Turnout (Fraction) -1.57 -3.71 12.00 -2.79 -3.65 -4.14 -12.05 - (1.21) (2.00)* (4.67)*** (2.56) (2.11)* (2.87) (2.73)*** - Democratic (Fraction) -3.47 -1.14 -0.00 -5.93 -16.08 -132.67 (1.37)** (1.75) (2.87) (3.71) (5.65)*** (30.80)*** log(Population) 0.97 1.43 51.84 (0.38)** (0.61)** (5.71)*** log(KM2) -0.26 -0.89 -2.50 -1.32 -1.37 -1.61 -2.77 -65.46 (0.19) (0.32)*** (0.71)*** (0.55)** (0.43)*** (0.63)** (0.77)*** (5.94)*** Median HH Income -11.92 -18.76 -44.02 -39.79 -26.62 -53.58 -24.59 -1061.39 (5.59)** (8.23)** (37.75) (17.50)** (23.88) (28.38)* (21.24) (139.27)*** Median Home Value 32.84 9.32 6.66 14.16 1.05 -142.63 (11.10)*** (5.05)* (5.54) (6.70)** (1.69) (20.65)*** Median Home Value > $1M 18.87 5.83 6.07 10.57 - (8.00)** (3.73) (4.38) (5.53)* - Owned Housing Units (Fraction) -1.64 -15.33 -4.45 7.35 -7.49 326.64 (1.13) (5.52)*** (2.69)* (4.09)* (4.24)* (30.07)*** More than High School (Fraction) -36.23 -2.73 -3.25 -11.56 5.99 554.68 (8.74)*** (4.14) (3.85) (5.27)** (6.82) (73.34)*** Median Age 0.84 0.15 0.08 -10.52 (0.17)*** (0.07)** (0.12) (1.27)*** Black (Fraction) -7.24 -4.31 6.43 140.91 (4.35)* (2.40)* (5.01) (21.20)*** Hispanic (Fraction) -4.93 6.42 (3.36) (5.55) 0 may, determin, inform, -1.16 4.49 -0.65 -2.13 -19.44 submit, addit (0.53)** (1.45)*** (0.64) (1.03)** (3.14)*** 1 remov, abandon, upon, -1.32 2.63 -0.87 2.29 3.61 1.30 -269.12 month, appropri (0.71)* (1.25)** (0.77) (0.75)*** (1.36)*** (1.28) (29.69)*** 2 review, consult, independ, 5.54 1.37 -0.78 1.25 -1.78 monitor, result (1.89)*** (0.56)** (0.47)* (0.78) (2.76) 3 tree, part, plant, 0.62 -3.91 0.90 1.99 2.20 -81.67 may, camouﬂag (0.25)** (1.49)*** (0.50)* (0.83)** (0.82)*** (9.28)*** 4 fcc, environment, assess, 1.02 -3.42 1.08 1.66 2.23 122.89 ﬁle, submit (0.44)** (1.73)** (0.50)** (0.51)*** (0.88)** (15.06)*** 5 will, design, upon, 0.74 1.90 -0.58 45.58 avail, provid (0.42)* (0.52)*** (0.76) (5.00)*** 6 height, feet, point, -1.30 -2.67 1.63 -1.89 -152.12 distanc, highest (0.66)** (0.89)*** (1.36) (1.27) (19.26)*** 7 rfr, measur, regul, 0.35 -0.38 -2.83 0.98 0.77 -55.60 meet, nois (0.21)* (0.36) (1.13)** (0.50)* (0.65) (7.65)*** 8 provid, adequ, capac, -1.95 -0.92 -0.71 -0.72 -78.37 demonstr, adequatecoverag (0.83)** (0.39)** (0.35)** (0.61) (10.47)*** 9 condit, bond, town, 0.43 0.85 1.36 1.32 73.88 may, remov (0.20)** (0.52) (0.61)** (0.91) (8.06)*** 10 minim, advers, potenti, 1.25 2.45 7.88 1.91 6.33 5.18 18.40 visual, protect (0.45)*** (0.72)*** (2.00)*** (0.82)** (1.63)*** (1.79)*** (6.25)*** 11 util, land, construct, -1.02 -0.99 -5.01 -1.26 -0.90 -5.00 -2.14 -22.67 access, instal (0.34)*** (0.48)** (1.82)*** (0.60)** (0.48)* (1.52)*** (1.13)* (5.83)*** 12 addit, user, commun, -1.74 10.43 -1.17 -0.86 -54.30 design, accommod (0.68)*** (2.29)*** (0.68)* (1.08) (5.91)*** 13 residenti, zone, commerci, 0.62 0.87 1.57 2.68 1.95 329.97 district, industri (0.40) (0.75) (1.89) (0.91)*** (1.29) (39.19)*** 14 owner, complianc, within, -0.46 -1.28 4.73 -1.93 -1.41 -115.70 mainten, maintain (0.26)* (0.45)*** (1.13)*** (0.62)*** (0.48)*** (12.62)*** 15 properti, subject, adjac, -0.46 -3.40 -2.03 -58.93 owner, name (0.23)** (1.35)** (0.89)** (8.07)*** 16 elev, grade, foot, -1.45 6.04 -0.94 -1.46 168.53 draw, mean (0.55)*** (1.62)*** (0.56)* (1.13) (18.31)*** 17 telecomtow, need, number, 0.41 1.32 -1.44 1.30 0.93 -1.43 0.89 185.86 futur, accommod (0.28) (0.45)*** (1.67) (0.57)** (0.47)** (1.02) (1.27) (24.57)*** 18 regul, commun, purpos, 0.93 -7.66 1.47 -0.83 1.81 155.69 safeti, gener (0.39)** (1.75)*** (0.56)*** (0.64) (1.00)* (19.62)*** 19 power, height, type, -0.37 -0.66 1.78 -0.88 -0.33 -0.43 -1.38 chang, frequenc (0.22)* (0.38)* (0.93)* (0.51)* (0.38) (0.63) (0.68)** 20 mount, roof, build, -0.66 1.87 0.79 -0.84 -1.18 23.75 ground, project (0.27)** (1.57) (0.45)* (1.32) (0.95) (7.83)*** 21 prefer, altern, technolog, 2.21 -4.45 1.50 2.90 order, feasibl (0.84)*** (1.63)*** (0.66)** (1.12)*** 22 feet, height, exceed, -1.40 -2.78 -1.11 20.60 inch, diamet (2.14) (1.13)** (1.11) (6.72)*** 23 regul, state, feder, 0.16 -1.53 1.76 0.64 1.86 1.18 -117.60 csc, local (0.42) (1.30) (0.57)*** (0.45) (0.85)** (0.91) (13.31)*** 24 unless, light, faa, -0.97 -1.72 -9.85 -1.14 -2.50 -1.83 -170.15 otherwis, paint (0.37)*** (0.66)*** (2.46)*** (0.73) (1.12)** (1.32) (17.75)*** 25 build, provid, follow, -0.80 90.18 watertow, standard (1.15) (9.51)*** 26 standard, compli, specif, -1.25 3.96 -2.73 -1.53 -0.95 -3.55 -175.61 Continued on next page

