Three Essays on Colombia's Telecommunications

Item Type text; Electronic Dissertation

Authors Velez-Velasquez, Juan S.

Publisher The University of Arizona.

Rights Copyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction or presentation (such as public display or performance) of protected items is prohibited except with permission of the author.

Download date 29/09/2021 08:38:41

Link to Item http://hdl.handle.net/10150/624582 THREE ESSAYS ON COLOMBIA’S TELECOMMUNICATIONS

Juan Sebastián Vélez-Velásquez

———————————————

A Dissertation Submitted to the Faculty of the

DEPARTMENT OF ECONOMICS

In Partial Fulﬁllment of the Requirements

For the Degree of

DOCTOR OF PHILOSOPHY

In the Graduate College

THE UNIVERSITY OF ARIZONA

2017 THE UNIVERSITY OF ARIZONA

GRADUATE COLLEGE

As members of the Dissertation Committee, we certify that we have read the dissertation prepared by Juan Sebastián Vélez-Velásquez, titled Three Essays on Colombia’s Telecommunications and recommend that it be accepted as fulﬁlling the dissertation requirement for the Degree of Doctor of Philosophy.

______Date: (4/20/2017) Gautam Gowrisankaran

______Date: (4/20/2017) Mauricio Varela

______Date: (4/20/2017) Mo Xiao

Final approval and acceptance of this dissertation is contingent upon the can- didate’s submission of the ﬁnal copies of the dissertation to the Graduate College. I hereby certify that I have read this dissertation prepared under my direction and recommend that it be accepted as fulﬁlling the dissertation requirement.

______Date: (4/20/2017) Dissertation Director: Mo Xiao

2 STATEMENT BY AUTHOR

This dissertation has been submitted in partial fulﬁllment of the requirements for an advanced degree at the University of Arizona and is deposited in the University Library to be made available to borrowers under rules of the Library. Brief quotations from this dissertation are allowable without special permission, provided that an accurate acknowledgement of the source is made. Requests for permission for extended quotation from or reproduction of this manuscript in whole or in part may be granted by the head of the major department or the Dean of the Graduate College when in his or her judgment the proposed use of the material is in the interests of scholarship. In all other instances, however, permission must be obtained from the author.

SIGNED: Juan Sebastián Vélez-Velásquez

3 Contents

List of Figures 8

List of Tables 9

1. Download Speed and Price Effects of the Comcel-Telmex Merger 13 1.1. Introduction ...... 13 1.2. Background ...... 15 1.2.1. The Telecommunications Sector ...... 15 1.2.2. Regulation ...... 17 1.2.3. The merger ...... 18 1.2.4. Events that (almost) overlapped with the merger . . . . 19 1.3. Data ...... 21 1.3.1. Characteristics of Internet plans ...... 22 1.3.2. Characteristics of the providers ...... 24 1.3.3. Download speeds over time ...... 24 1.3.4. Price over time ...... 25 1.4. Empirical Strategy ...... 28 1.4.1. The nature of the diff-in-diff ...... 29 1.4.2. Different specifications ...... 31 1.5. Results ...... 33 1.5.1. Effects on markets ...... 33 1.5.2. Effects on rivals ...... 37 1.5.3. Effects on merged ...... 39 1.6. Comments ...... 40

2. Merger Eﬀects with Product Complementarity: Evidence from Colombia’s Telecommunications 43

4 2.1. Introduction ...... 43 2.2. Mergers: horizontal, vertical and with complements ...... 49 2.2.1. Types of mergers ...... 49 2.2.2. What we know about these mergers ...... 49 2.3. Relation to the Literature ...... 51 2.4. Colombia´s telecommunications sector ...... 52 2.5. Data ...... 55 2.5.1. The CRC’s Form 5 data ...... 55 2.5.2. Cities, strata and markets ...... 60 2.5.3. The GEIH data ...... 61 2.6. Empirical model ...... 62 2.6.1. Empirical demand ...... 63 2.6.2. Utility ...... 63 2.6.3. Instruments ...... 66 2.6.4. Identiﬁcation of complementarity ...... 67 2.7. Supply ...... 68 2.8. Counterfactuals ...... 69 2.9. Results ...... 70 2.9.1. Bundle characteristics ...... 70 2.9.2. Substitution patterns ...... 71 2.10. Counterfactual scenarios ...... 74 2.10.1. A merger with complements ...... 75 2.10.2. Splitting up a consolidated merger ...... 78 2.11. Comments ...... 80

3. The Costs of Banning Price Discrimination under Imperfect Com- petition: Evidence from Colombia’s Telecoms 83

5 3.1. Introduction ...... 83 3.2. Literature ...... 86 3.3. Best response asymmetry in Colombia’s telecoms ...... 88 3.4. Strata: what are they and how ﬁrms use them ...... 89 3.5. Data ...... 90 3.5.1. Form 5 ...... 91 3.5.2. Summary statistics ...... 92 3.5.3. Household Survey ...... 96 3.6. Empirical demand ...... 97 3.6.1. Utility ...... 98 3.6.2. Instruments ...... 100 3.6.3. Demand Estimates ...... 102 3.6.4. Marginal costs ...... 105 3.7. Uniform pricing ...... 106 3.7.1. Results ...... 107 3.7.2. Consumers ...... 107 3.7.3. Firms ...... 110 3.8. Comments ...... 111

4. Conclusions and Future Directions 113 4.1. Conclusions revisited ...... 113 4.2. Future research directions ...... 114

A. Appendix 116 A.1. Chapter 1 ...... 116 A.2. Chapter 3 ...... 120 A.2.1. Demand estimates ...... 120 A.2.2. Counterfactual prices and consumer surplus ...... 121

6 References 123

7 List of Figures

1. Average download speeds over time ...... 25 2. Average download speeds over time(weighted by subscribers) . . 26 3. Average unit prices over time ...... 27 4. Average unit prices over time ...... 27 5. Correlations between standalone goods ...... 60 6. Distribution of prices after merging ETB and Avantel ...... 76 7. Distribution of prices after breaking up Claro ...... 79 8. Demand estimates ...... 104 9. Compensating Variation ...... 109 10. Distribution of prices under diﬀerent scenarios ...... 116 11. Price histogram ...... 122

8 List of Tables

1. Descriptive statistics for wired Internet plans (Means and Stan- dard Deviations) ...... 23 2. Descriptive statistics for Internet providers (Mean and Standard Deviations) ...... 24 3. Average download speeds before and after ...... 26 4. Median download speeds before and after ...... 26 5. Median download speed in the market ...... 34 6. Minimum download speed in the market ...... 35 7. Maximum download speed in the market ...... 35 8. Unit price of plan offering the median speed ...... 36 9. Median speed (Pooled) ...... 38 10. Median speed ...... 39 11. Median speed for plans offered ...... 40 12. Bundle characteristics ...... 57 13. Price and market share of bundles ...... 58 14. Summary statistics for firms ...... 59 15. Household characteristics by stratum ...... 62 16. Bundle characteristics ...... 72 17. Estimated substitution patterns for selected goods ...... 73 18. Price distributions for standalone goods ...... 78 19. Price distributions for bundled goods ...... 78 20. Price distributions for standalone goods ...... 80 21. Price distributions for bundled goods ...... 80 22. An example of Price Discrimination ...... 86 23. Price of bundles ...... 92

9 24. Price of bundles by stratum in 2015 dollars ...... 93 25. Internet and Phone characteristics by stratum (Means and S.D.) 94 26. Bundle characteristics by stratum (Weighted Means and S.D.) . 95 27. Characteristics TV (Percent of plans with) ...... 96 28. Household characteristics by stratum ...... 97 29. Demand estimates: Standard deviations ...... 105 30. Monthly compensating variation ...... 108 31. Bundle characteristics by stratum (Weighted Means) ...... 110 32. Estimated substitution patterns for goods included in a given bundle ...... 117 33. Substitution patterns (Bundles of 2) ...... 117 34. Substitution patterns (Bundles of 3) ...... 118 35. Substitution patterns (Bundles of 4 and 5) ...... 118 36. Correlations between standalone goods ...... 118 37. Compensating valuations after merging ETB and Avantel . . . . 119 38. Compensating valuations after splitting up Claro ...... 119 40. Compensating variations ...... 120 39. Demand estimates (Means) ...... 120 41. Counterfactual prices ...... 121

10 Abstract

Colombia’s telecommunication industry has changed drastically in the last decade. Among the most salient events, a series of mergers between some of the industry’s largest providers which resulted in a reduced number of competitors. One would expect that this reduction in the number of competitors would translate into higher prices. However, competition is at such high level that the media even talk about a "price war". My dissertation aims to shed light on the causes of this apparent inconsistency between a smaller number of competitors and more competitive outcomes. I start by showing that, in effect, the latest of these mergers, the one between Comcel and Telmex, had a pro-competitive effects on the provision of broadband. Next I show that the services provided by Comcel and Telmex were complements and that the pro-competitive effects of the merger can be explained by this complementarity. Finally, I study the effects of price discrimination under oligopolistic competition. In chapter 1 I assess the ex-post short-run effect on broadband provision of the Comcel-Telmex merger. Employing administrative data about the universe of plans and firms providing wired telecom services, I use several difference- in-difference specifications to obtain estimates for the effect of the merger on price and download speeds of plans provided by the merging firm and its rivals. My estimates suggest that, in markets affected by the merger, download speeds rose by .6 Megabits per second on average. The average increase in the markets resulted from increases in the plans offered by the merging firm (1.2 Mbps) and increases in the speeds of plans provided by its rivals (.5 Mbps). In chapter 2 I study mergers of firms producing complementary goods. Merg- ers of firms producing complementary products have ambiguous effects on consumer welfare. The merged firm may lower prices because the merger in-

11 ternalizes the profits originated by the complementarity. But with the merger the firm gains the ability to bundle and with bundles the firm can exert price discrimination, increasing the prices of standalone products. I employ a comprehensive, administrative data set, which records prices, market shares, and plan attributes of the universe of Colombia’s telecom carriers, to assess which effect dominates. I estimate a random-coefficient discrete choice model of consumer demand model for bundled and standalone telecom products, in which the degree of substitutability or complementarity among products is an essen- tial parameter of interest. I find that major telecom products display a mix of substitutability and complementarity, but in general hardwired and mobile services are perceived as complements by Colombian households. My counterfactual experiments using the estimated model, indicate positive net effects of mergers with complements: despite a small increase in the price of standalone goods, consumer surplus increases by around 7 million dollars per quarter. Finally, in chapter 3, I study price discrimination in an oligopolistic setting. Economic theory is not conclusive about the effects of banning third degree price discrimination under imperfect competition. Price discrimination can enhance competition if the firms practicing don’t agree on the ranking of their markets. In this case, price discrimination can lead to lower prices in all markets. Thus, forcing the firms to charge uniform prices can increase their profits and reduce consumers’ surplus. Using data on prices, market shares and characteristics of telecommunication services sold under price discrimination by Colombian telecom providers, I estimate a model of competition. The estimates allow me to simulate a counterfactual scenario in which firms lose their ability to exert price discrimination within a city. Simulating a ban on price discrimination has negligible effect on consumer surplus and increases profits slightly.

12 1. Download Speed and Price Eﬀects of the

Comcel-Telmex Merger

1.1. Introduction

For the telecommunications industry of Colombia, 2012 was a tumultuous year. However, the Claro merger was, by far, the most important event. At the end of the second quarter, Colombia’s largest ISP (Telmex) and Mobile carrier (Comcel) merged under the Claro brand, despite announcing early in the year they had no plans to do so. At the time of the merger, Telmex was providing wired services in 77 cities in 24 departamentos.1 In each of those cities, Telmex played a prominent role with market shares averaging over 39%. Comcel, in part due to its wireless nature, provided mobile services in over 90% of the cities and provided services to over 29 million subscribers (65% of the market). The added revenue of the merging firms was the second largest in the country (around $2.7 billion). Despite the size of the firms involved in the merger, regulators said little and did less.2 Among the few things regulators mentioned was that evaluating the effects of the merger ex-ante would be a strenuous and difficult task. And they were right. The merger was not the usual horizontal merger, in which the merging firms produce substitute goods. Comcel and Telmex did not sell components of a final good, either. Each merging firm sold subscriptions to several services, and many households had subscriptions to both companies.3

1Departamentos are the administrative analog of states in the United States. There are over 1000 cities in Colombia., however, the vast majority of those are very small. The 77 cities where Telmex was present, account for over 70% of the population of the country. 2Carlos Rebellón, director of the CRC at the time said: ’should a scrutiny be required, it would be done after detecting an eﬀect in the markets, then we must act’. (http://archive.is/sLvnC) 3Telmex provided Cable TV, land-line phone and Internet while Comcel provided cell phone and Mobile Internet.

13 In this kind of merger, defining the market is not straightforward, and the direction of the price effects is harder to grasp than when reviewing horizontal mergers. As shown in Choi (2008), because the goods are complements, the merged firm has incentives to lower the prices, relative to the prices before the merger. However, the merged firm has incentives to use bundling for price discrimination, charging a low price for the bundle and a high price for standalone goods. In this paper, I study how the merger affected the provision of broadband Internet. I look at its effect on prices and download speeds for broadband plans provided by Telmex -the Internet Service Provider (ISP) involved in the merger- and its rivals. I use a difference-in-difference approach to assess these effects. I capitalize on the fact that not all markets were affected by the merger in the same way. Although Comcel, the mobile merging company, operated in every city, Telmex, the ISP, operated only in a couple dozen cities. From the standpoint of the broadband providers, the merger only affected broadband markets where Telmex operated. The difference-in-difference approach compares download speeds (and prices) in markets affected by the merger (i.e., where Telmex operated) to similar markets unaffected by it. Similar to the findings in Howell (2012), I find no evidence of anti-competitive effects. Moreover, the data seem to suggest the merger produced pro-competitive effects. First, there is no evidence the prices of broadband plans sold by the merged company or by its rivals increased more in markets where the merger occurred. Second, the average download speeds increased in markets affected by the merger. Speeds increased everywhere around the time of the merger, among other things, due to policies aiming at improving adoption. I find evidence the increases were higher in cities where Telmex was present. The resulting increases in speeds in those markets appear to tell a story about product

14 differentiation. The merged entity and it largest rivals increased the top tier speeds on their menus. Small providers competing with the merged increased the speeds of their low tier plans. My estimates suggest that, in cities where Telmex operated during the merger, the average download speeds increased by .6 Megabits per second (Mbps) more than in other cities during the 2 quarters following the merger. In particular, the merged firm increased speeds in those cities by 1.2 Mbps, while its rivals increased it by approximately .5 Mbps. Beyond their magnitude, the increase in speeds was remarkable in at least two ways. First, it happened abruptly. Download speeds had been growing at about 0.3 Mbps for the last 5 years prior to the merger (). Only one quarter after the merger, the average speed in markets affected by it doubled that rate of growth. Second, the increase was widespread. Speeds jumped from one quarter to the next in every city affected by the merger. The abrupt nature of the increase in speeds dismisses technological explanations -building new networks, accessing new equipment, or the mobile firm’s networks- as a reason for the higher speeds. The merged entity gained access to a new submarine cable (AMX1), which could have explained the increase in download speeds, but that did not materialize until several quarters after the merger.

1.2. Background

In this section, I describe the institutional setting, the main players in the telecommunications sector, and some policies contemporaneous with the merger. I also provide a brief timeline of all the events leading to the merger.

1.2.1. The Telecommunications Sector

In 1947, Colombia nationalized Marconi Wireless Telegraph and The Radio (Telegrafía) Nacional under the name Telecom (Empresa Nacional de Teleco-

15 municaciones). Since its foundation, Telecom held a monopoly on national and international services, while several regional monopolies provided wired phone services locally. After the 1991 constitution, which emphasized economic deregulation, the sector became more open, even more so, after 1994, with the creation of the Comisión Reguladora de Comunicaciones (CRC), in a bid to foster competition and quality in the sector. The same year, the first 3 cell-phone companies were established. Today, 8 mobile carriers operate nationwide, 4 of which do so as Mobile Virtual Network Operators (MVNO). There are well over 100 ISP and cable providers. Despite the large number of firms providing wired and wireless services, the country has deeply concentrated mobile and hardwired markets. Concerning its size, the sector generated 14 billion dollars in revenue in 2012. These revenues would rank higher than OECD countries, like Belgium or Portugal. The country’s telecom infrastructure is well-developed in the main urban centers. In cities, service availability ranks high when compared to the rest of Latin America. In small urban areas, however, telecom facilities remain inadequate. The government has tried to close the rural-urban gap in broadband adoption by carrying out a series of policies. These policies were designed with two goals in mind, first, to facilitate the entry of operators to the mobile market to improve competition. The second is extending availability of broadband nationwide.4 The national market structure resembles the structure of most countries: a few large firms providing services nationally and numerous small firms with presence in local markets. Today, 4 firms lead the hardwired markets: Claro (the merged company this paper addresses), UNE-EPM, Movistar, and ETB.

4The government’s main initiatives have been: Computers for Schooling (which provides computers to schools), Compartel (which aims at providing every Colombian with access to telephone and Internet services), Vive Digital (subsidies for Internet access for schools, low-income households, and rural areas), and the National Fiber Optic Project.

16 Claro provides cable modem access, UNE offers both DSL and cable modem services, whereas Movistar and ETB use DSL technology. The cable TV sector has undergone major consolidation, and Claro and UNE-EPM, together, control 89% of the national subscriptions to cable TV. The mobile sector grew rapidly during the first decade of the century, but has slowed in recent years. However, there is still room for growth, in particular for mobile Internet, where penetration still lags. In rural areas, there is great potential for adoption in mobile Internet because of the inadequate hardwired infrastructure. Mobile markets are dominated by three carriers: Claro, Movis- tar, and Tigo. Long before the merger, when the firm was still named Comcel, Claro held a leading position, but its market share has dwindled steadily since 2009. Avantel, a company that specialized as a trunking service provider for a few corporate customers via iDEN technology, has recently gained terrain. After winning a spectrum auction in June 2013, Avantel’s LTE networks have grown steadily in the main urban centers, making it another important player. A large competitive push has come from Mobile Virtual Network Operators. Virgin Mobile alone, for instance, has almost 2.6 million subscribers in the last quarter on the dataset.5 As I am writing this paper, DirecTV a long-time satellite TV provider, is adding LTE Internet to its menu.