276 Continued from previous page (1) (2) (3) (4) (5) (6) (7) (8) All Towers Rooftops AT&T Verizon Sprint T-Mo. Metro.

amend, gener (0.56)** (3.12) (0.78)*** (0.79)* (0.84) (1.57)** (21.55)*** 27 servic, person, regul, 0.58 0.40 1.00 0.33 1.27 21.04 eﬀect, provis (0.36) (0.37) (0.41)** (0.68) (0.56)** (2.94)*** 28 setback, proplin, greater, -0.04 8.63 2.20 1.79 lot, zone (0.31) (2.70)*** (1.02)** (0.99)* 29 develop, histor, area, 0.99 -6.25 -0.99 -0.90 -4.04 preserv, scenic (0.53)* (1.63)*** (0.73) (0.72) (4.43) 30 approv, specialpermit, 4.37 -1.08 -1.11 -0.96 -52.23 construct, time, grant (1.12)*** (0.44)** (0.60)* (0.67) (6.58)*** 31 servic, commun, limit, 0.69 1.18 2.33 1.25 0.96 1.69 -40.54 transmiss, gener (0.25)*** (0.50)** (1.14)** (0.55)** (0.43)** (0.80)** (5.10)*** 32 plan, show, map, 1.20 -4.83 -1.95 -2.44 -118.72 coverag, scale (0.63)* (1.34)*** (1.04)* (1.06)** (15.52)*** 33 fenc, landscap, screen, -1.39 -2.21 -3.18 -2.42 -8.00 -3.81 -118.90 surround, veget (0.49)*** (0.65)*** (0.91)*** (0.77)*** (2.18)*** (1.33)*** (13.52)*** 34 date, time, day, -3.96 -2.51 -188.62 prior, balloon (1.54)*** (0.91)*** (22.01)*** 35 support, monopol, design, 6.68 -2.30 -1.47 -83.27 lattic, exampl (1.88)*** (1.17)* (0.67)** (8.68)*** 36 will, telecommun, report, 3.79 1.67 0.97 1.47 -0.44 26.97 indic, system (1.72)** (0.52)*** (0.44)** (0.85)* (0.84) (5.82)*** 37 area, base, within, 0.28 11.93 0.57 -16.10 fallzon, wetland (0.24) (3.15)*** (0.61) (2.55)*** 38 engin, plan, licens, -2.25 121.17 equip, document (0.90)** (15.31)*** 39 within, public, radiu, -4.38 0.96 -0.86 -1.22 -1.06 114.06 mile, owner (1.46)*** (0.64) (0.55) (0.79) (0.85) (17.10)*** 40 co, can, feasibl, -5.92 0.52 -33.10 share, licensedcarri (1.56)*** (0.54) (3.21)*** 41 build, equip, accessori, 0.35 0.58 2.88 2.85 - box, cabinet (0.35) (0.55) (1.12)** (0.99)*** - 42 siteplan, follow, regul, 7.94 1.48 -1.48 208.46 specialpermit, specialexcept (2.54)*** (0.53)*** (0.93) (24.62)*** 43 color, materi, detail, 0.02 2.93 1.71 3.51 2.80 0.82 2.43 158.68 build, possibl (0.35) (0.69)*** (2.23) (1.04)*** (0.72)*** (0.97) (1.33)* (18.59)*** 44 visibl, design, visualimpact, 0.41 0.80 4.71 3.62 101.98 visual, simul (0.32) (0.68) (1.56)*** (1.39)*** (9.93)*** 45 oper, provid, provis, 3.87 0.71 0.59 -1.23 - insur, eﬀect (1.55)** (0.44) (0.78) (0.65)* - 46 receiv, transmit, rfr, -4.83 0.58 1.35 1.33 116.39 equip, signal (1.64)*** (0.41) (0.71)* (0.76)* (11.46)*** 47 area, contain, signal, -0.25 -0.29 -1.72 1.10 1.73 -173.10 analysi, squarefeet (0.31) (0.48) (1.12) (1.03) (0.86)** (17.32)*** 48 town, provid, within, -0.29 -0.96 -1.09 -1.93 -0.83 - written, abut (0.22) (0.44)** (0.39)*** (0.79)** (0.63) - 49 permit, may, allow, -1.14 -1.71 -2.08 -3.64 - except, sign (0.62)* (1.79) (0.66)*** (1.46)** - Observations 553 267 286 261 226 215 226 123 Loglikelihood 371.08 154.66 140.57 152.57 126.07 119.85 123.40 25.68 Pseudo-R2 0.112 0.241 0.344 0.245 0.258 0.266 0.285 0.725

∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1 Blank entries are variable not selected by LASSO. Entries with “−” were removed ex ante due to multicollinearity issues. the penalized regressions.4

Table C.9 reports the result of the Presence regressions with 50 topics. In stark contrast with the Focus regressions, the new regressions generally select variables that are quite diﬀer- ent from the regressions with 30 topics. Similar topics that are often selected by subsamples in both include only radiation standards (before Topic 17 and now Topic 7), removal bonds

4That is, perhaps adding more topics is akin to replacing a variable with two variables that add up to the old variable. In the true model, the correct coeﬃcient on each of the two variables should be the same. But to LASSO, this means adding another variable. As a result, LASSO may prefer to add in another variable instead of adding both of the split variables.

277 (before Topic 13 and now Topic 9), and visual impact (before Topic 11 and now Topic 10). In the latter two cases, the coefficients on the topics actually have opposite signs to the previous results. Moreover, several presumably burdensome regulations have statistically significant positive impacts, including the aforementioned topics 9 and 10, and also Topic 3. Topic 3 is especially interesting as it appears to measure landscaping regulation and so is therefore redundant with Topic 33. Topic 33’s presence has a statistically significant negative coefficient when selected. In the original regressions, there is only one landscaping topic, whose Presence measure does not have a statistically significant effect. Thus the “splitting” of topics may be generating spurious statistically significant impacts through the inclusion of multiple proxies for the same idea. Overall then, the results for the presence measures with 50 topics are generally harder to interpret than the results for the Focus regressions. A partial reason for this may be how I construct the presence measures. Presence is measured as the topic having any sentence in a town with a share of that topic greater than the 97.5th percentile of shares of all topics in all sentences. Thus as the number of topics increases, the percentile scales, so that about the same number of successes of the indicator is achieved for each topic. This is why the number of successes of Presence measures are similar on average in both 30 topics and 50 topics regressions. However, as the topics now represent “less” words, the successes should decrease as there should be less Presence. In contrast, the current measure is now effectively more permissive as the number of topics increases. The Presence measures with 30 topics seem to proxy for regulation well enough, though it may be that the 97.5th percentile is simply the right cutoff for 30 topics and the optimal cutoff is actually sensitive to the number of topics and sentences. As determination of what an optimal cutoff might be is beyond the scope of this study, I leave it for future work to consider. As a final note, the multicollinearity with 50 topics actually seems to be less problematic

278 in the Presence regressions than with the Focus regressions. I only need to remove variables ex ante from the MetroPCS regressions, though the estimated parameters, like the Focus measures, are implausibly high. All the regressions retain more variables than with 30 topics, Rooftops and MetroPCS especially, but the other subsamples do not retain as much as the Focus regressions. In summary, considering that with 50 topics the Focus results are not more informative and the Presence results seem less interpretable, I use 30 as my preferred number of topics.

279