1.2.2. Regulation

Colombian authorities realize the need for liberalization and eﬀective competition. Regulators have made enormous progress, regarding the era of state monopolies that characterized the country prior to 1991. The existence of an

5Other MVNOs are: Uff! Móvil and Móvil Éxito who offer mobile voice and Internet services over Tigo’s network. Metrotel and Virgin Mobile have wholesale agreements with Movistar. Since Tigo merged with UNE-EPM it has been able to offer converged services across fixed and mobile networks.

17 additional consumer authority aids the CRC in competition issues, allowing for a second scrutiny to ensure operators abide by the principles of competition. The recent adoption of “unique licensing”, which reduced greatly the entry barriers, produced incentives for market entry. Regulators in Colom- bia also understand that, when scrutinizing mergers, they should not examine competitiveness in markets in isolation. When reviewing deals, they usually consider that, as Prieger (2015) points out, Multi-Market Contact (MMC) not only facilitates collusion but plays a role in entry decisions. Merger analysis usually accounts for the networks created by MMC. Despite the solidly designed institutions, the policy and regulatory framework for Colombia’s telecom has problems. The lack of independence of the CRC and the high turnover of its commissioners and directors is worrisome. The Information and Communication Technology Minister is the Chairman of the regulatory agency, which cannot meet without his presence. The government owns 30% of the second largest operator in the country. In addition, the Superintendent for Trade and Commerce, the other consumer and competition authority, is appointed by the president. While regulators have paid a great deal of attention to mobile markets, they have allowed hardwired markets to operate virtually unregulated. In Colombia, like in the rest of the world, services converge, and providers offer voice, data, and video altogether. The regulators appear not to realize that and still regulate hardwired and mobile services as if they are independent services. In particular, they should consider the firms’ ability to bundle when defining markets.

1.2.3. The merger

During the third quarter of 2012, Comcel and Telmex merged under the name Claro. Comcel was launched in 1992 as a mixed capital company, with the

18 Colombian state owning almost 80%. The company, however, did not provide mobile services until July 1994. By then, the State was not the main share- holder anymore, and 50% of Comcel was owned by Bell Canada. At the end of their ﬁrst year of operations, Comcel provided services to 75,000 subscribers. In 1999, América Móvil bought Bell Canada, acquiring its shares on Com- cel. However, it was not until 2003, after buying Telecom’s remaining 4.3% of shares, the conglomerate seized 100% of the ownership. By June 2013, just before the merger, Comcel had over 30 million subscribers. Telmex arrived in Colombia as a provider of telecommunications for businesses at the beginning of the 2000s. It only started providing services for homes after acquiring TV Cable Bogotá and Cable Pacíﬁco. Further acquisition of local telephone companies aided in consolidating Telmex’s ranking as one of the leaders in the wired services market. By 2010, Telmex had 21% of Colombia’s cable TV subscriptions, second only to UNE, with 22%. By mid-2012, Telmex was the third provider of landline phone, nationally. Its market share in the cities where it provided broadband averaged 40%.

1.2.4. Events that (almost) overlapped with the merger

Several events aﬀected the broadband markets around the time the merger took place. These events may seem, at ﬁrst, as confounding factors able to explain the observed increase in speed. In this section, I describe those policies in some detail. First, Colombia’s Government championed several initiatives to increase adoption of broadband by poorer households. The 2012 tax reform, for instance, added tablets to the list of devices exempted from value-added tax to facilitate access to the web.6 In addition, the reform made households in strata 1, 2, and 3 exempt from paying value-added tax on their subscriptions

6Laptops and desktops were already exempt.

19 to wired Internet. With over 80% of Colombian households in the exempt strata, this measure was an effective tool to boost broadband adoption. Both measures were in effect starting January 2013, coinciding with the period after the merger. In 2012, the National Fiber Optic Project launched. The project, with a budget of $600 million dollars, aimed to deploy over 15,000 kilometers of optic fiber by 2014, with the goal of connecting 1,078 of Colombia’s 1,123 municipalities. Although the project started deploying fiber before the merger took place, most of its progress spanned well into 2013 and 2014, overlapping the period after the merger. A highly successful 4G auction in June 2013 produced five winning bidders: Claro, Movistar, Tigo, Avantel, and DirecTV; the latter being a new entrant into the mobile market. These license winners had to launch LTE services before mid-2014, and Movistar started to do so at the end 2013. Another important event that affected the mobile markets was the launch of two MVNO: Virgin Mobile in April and Éxito Movil one month later. Given that both policies aimed at increasing the demand for broadband, firms could have responded to the increased demand with higher speeds. For instance, if economies of scales were associated with the provision of broadband, it may be optimal for firms to increase speeds. Second, a couple quarters after the Telmex-Comcel merger, Telefónica seized full control of Colombia Telecommunications, and firms may have reacted to that move by increasing speeds. However, these events are incapable of compromising the effect of the merger captured in this paper. The broadband side of the merger only affected those markets where Telmex operated. The geographic nature of the merger generated a discontinuity that affected some markets, while leaving others unaffected. I compare the speeds in cities affected by the merger to cities where the merging ISP did not operate at the time. The difference-in-difference nature of the exercise guarantees the events mentioned above are differenced out,

20 removing any doubt about the cause of the increase in speeds. Such approach gets rid of anything that affects both the rivals of the merging firm and the controls. The same comparison was made dropping plans in strata 1 and 2, and the estimates remained unchanged. Similarly, difference-in-differences for rivals of the merged ISP and similar firms that faced competitors other than Telmex. With this exercise, I aim to measure the effect of the merger on rivals. Finally, I compare the increase in speeds offered by the merged company to those of similar rivals to obtain a lower bound on the increase caused by the merger.

1.3. Data

The CRC provided the data employed in this study. The CRC collects the data quarterly from information reported by the operators through Forms 30, 31, 34, and 35. Firms that provide landline phone use Forms 30 and 35 to report local and long distance plans, respectively. In Form 31, ISPs report the characteristics of the broadband offered. Form 31 requires firms to specify for every plan they offer. For every plan, firms also report the plan’s download and upload speeds, the connection used for delivery, its price, and the city and strata where the plan is available. Similarly, cable TV operators report on Form 34 the main characteristics of the plans, like having HD or premium channels. For every plan in every Form, the firms report the number of subscribers. The data spans 18 quarters, starting in the first quarter of 2010, and contains information for all the providers and all the cities. The providers had to report the information within 30 days of the end of each quarter. Besides reporting active plans offered during the quarter, firms also had to report plans that became inactive. The dataset contains information on residential and business plans. I focus on residential services. The dataset

21 only contains information about plans that companies offer massively. If a firm caters to a specific customer and designs a plan for them, that plan does not show in the dataset. With the data, the CRC provided all the documentation to read the dataset. The data, the CRC warned, was provided “as is”. The CRC performed no quality control on the data, so they gave me the information as reported by the providers. For instance, some plans reported as offered to residential users have a unique subscriber, are delivered via clear-channel, and the prices are several hundred times higher than the average residential plan. These characteristics are typical of internet services offered to firms. I drop such plans, because they seem to have been misreported as residential when they were business plans. There is some missing information about download speeds and some characteristics of TV plans. For the most of those, it was possible to input values because the codes that identified the plan contain such information.

1.3.1. Characteristics of Internet plans

The main attributes of wired broadband plans and how they evolved over time appear in table 1. The average price of Internet plans increased over the period due to usual inflationary pressure and because firms offered increasingly better plans. For instance, the slowest plan observed in the first quarter of the data is a 0.128 Mbps offered by Edatel in Ebéjico, Antioquia. The minimum speed offered by the same company in that municipality is 0.6 Mbps in the last quarter on the data. A little under 200 people subscribe to the average plan. But the number of subscribers varies enormously, as the large standard errors made it evident. For instance, the number of subscription ranges from 9 (a DSL plan provided by Coltel in Garagoa, Boyacá) to over 300.000 (a DSL plan offered by ETB in

22 Table 1: Descriptive statistics for wired Internet plans (Means and Standard Deviations) 2011Q3 2011Q4 2012Q1 2013Q1 2013Q2 2013Q3 Average price 19.78 20.12 19.84 20.13 20.22 20.47 (4.97) (4.39) (4.65) (4.85) (4.84) (4.92) Average subscribers 146.53 186.02 195.74 207.40 183.14 189.69 (167.80)(188.21)(196.18)(199.35)(172.18)(169.72) Download speed 2.20 2.41 2.43 3.12 3.22 3.29 (0.91) (0.88) (0.88) (1.10) (1.12) (1.01) Upload speed 0.72 0.77 0.78 0.97 1.15 1.07 (0.19) (0.21) (0.21) (0.30) (0.47) (0.38) Price in 2015 dollars. Download speed and Upload speed in Megabits per second.

Bogotá). In terms of subscription, no clear time trend arises, because although overall subscription increased over the period, the variety of plans offered also increased, resulting in fewer subscriptions per plan. The average download speed exhibits a clear upward trend over time, not caused by the merger. Average download speed increased by about 0.2 Mbps between the third quarter of 2011 and the first quarter of 2012; two quarters before the Claro merger occurred. The same trend can be observed after the merger, as the average download speed increased by about 0.2 Mbps between the first and third quarters of 2013. Perhaps, the most salient feature regarding download speeds is the increase around the time of the Claro merger. Between 2012 Q1, two quarters before the merger, and 2013 Q1 the average download speed increased by about 0.7 Mbps. This increase is almost 0.5 Mbps larger than the increases observed before and after the merger. Some of the difference between the secular increase of about 0.2 Mbps and the 0.7 Mbps increase during the merger was caused by the merger. As seen in the last row of table 1, the evolution of upload speeds over time exhibits the same pattern.

23 Table 2: Descriptive statistics for Internet providers (Mean and Standard Deviations) 2011Q3 2011Q4 2012Q1 2013Q1 2013Q2 2013Q3 Markets 27.31 28.03 28.01 27.22 27.97 2715.62 (61.04) (54.39) (55.21) (54.32) (55.92) (56.58) Plans 27.44 30.74 30.75 24.63 29.24 28.01 (29.53) (35.70) (36.40) (28.51) (41.18) (31.51) Subscribers (×1000) 95.29 108.21 113.86 105.83 111.21 113.90 (225.90)(241.08)(257.14)(281.76)(295.98)(304.74) Markets refers to the number of markets am ISP operates in. Plans refers to the number of plans the ISP oﬀers in a market.

1.3.2. Characteristics of the providers

81 firms provide services in 703 municipalities.7 Table 2 shows the main characteristics of the average ISP. The average firm provides services in fewer than 29 markets. However, large firms like Coltel are present in over 300 and a few small ones operate in single cities. A firm offers, on average, less than 30 plans. It is worth noting firms´ menus are heterogenous. EPM, for instance, has over 130 plans in a quarter, whereas the menus of many small providers comprise a single plan. The same heterogeneity applies to the subscriptions. For instance, Telmex has over 1.3 million subscribers, whereas Delcatel S.A. reports 32 subscribers during the third quarter of 2013.

1.3.3. Download speeds over time

Figure 1 displays the average download speed in Mbps for the four largest wired Internet providers in the country during the period surrounding the Claro merger. The ﬁrst salient feature of the plot is the change in trends around the time of the merger. Before the merger, average download speeds were increasing, albeit at a low rate. After the merger, the increase in speeds

7There are 1035 municipalities in the data. The smallest towns are not included in the analysis.

24 Figure 1: Average download speeds over time

3.5

3.0 Provider ETB Telefónica Mbps 2.5 EPM Telmex

2.0

2011Q3 2012Q1 2012Q3 2013Q1 2013Q3 2014Q1 Quarter was much faster for all operators except ETB. Another feature of the data, distinctly visible in the plot, is the change in Telmex’s (the ISP involved in the merger) position relative to the other ISPs. Telmex’s average speed before the merger was the slowest if one looks at the last quarters of 2011 or the first quarter of 2012. Just one quarter after the merger, Telmex was offering, on average, much higher download speeds than its major rivals. All providers increased their averages speeds over this period, as it coincided with a higher demand for faster broadband. However, Telmex increased its speeds at a higher rate at a time that coincided with the merger. Figure 2 tells a similar story, using the number of subscribers as weights. The plot suggests, because of the merger, the merging firm increased quality as measured by broadband download speeds.

1.3.4. Price over time

Firms sell differentiated internet plans. Plans may differ along many dimensions. They may differ in the technology chosen for their final delivery (ADSL, WiFi, fiber, etc.); they may grant access to cloud storage or email, etc. How- ever, the most important difference between internet plans is their download

25 Figure 2: Average download speeds over time(weighted by subscribers)

3.5

Provider 3.0 ETB Telefónica

Mbps EPM Telmex

2.5

2.0 2011Q3 2012Q1 2012Q3 2013Q1 2013Q3 2014Q1 Quarter

Table 3: Average download speeds before and after Before After Difference Affected by merger 2.28 3.02 0.74*** 0.03 0.04 0.04 Not affected by merger 1.21 1.39 0.18*** 0.01 0.01 0.01 Difference 1.07 1.63 0.56*** 0.11 0.15 0.13 Affected by merger refers to cities where the merger Telmex was operating at the time of the merger. Download speeds in Mbps. Before: 2011Q3-2012Q1. After: 2013Q1-2013Q3. Mean speed is weighted by number of subscribers.

Table 4: Median download speeds before and after Before After Difference Affected by merger 2.13 3.03 0.92*** 0.03 0.03 0.04 Not affected by merger 1.14 1.44 0.31*** 0.01 0.01 0.01 Difference 0.99 1.6 0.61*** 0.12 0.15 0.14 Affected by merger refers to cities where the merger Telmex was operating at the time of the merger. Download speeds in Mbps. Before: 2011Q3-2012Q1. After: 2013Q1-2013Q3. Mean speed is weighted by number of subscribers.

26 Figure 3: Average unit prices over time

● 30 ● ●

●●●●●●● ●●●●●●●●●● ●● ●●●●● ● ●● ●●

●●●●● ●● ● 20 ●●● ● ●

●

● ●●● ● ● ● ●● ● ● ●● ●● ● ● ●● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●●●●●●●●●●●●● ●●●●●● ●●● ● ●● ●● ●●●

●●●●●●● ●●●● ●● ●●●● ●●● ●●●● ●●●● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● 10 ● ●● ● ●● ● ●●●●●●●● ● ● ●●●●●●●● ●●●●●●●●● ●● ●●●● ●●● ●●●●●●●●●●● ● ●● ● ●● ● ● ●● ● ● ● ●● ●● ● ●●● ●● ● ● ● ● ●

Download speed (Mbps) Download ● ● ●●●●●●●●● ● ●●●●● ●●●●● ●● ● ●● ● ●●● ●● ●● ● ● ●● ●●●●● ●●● ●● ●● ●● ●●● ●●● ● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ● ●●●●● ●●●● ● ●●●●●● ●●●●●●●●● ●●● ●● ●● ●● ●●● ●●● ● ●● ● ● ●●● ● ●●● ●●●● ● ●● ●● ● ● ● ●● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●●● ●●●● ●●●●● ● ● ●●●●● ●● ●●●●● ● ●● ● ●●●● ● ● ●● ●●●●● ●●● ●● ●●● ●●● ●● ●● ● ● ●●● ● ●● ●●●●● ●● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●● ●● ●●●● ● ●●●●●●●●●●●●●● ●●●●● ● ● ●●●●●●●●●●●●●● ●● ●●●●● ●●● ●●● ●● ● ●●●● ● ●●●●●●●● ● ● ●●● ●● ● ● ●●● ● ● ● ●●●● ● ●● ●●● ●● ● ● ● ●● ● ● ● ●●●● ● ● ● ●● ●●●●●● ● ●● ● ● ● ●● ● ● ●● ●●●● ● ●● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●●●●● ●●●●●●● ● ●●●●●●●●●●●● ● ●●●● ●●●●● ●●●●●●● ● ● ● ●●●●● ● ●●●● ● ● ●●●● ● ● ●●●●●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●●●●●● ●● ● ● ●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●● ●●● ●●●●●●●●●●●● ●●●● ●●●● ●●●●●●● ●●●●● ●●●●●●●●● ●●●●●●●● ●●● ●● ●●● ●●●●●●●●●●●●●●● ●● ●●●●●● ●●●●●●●●●●●● ●●●●●●●● ●●●●● ●● ●●●●●● ●●● ● ●● ●●● ●● ● ● ●●●●●●●●●● ● ● ● ● ● ●● ● ● ●●● ● ● ● ●●● ● ● ●● ● ●● ●●● ● ●● ●● ● ● ●●● ● ● ● ●●●●●● ● ●●●●● ● ●●●● ● ● ●●● ●●● ●●● ●●● ● ● ● ●● ●● ● ●● ● ●● ● ● ●● ● ● ●●● ●●● ●●● ● ● ●●● ● ●●●● ●●● ●● ● ●●● ● ● ● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●● ●●●●●●●● ●●● ●●●● ● ●●●●●●●● ●●●●● ● ●●●●●●●●●●● ●●●●●●●●●● ● ●●●● ● ●● ● ● ●● ●● ●●●●● ●● ● ●● ●●●●●●●●●●● ●● ● ●●●●●●● ●●●● ● ●●● ● ● ● ●●●●● ●● ● ●●● ●●● ●●●● ● ●● ● ● ●● ●●● ●●● ● ●●●●●●●●●●●●●●●● ●●●● ●●●● ●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ●●● ● ●●● ●● ●●● ●●● ●●●● ● ●●●●●●●●●●● ● ●●●● ●● ●● ● ●●● ●●●● ● ● ●●●● ●●●● ●●● ● ● 0 ● ● ● ●● ● ●●● ● ● ● ●●● ● ●● ●● ● ● ●● ● ● ● ●●●● ● ●

0 20 40 60 80 Dollar per Mbps

Figure 4: Average unit prices over time 4.4

4.2

Provider EPM 4.0 ETB Telefónica Telmex Dollars per Mbps 3.8

3.6

2011Q3 2012Q1 2012Q3 2013Q1 2013Q3 2014Q1 Quarter speed. As download speeds vary across plans, prices do so too, reflecting such vertical differentiation. Plans with a download speed of 40 Mbps are naturally more expensive than plans offering 20 Mbps, although, typically, not twice as expensive. Figure 3 shows there is a negative relation between the price paid by Mbps and total download speed of the plan. Given that download speeds increased during the quarters in the data, analyzing the sticker price of plans without accounting for the rise in speeds would overstate any increase in prices. Here, the quantity I want to analyze is the unit price, defined as dollars paid per Mbps, which conveys far better information about the evolution of prices.

27 Figure 4 shows the average unit price for wired internet provided by the four major ISPs. Average unit prices for all ISP stayed between $3.6 and $4.4 during the quarters in the data. The main feature of the price data is there is no evident change in the period following the merger. Telmex’s average unit price oscillated from $3.7 to $4.0, but its relative position in the first and last quarters is the same regarding other prices. This suggests the merger did not affect the prices. The general ranking of the average unit prices remains the same through- out the period surrounding the merger. EPM charges the highest unit price followed by Telefónica and Telmex. The average cheapest price per Mbps cor- responds to ETB. However, one should not interpret this ranking of average unit prices as evidence of ETB charging more for their plans. It is worth remembering that EPM and Telefónica provide their services in smaller cities and have more plans with low speeds than their counterparts. As seen in figure 3, unit price is decreasing in the amount of Mbps included. For a plan with a download speed of 2 Mbps, a firm will charge less than twice the price it charges for a plan with 1 Mbps of download speed.

1.4. Empirical Strategy

In this section, I assess the effects of the mergers. I start by looking at how speeds and prices in markets affected by the merger changed in relation to markets not affected by it. This way, I can obtain an estimate of the overall effect of the merger on markets where it took place. Second, I compare speeds and prices of plans provided by small firms that competed with the merging firm (Telmex) to similar firms that did not compete with it, because they operated in markets not affected by the merger. This exercise provides a measure of the effect of the merger on rivals. Finally, I compare the speeds of

28 the merging firm to its rivals to get a lower bound on the effect of the merger on the merging firm. The following subsections describe the three approaches in more detail.

1.4.1. The nature of the diﬀ-in-diﬀ

In all three exercises, to gauge the effect of the merger, I employ the same strategy, difference-in-differences (DID). In every case, I will apply difference- in-differences to a quantity of interest y. The first difference refers to the change in y before and after the merger. The second difference compares these changes between control and treated groups. When assessing the effect of the merger on markets, the treated group comprises cities where, at the time of the merger, Telmex was operating. The control group contains cities where Telmex was not operating when the merger happened. To look at the effect of the merger on rivals, the treated group is small ISPs that compete with Telmex, while the control group will be small ISPs that do not face Telmex as a competitor. Finally, I assess the effect on the merged by comparing Telmex to the other large ISPs it competes with. Since the definitions of before and after are arbitrary, I will try alternative definitions and show the direction of the results is robust to such definitions. The conventional assumption to obtain a good difference-in-difference estimator requires that, absent the treatment, the average outcomes for the treated and control groups would have followed parallel paths. As seen in figures 1 and 3, that is the case in the data. Another arbitrary decision is how long the before and after periods should be. The length of the before period is bounded (downwards) by events that must be avoided if one wants a clean identification of the effect of the merger. I limit upwards the length of the after period to make sure the effects obtained can be attributed to the merger

29 and not to other events that may have occurred after the merger. The longer the period after the merger, the more likely that something unobserved could have affected download speeds and prices. Since the definitions of before and after are arbitrary, I will try alternative definitions and show that the direction of the results is robust to such definitions. The conventional assumption to obtain a good difference-in-difference estimator requires that, in the absence of the treatment, the average outcomes for the treated and control groups would have followed parallel paths over time. As seen in figures 1 and 3, that is the case in the data. Another arbitrary decision is how long the before and after periods should be. The length of the before period is bounded (downwards) by events that have to be avoided if one wants a clean identification of the effect of the merger. I limit upwards the length of the after period to make sure that the effects obtained can be attributed to the merger and not to other events that may have occurred after the merger. The longer the period after the merger, the more likely that something unobserved could have affected download speeds an prices. Besides, there is another reason for wanting a short after period. The merger could affect prices and download speeds via two mechanisms. The merger may affect the production function of the ISP, for instance, if the ISP can reduce its cost of providing a speed by having access to the mobile carrier’s networks. However, the merger could also affect speeds and prices provided on a purely strategic basis. If the products provided by the ISP and the mobile carrier are complements, an increase in broadband download speeds increases the demand for mobile services. The merger internalizes the extra profits created by that complementarity. So, even with no cost synergy, the merger can affect prices and download speeds. Using a short after period ensures the effect captured by the DID is due to the internalization of extra revenues. Should there be any

30 cost synergy, the DID would underestimate the eﬀects of the merger, which is always safer than overestimating. I go into more detail below.

1.4.2. Diﬀerent speciﬁcations

First, I want to learn about the effect of the merger on the merged firm. Observing the variable of interest before and after the merger is not enough. Without knowing what the firms would have done had they not merged, a comparison between the pre- and post-levels of download speed or price is futile. Instead, I compare the changes in the variables of interest for the merging firm to those of its rivals by estimating

yfct = φ0 + φ1T ELMEXf + φ2AF T ERt + φ3T ELMEXf × AF T ERt + x

µ + γ1F ACES_MERGERf + γ2AF T ERt+ (1)

The indexes in equation ?? are f for ﬁrms, c for cities and t for quarters.

T ELMEXf is an indicator function that takes value one when the ﬁrm is

Telmex and zero otherwise. Similarly, AF T ERt is an indicator function for quarters after 2012 Q2. Finally, the interaction T ELMEXf × AF T ERt identifies observations that share two characteristics: the provider is Telmex and happened in quarters after the merger. The value of the estimate on φ3, captures the effect of the merger on the merged ISP. Arguably, a more interesting measure of the impact of the merger is the response of firms that competed with the merging ISP. Changes in the characteristics of services provided by the merged firm might result from cost reduction caused by the merger, whereas rivals would respond strategically to the new market structure. I compare the performance of the merging firm’s rivals to

31 the performance of similar ﬁrms operating in cities unaﬀected by the merger. For such purpose the equation I estimate is

yfct = µ + γ1F ACES_MERGERf + γ2AF T ERt + (2)

γ3F ACES_MERGERf × AF T ERt + δst + αf + εfct

In equation 2, yfct is a measure of download speed -or price- for firm f in city c during quarter t. F ACES_MERGERf is an indicator for firms that compete with the merging firm in their markets. AF T ERt is an indicator for quarters after the second quarter of 2012. Last, δst are state-quarter fixed effects. The parameter γ3 captures the effect of the merger on speeds offered by firms competing with the merged entity. The last task is to get a robust estimate of the average effect of the merger for the cities that experienced it. For that purpose, I estimate the following equation:

yct = α0 + ξc + α1AF T ERt + α2MERGERc + α3AF T ERt × MERGERc + εct (3)

The variable of interest, yct, is regressed on a constant α0, a market fixed effect ξc and a set of indicator functions. The market fixed effects ξc control for city specific characteristics that are unobserved and time invariant, like orographic conditions. The indicator function AF T ERt takes value one for quarters after 2012 Q2. Another indicator function, MERGERc, is equal to one just in markets where Telmex operated during the merger. The interaction

32 AF T ERt × MERGERc will take value 1 for quarters after 2012 Q2 and markets where Telmex was providing its services. After estimating 3, the estimate for α3 should capture the eﬀect of the merger in cities where the merger took place, relative to those cities that did not experience it.

1.5. Results

This section presents the results of estimating the equations presented above. I begin by showing the aggregate effect of the merger on markets. Then I present the estimated effect the merger had on its rivals. Finally I provide a lower bound for the effect of the merger on the merged firm.

1.5.1. Eﬀects on markets

Table 5 presents results from the fixed effects analysis described above. The estimation contains fixed effects at city level, a linear time trend, and an indicator for the presence of Coltel in the market. The large negative and statistically significant intercept reflects how markets differ, regarding download speeds. The coefficients on After and Merger are not statistically different. The coefficient on their interaction is statistically significant. Jointly, these two facts provide important evidence. First, it implies neither the quarters after the merger nor the markets where the merger happened are special. Had the merger not happened, the average speeds in markets where Telmex operated would have stayed constant. Second, they imply the merger affected the speeds of those markets affected by it. The average effect of the merger on a city’s mean download speed is .64 Mbps. Before the merger, the average speeds in the markets affected by the merger was 3.1, so the merger increased average speeds a little over 20%. Measures of central dispersion, such as the median and the mean, however,

33 Table 5: Median download speed in the market Model 1 Model 2 Intercept -146.61*** -146.92*** 68.95 68.92 After 0.04 0.04 0.07 0.07 Merger 0.12 0.10 0.11 0.11 After/Merger 0.64*** -0.64*** 0.12 0.12 Coltel -0.06 0.76 Adj. R-squared 0.76 0.76 Observations 2,993 2,993 The dependent variable is the median download speed offered in the market in megabits per second. Column 2 adds a dummy for markets where Coltel was present. After is a dummy for quarters after the third quarter of 2012. Merger is a dummy for markets where the merged firm operated in the second quarter of 2012. are not enough to paint a realistic picture of the effect the merger had on the markets’ speeds. For instance, mean speeds could rise if firms retire the slowest of broadband plans, despite not increasing speeds for the rest of their plans. Tables 6 and 7 show the results of estimating similar specifications to 3 with the minimum and maximum speeds available in the market as variable of interest, respectively. The data show minimum speeds increased by about 0.2 Mbps because of the merger. Perhaps more importantly, the data show, because of the merger, the maximum speeds in markets where the merging firm operated increased by almost 0.7 Mbps. Next, I consider how the merger affected pricing. As shown above, downloads speeds rose due to the merger. As plans get better, they become more expensive. Instead of looking at the advertised prices, I look at unit prices, defined as advertised price divided by download speed. Table 8 shows the results of regressing unit price on dummies accounting for

34 Table 6: Minimum download speed in the market Model 1 Model 2 Intercept -42.63** -41.32** 28.63 28.57 After 0.13 0.12 0.18 0.18 Merger 0.12 0.10 0.15 0.13 After/Merger 0.20* -0.19* 0.13 0.13 Coltel 0.04 0.23 Adj. R-squared 0.72 0.73 Observations 2,993 2,993 The dependent variable is the minimum download speed offered in the market in megabits per second. Column 2 adds a dummy for markets where Coltel was present. After is a dummy for quarters after the third quarter of 2012. Merger is a dummy for markets where the merged ﬁrm operated in the second quarter of 2012.

Table 7: Maximum download speed in the market Model 1 Model 2 Intercept -146.61*** -146.92*** 68.95 68.92 After 0.04 0.04 0.07 0.07 Merger 0.12 0.10 0.11 0.11 After/Merger 0.64*** -0.64*** 0.12 0.12 Coltel -0.06 0.76 Adj. R-squared 0.76 0.76 Observations 2,993 2,993 The dependent variable is the maximum download speed offered in the market in megabits per second. Column 2 adds a dummy for markets where Coltel was present. After is a dummy for quarters after the third quarter of 2012. Merger is a dummy for markets where the merged ﬁrm operated in the second quarter of 2012.

35 Table 8: Unit price of plan offering the median speed Model 1 Model 2 Intercept 35.72*** 37.13*** 2.68 4.34 After -0.08 -0.04 0.74 0.8 Merger -1.64*** -1.64*** 0.69 0.67 After/Merger 5.20 5.18 5.35 5.32 Coltel -1.43 4.71 Adj. R-squared 0.87 0.87 Observations 2,746 2,746 The dependent variable is the advertised price of the plan divided by the speed in Mbps. Column 2 adds a dummy for markets where Coltel was present. After is a dummy for quarters after the third quarter of 2012. Merger is a dummy for markets where the merged firm operated in the second quarter of 2012. the merger and fixed effects. The coefficient on Merger is negative and statistically significant. The reason behind this is the markets where the merger took place are larger than the markets in the control group. The broadband plans sold in those markets are better and have faster download speeds. The negative sign captures the fact that extra megabits per second of download speed are cheaper. The coefficient on the interaction between After and Merger, which should capture the effect of the merger, is positive but not significant. This means that unit prices of plans offered by the merging firm did not increase after the merger. Interpreting this result warrants caution. Unit prices set by ISPs are usually non-linear. Consumers subscribed to faster plans pay less per unit of speed (Mbps). This results in the unit price of Internet plans decreasing in speed (the outlay per Mbps is usually lower for a 50 Mbps than for a 20 Mbps plan). There are several plausible explanations for why the unit price

36 remained unchanged while the speed increased. For instance, if firms could extract more consumer surplus after the merger, they could do it by offering faster download speed but charging higher prices than before. In this context of non-linear prices it is hard to make an assessment about whether the effects on prices were pro or anti-competitive. One other way of seeing the effect is by looking at the ratio between the increase in speeds and the increase in prices. The ratio of the change in average speed to change average prices is almost one. On average, speeds increased by 1.2 Mbps (20%) whereas prices increased by about 4.5 dollars (22.5%).

1.5.2. Eﬀects on rivals

The next step is to gauge the effect of the merger on firms that operated in markets affected by it. I will refer to firms competing with the merged firm as rivals. Here, firms with similar characteristics to Telmex rivals that operated in markets unaffected by the merger will serve as control group. Table 9 shows the estimated coefficients resulting of estimating differences in differences for the median speeds offered by rivals and the control group. The positive and significant estimate on Merger is expected. Telmex, the merged company, provided larger markets where households have higher incomes. Because they are richer, households in those markets demand higher speeds. The positive coefficient is capturing that Internet plans are faster in the markets where the merger took place. The coefficient on the interaction of After and Merger is positive and significant. The rivals’ average response to the merger was to increase their median speeds by almost half Mbps. ISPs are very diverse. One can naturally divide them into two groups: a small group of large ISPs with presence in many markets across the country and a large group of small ones that are almost exclusively local. It is reasonable

37 Table 9: Median speed (Pooled) Model 1 Model 2 Intercept -58.43* -58.27* 54.54 51.54 After -0.19 -0.19 0.19 0.19 Merger 0.28*** 0.28*** 0.12 0.12 After/Merger 0.46** 0.46** 0.19 0.19 Coltel 0.01 0.01 Adj. R-squared 0.73 0.78 Observations 7,536 7,536 The dependent variable is the median speed offered by the firm in a market. Column 2 adds a dummy for markets where Coltel was present. After is a dummy for quarters after the third quarter of 2012. Merger is a dummy for firms that compete with the merged. to believe firms in these two groups will react differently when competing with the merged firm. Column 1 on table 10 shows the estimates for the sub-sample of small rivals, while column 2 does it for large rivals. The main takeaway from the table is that Telmex’s larger rivals responded to the merger by increasing the speeds, whereas small rivals kept the speeds offered constant. Although not shown in the table (available in the annex), I ran a regression similar to column 1 with the minimum speed as dependent variable, instead. The results of such exercise provide evidence that small rivals increase the minimum speeds they provided because of the merger. It looks as if small rivals perceive low speed plans as their strong product. They appear to have distanced themselves from the higher speeds offered by the merged and the large rivals to reduce competition.

38 Table 10: Median speed Small Rivals Big Rivals Intercept -35.33* -41.12 34.29 51.54 After -0.11 -0.26 0.14 0.39 Merger -0.10 0.18 0.14 0.23 After/Merger 0.27 0.56** 0.29 0.29 Adj. R-squared 0.71 0.68 Observations 4,628 2,908 The dependent variable is the median speed offered by the firm in a market. Column 2 adds a dummy for markets where Coltel was present. After is a dummy for quarters after the third quarter of 2012. Merger is a dummy for firms that compete with the merged.

1.5.3. Eﬀects on merged

Capturing the effect of the merger on the merged firm is challenging. As described in the previous subsection, the merging firm’s rivals increased their speeds because of the merger. Ideally, the control group’s behavior should remain the same with or without the merger. However, using the merged rivals as control can still provide information about causal changes in the speeds provided by the merging ISP. In particular, the proposed exercise should provide one with a lower bound on the causal effect of the merger on the merged. The results of such approach are shown in table 11. Column 1 includes all plans Telmex offered during the quarters in the data. The estimates on column 2 are obtained after dropping plans in strata 1 and 2. Households subscribing to broadband Internet plans in those strata received subsidies around the same time the merger took place. It is plausible the increased demand generated by the subsidies could have affected the speeds offered by the ISPs. The only noticeable difference between columns 1 and 2 is the slightly smaller intercept

39 Table 11: Median speed for plans offered 1 2 Intercept 0.53*** 0.79*** 0.17 0.20 Coltel 0.02 0.16 0.17 0.20 After 0.31 0.33 0.35 0.36 Telmex 1.41*** 1.53*** 0.19 0.22 Coltel/After -0.32 -0.51 0.17 0.20 Telmex/After 1.29*** 1.25*** 0.15 0.26 Adj. R-squared 0.69 0.65 Observations 7,891 7,029 The dependent variable is the median speed offered by the firm in a market. Column 2 removes subsidized plans. After is a dummy for quarters after the third quarter of 2012. Telmex is a dummy for plans offered by the merged firm. in column 1. The smaller intercept is expected, as column 2 keeps more expensive plans that usually provide higher download speeds. In both columns, the coefficient on After is positive, but statistically not significant. The coefficient on Telmex, the dummy accounting for plans offered by the merged firm, is positive because Telmex provided faster downloads. The coefficient of interest, the interaction of Telmex/After, is over 1.2 and statistically significant. So, because of the merger, Telmex increased its median speeds by at least 1.2 Mbps on average. Given that Telmex’s overall average speed was 5.1 Mbps the quarter before, the increment was a 23% increment.

1.6. Comments

The ﬁndings in this paper suggest the changes caused by the 2012 merger between Comcel and Telmex were not as harmful as the popular opinion claims.

40 It appears as if the merger produced pro-competitive effects, at least concerning the provision of wired broadband services. In particular, markets where Telmex operated during the merger, that is, the markets affected by it, saw their download speeds go up. Median speeds in those markets rose by 20% (.64Mbps) relative to similar markets unaffected by the merger. That increase in the median speed of the market was, in part, due to the response of Telmex’s rivals to the merger. Firms competing with the merged increased the download speeds of their plans relative to firms that did not compete with it. Not all firms facing the merger responded equally. There seemed some sort of product differentiation to avoid strong competition. Small rivals increased the speeds of their slowest plans, while large rivals increased the speeds of their faster plans. Finally, the data suggest the merging ISP increased its median download speed by around 1.2 Mbps (23% increment). All these increases in speeds seem to have happened with no important increase in unit prices. The reasons for this apparent pro-competitive effects are out of the scope of this paper. However, one can venture certain possible scenarios. The merger may have internalized profits, reduced costs, or made investments possible (Grzybowski et al. (2014)). The merging firms talked about unifying their customer service as one thing that could enable them to reduce costs. Cost synergies arising from common use of facilities or engineers is not likely in the short-run (Aarnikoivu and Winter (2006)). The increase in speeds could be due to the fact that Telmex was the dominant firm in the markets where it provided services around the time of the merger. Telmex’s dominance could have translated into speeds that grew faster than its rivals. However appealing, this explanation is unable to account for the abrupt nature of the changes. Mobile networks naturally suffer from congestion. The tendency worldwide has been operators that, to relieve some of the pressure imposed on the mo-

41 bile network, allow users to access hot spots. Finally, the effect of the merger can extend beyond download speeds and prices. Research and Development and investment depend on the firm’s profits and its expectations about them. Because the merger altered the profits of the merging firm, an effect on investment and R&D expenditure is highly likely. As pointed out by Hausman et al. (2001), the merger may also provide the merging cable provider with a tool to control alternate channels of distribution.

42 2. Merger Eﬀects with Product Complementarity:

Evidence from Colombia’s Telecommunications

2.1. Introduction

In recent years, there has been a wave of merger among firms producing differentiated and potentially complementary products. Take for instance the merger between Apple and Beats.8 Smartphones and headphones are separate products. In spit of the fact that some consumers enjoy having both, others are content having just one or the other. For consumers who enjoy having both, the utility of consuming them together is typically higher than the sum of their individual components. The products sold by Apple and Beats are textbook examples of complementary products, and the merger between these two firms is not a traditional horizontal or vertical merger. Mergers like this, in particular among the market leaders in their respective fields, often stir a policy debate fueled by concerns about anti-competitive effects. Take Colom- bia’s telecommunications industries as an example. The Claro merger involved Comcel, the largest mobile carrier, and Telmex, the largest Internet Service Provider (ISP) in the country. Consumers were understandably concerned not only due to the massive market power of the merging parties, but because of the cunning nature of the deal.9 Regulators, on the contrary were surprisingly pithy. 10

82015 saw many mergers of these mergers. Other examples are: AT&T-DirecTV, Dell- EMC, IBM-Explorys. 9The merging firms dismissed the possibility of a merger in February 2012 (http://archive.is/KLvxs). By June the same year the companies announced the uni- fication of their brands (http://archive.is/exIwu). By the end of that summer no one doubted they had effectively merged. 10Carlos Rebellón, director of the CRC at the time, said that there was no need to scrutinize the merger (http://archive.is/sLvnC). That is the only statement made by the CRC about the case.

43 The effect of mergers with complementary goods is often uncertain. As shown in Choi (2008), the merger may cause a decrease in prices. When two firms produce complementary goods, their demands are increasing in each other’s prices. A reduction in the price of either firm, increases the profits of the other. Before the merger, the firm reducing the price does not reap the extra profits it generated. Hence, the firms have no incentives to reduce prices. However, the merger internalizes those extra profits making the price reduction optimal. This price reduction causes an increase in economic welfare. But there is also the standard upward pricing pressure that comes with any merger. Besides, the merged firm may use bundling to exert price discrimination. The merged firm could find it optimal to decrease the price of bundles and increase it for standalone goods. If enough people prefer standalone goods, price discrimination may cause a reduction in welfare. The telecommunications sector in Colombia is a great setting for studying these mergers. First, on the grounds that it has been rocked by groundbreaking occasions recently, with the most recent few years being especially tumultuous, including a few mergers of companies delivering complementary services. 11 The other reason that makes the sector attractive is the availability of detailed data. Despite the fact that the regulator’s attitude regarding the scrutiny of all these events can only be described as tepid, they have been very resolute in collecting information about the industry and making it publicly available. This paper explores the dual effects of mergers with complementary goods using panel data from Colombia on market shares, prices, and product characteristics for the universe of telecommunications providers. I begin by esti-

11For instance, early in 2012 Telefónica consolidated its control over Coltel and a few months later Telmex and Comcel merged under the Claro banner. 2013 was the year of the Mobile Virtual Network Operators (MVNO), with Éxito and Virgin entering the mobile market. In 2014 Tigo and Une-EPM merged and DirecTV started providing mobile Internet. Furthermore, there will be soon another merger, as the government of Bogotá decided to sell ETB.

44 mating the demand for five major telecommunications products (Internet, cell phone, TV, landline phone and mobile Internet), in a discrete choice framework that permits complementarities between the different products. I then use the estimated demand model together with an assumption of Bertrand competition to evaluate whether potential mergers between wired and wireless providers in this market would be pro- or anti-competitive. I employ a comprehensive administrative data set put together by the Comisión de Regulación de Comunicaciones (CRC).12 Every provider of telecom services in Colombia is obligated to file CRC’s Form 5 quarterly, and as a result the data set contains information about the universe of firms and plans offered. In Form 5, providers report bundled and standalone services provided in every market, the number of households subscribed to each service as well as the service’s most relevant characteristics. The key feature of the data, besides the high detail of information about characteristics of the products, is the ability to observe market shares for standalone and bundled goods, as this is critical to identify whether the services are complements or substitutes. The framework I use to obtain the substitution patterns builds on Gentzkow (2007). I estimate a random coefficients discrete choice model for telecommunications services. I specify the random utility in a way that allows for consumption of standalone and bundled goods. I additionally adjust the indirect utility function to incorporate the additional utility of joint consumption. Consequently, the indirect utility for bundles has a parameter containing information about the substitution patterns. An estimated negative (positive) sign means that the goods in the bundle are substitutes (complements). The main challenge is to avoid confusing complementarity with correlated tastes or with the fact that the goods are simply bad substitutes. Observing

12CRC is the Colombian analog to the Federal Communications Commission in the US.

45 that frequent subscribers to cable TV often subscribe to broadband as well can be evidence that the two are complements. But it could also mean that the preferences for both are correlated. For instance, a household of cinephiles subscribes to cable TV and Internet because both are media for watching movies. Also, one usually observes joint purchases of goods that are bad substitutes. Households purchase salt and detergent jointly because they are bad substitutes, not because they are complements. With the appropriate data, the model can distinguish between complementarity and correlated preferences or bad substitutes. The first requirement is to have market shares for both bundled and standalone goods. Additionally, the data should include variation on characteristics of the goods since it aids the identification of the substitution patterns. Iden- tification is further enhanced when characteristics that shift the utility of one good do not enter the utility of the other. Luckily for me, the data collected by the CRC has this desirable traits. I find that Colombian households derive extra utility from consuming some hardwired and mobile services together. For instance, there is a strong complementarity between Internet and cell phone services. Every quarter, households get almost $3.13 worth of utility for their joint consumption. Households also perceive TV, mobile and hardwired Internet as complements. They get extra utility worth almost $1.45 from subscribing to all three. In contrast, mobile Internet and land-line phone are strong substitutes. Knowing that wireless and wired services are complementary goods allows me to simulate different scenarios of interest. The first scenario is a merger between ETB and Avantel which provide hardwired and mobile services respectively. The interest in this simulation is twofold. First, it contributes to the understanding of the effects of mergers with complements. Second, it

46 adds to the ongoing discussion around the sale of ETB. This simulation however, cannot account for bundling, since there are no baseline market shares for bundles between services provided by ETB and Avantel. In the second scenario simulated, I break up Claro, a consolidated merged firm, in its two original parts: Telmex and Comcel. Simulating the break-up of Claro provides an insight into the bundling aspect of the mergers. I start by making assumptions about the game that generates the observed distribution of prices. I assume that the firms engage in Bertrand-Nash competition. This solution concept assumes that the firms strategically decide prices for their multiple differentiated products and that the characteristics of the goods are exogenous. I use the First Order Conditions (FOC) of the game to simulate alternative scenarios and solve for the new equilibrium prices. The resulting distributions of prices confirms the predictions from the theory. A merger causes a decrease in prices of bundled services. There is evidence of small increase in prices for standalone goods. The total effect on consumer welfare, however, is positive. The merger increases total consumer surplus by over $7 million dollars every quarter. These findings have momentous implications. Often, regulators prescribe divestiture and proscribe bundling when allowing mergers. Given the potential for pro-competitive effects, the acquisition of firms producing complements could be an alternative remedy to divestiture. If the merger under scrutiny is a merger with complements and consumers have strong preferences for bundles, the regulatory agencies can be more lenient concerning bundling. For the particular case of selling ETB, those deciding whether to sell it or not need only decide to whom. The relation between ETB’s products and those of its eventual buyer matter. My findings imply that it is more desirable to sell the company to a mobile carrier. The results show that in that case consumer

47 surplus increases. Studying the effect of mergers with complements empirically is very important for two separate reasons. First, there is an important economic theory question of what occurs with these mergers. It is well understood that mergers between complements can solve double-marginalization problems and hence are pro-competitive. When the levels of complementarities are large enough, they can be pro-competitive even in the absence of cost synergies. But do products in different sectors have enough complementarities to overcome the standard upward pricing pressure effects from mergers? Economic theory alone cannot resolve this question and hence empirical research can add valuable evidence. Second, in both the U.S. and Colombia, there have been some proposed mergers between providers of complementary services, in the telecommunications and other sectors. For instance, in the U.S., AT&T and Comcast recently announced their intentions to merge. Exactly as in the Claro case, AT&T focuses on wireless services while Comcast focuses on wired services. To evaluate the price impact of any of these mergers, one would need to understand the level of complementarity between these providers and how this level translates into changes in price following the merger. The structure of the rest of the paper is the following. The next two sections define in detail mergers with complements and describe how my work relates to the literature. Next I introduce the data and the main features of Colombia’s telecommunications sector. The next section presents the empirical model of demand used to estimate substitution patterns. Finally, I present and discuss the results of the estimating the model and the counterfactuals.

48 2.2. Mergers: horizontal, vertical and with complements

2.2.1. Types of mergers

There are four types of mergers. Conglomerate mergers, which happen between firms producing goods that are not related. Horizontal mergers, in which a firm acquires a competitor. In typical horizontal mergers the goods produced by the merging firms are very close substitutes; frequently just horizontally differentiated renditions of the same good. Vertical mergers occur between firms that produce a component of a final good. From the standpoint of the final consumer the individual components produced by vertically integrating firms have no use. The consumer can only derive utility once the components are assembled into the final good. Finally, in mergers of firms producing complementary products, consumers derive utility from the standalone goods produced by the merging parties. However, the utility of joint consumption is higher that the sum of the individual utilities derived from the goods.

2.2.2. What we know about these mergers

Research into mergers is a central topic in Industrial Organization and as such has produced many papers.13 The focus, however, has been on vertical and horizontal mergers. With so much consideration given to them, is not surprising that they are well-known phenomena. With some ease, regulators can anticipate whether the effects of one such merger will be anti-competitive or not. Regulators know that the outcome will depend upon two types of forces at odds. Forces that increase welfare and forces that reduce it. The relative strength of these forces will determine the total effect on welfare. In horizontal mergers, cost synergies between the merging firms can increase

13Just to name a few: Salant et al. (1983); Perry and Porter (1985); Salinger (1988); Ordover et al. (1990); Nevo (2001); (Chen, 2001)

49 welfare. The merging entity could produce at lower cost than the sum of the independent costs. However, the merger increases the market power of the resulting firm. The firm can use the extra market power to extract more consumer surplus. The net welfare effect will depend on how the cost reduction and the increased market power balance each other. In a comparable manner, vertical mergers can increase welfare by avoid- ing a double marginalization. Before the merger, the downstream firm’s price includes a markup over marginal cost. The upstream firms also includes a markup in the price it charges to the downstream firm. As a result, there is a double markup in the price faced by the final consumers. The vertical integra- tion inhibits a markup within the merged firm. After the firms are integrated there is only one markup over total marginal cost. However, these mergers can also lead to foreclosure, which can cause competitive harm. Whether the merger increases or reduces welfare will depend on which effect preponderates. Mergers of firms producing complementary goods are not free of ambiva- lence. Suppose that a smartphone manufacturer reduces the price of their smartphones. The reduction in price induces an increase in the demand for smartphones. Indirectly, it will also induce an increase in the demand for headphones. This happens because some consumers enjoy having smartphones and headphones. Thus, the price reduction of smartphones increases the headphones manufacturer’s profits. A merger between these firms internalizes those extra profits. Hence, a reduction in price that is not optimal prior the merger, is optimal after it. This is the reason why the merger with complements can potentially increase economic welfare. But the merged firm may use its recently gained ability to bundle to exert price discrimination. The firm may want to increase the price of standalone goods (because some consumers may want substitute toward the bundle) and

50 decrease the price of the bundle. Under extreme circumstances, the price discrimination may totally exclude buyers wanting to buy standalone goods. This could have an adverse impact on impact on total consumer surplus. Thus, the net eﬀect of the merger will depend on how eﬀective the price discrimination is and the proportion of consumers that prefer standalone goods.

2.3. Relation to the Literature

The literature on the effects of mergers of firms producing complementary goods is sparse. In particular, it pales in comparison with the literature studying vertical and horizontal mergers. Some theory papers examine the behavior of firms that engage in mixed-bundling after the merger. For instance, Choi (2008) provides an analytical framework to study how a merger affects prices when the merged firm can bundle. Choi derives the welfare implications for such mergers under different sets of assumptions. Besides, he examines how the mergers impact the incentives for R&D in the long run. A similar model is presented by Anderson et al. (2010). They show under which circumstances the merger is profitable for the merging parties. They also explore the conditions that may lead to competitive harm. Both models assume that there are no cost synergies. The incentives for the merger stem from the internalization of profits. To the best of my knowledge, no studies are testing the predictions of the theoretical models described above. This paper aims at filling that void. To do so, I draw from Gentzkow (2007) which devises a method for estimating complementarity between goods. In short, he proposes to write the utility of a bundle as the sum of two parts. First, the sum of the utilities of the standalone goods included in the bundle. Second, a term that varies by bundle and whose sign defines a substitution pattern. A negative estimate for the term

51 means that the goods are substitutes. A positive estimate means that they are complements. In a similar exercise, Ribeiro and Vareda (2010) found that in the UK phone and cell phone phone are complementary goods. My estimates confirm that Colombian households, much like their British counterparts, perceived them as complements as well. This paper is similar to Grzybowski and Pereira (2007) in that they assess the unilateral effects on prices of a merger in the Portuguese mobile telephony market. They find that the Portuguese merger caused anti-competitive effects. The relation between mobile and fixed telecom services is likely to vary around the world (Banerjee and Ros (2004)). Several papers attempt to estimate substitution patterns in the telecom industries. Srinuan et al. (2010) try to estimate substitution patterns between Mobile and Fixed broadband for Sweden, using a multinomial logit. They find that Mobile and Fixed broadband are independent. Their results, though, are probably due to the fact that substitution between goods in the multinomial logit depend exclusively on the goods market shares. Andersson et al. (2009) estimate a demand model for text messages and voice with data from Norway. Their results are inconclusive as they find that voice is a substitute for text messages for small network sizes, and a complement for large network sizes. More recently, Grzybowski et al. (2014) investigate whether mobile and fixed telecom are complements using data from Slovakia.

2.4. Colombia´s telecommunications sector

Historically, Colombia had a number of local and regional monopolies providing ﬁxed phone services while Telecom held a monopoly on national and international services. After the 1991 constitution, which emphasized economic deregulation the sector became more open, especially after 1994 when

52 the CRC was created in a bid to foster competition and quality in the sector. Presently, the country has very concentrated mobile and hardwired markets. Regarding its size the sector generated 14 billion dollars in revenue in 2012, which would rank it higher than OECD countries like Belgium or Portugal. Colombia’s telecom infrastructure is well developed in the main cities. There, service availability is high for Latin America’s standards. In small urban areas, however, telecom facilities remain inadequate. The government has vowed to address this with some policies. These policies have had two main goals. First, facilitate the entry of operators in the mobile market in an attempt to improve competition. The second is extending availability of broadband nationally.14 The sector is characterized by a small number of large firms providing services nationwide, and numerous small firms with local presence. Colombia’s leading hardwired providers are Claro, UNE-EPM, Movistar, and ETB. Claro provides cable modem access, UNE offers both DSL and cable modem services, whereas Movistar and ETB use DSL technology. The cable TV sector has undergone significant consolidation and, as a result, Claro and UNE, together control 89% of the national subscriptions to cable TV. After growing fast during the first decade of the century, the mobile market has slowed in recent years. However, there is still potential for growth, especially in the mobile broadband sector where penetration is still low. There is great potential for the adoption of mobile Internet in rural areas were hardwired infrastructure is still poor. Three carriers dominate the mobile market: Claro, Movistar , and Tigo . Claro has been for a long while the leading player, but its market share has fallen steadily since 2009. Another important player is Avantel. It started as a trunking service provider for a small number of cor-

14The government’s main initiatives have been: Computers for Schooling (which provides computers to schools), Compartel (which aims at providing every Colombian with access to telephone and Internet services), Vive Digital (subsidies for Internet access for schools, low-income households, and rural areas), and the National Fiber Optic Project.

53 porate customers using iDEN technology. After winning a spectrum auction in June 2013, Avantel’s LTE networks have grown steadily in the main urban centers making it another important player. In recent years, there has been considerable competition from Mobile Virtual Network Operators. Virgin Mo- bile alone, for instance has almost 2.6 million subscribers in the data.15 As I am writing this paper, DirecTV a long time satellite TV providers is adding LTE Internet to its menu. In general, Colombian authorities clearly understand the need for liberalization and eﬀective competition and have advanced considerably in regulatory terms. The existence of an additional consumer authority aides the CRC in competition issues, allowing for a second scrutiny to ensure that operators do not infringe the principles of competition. Market entry was recently incen- tivized by the adoption of “unique licensing” which reduced the entry barriers greatly. Most of the time, the public concern is matched by the political will and technical implementation that make possible the improvement of market conditions. In spite of the solidly designed institutions, the policy and regulatory framework for Colombia’s telecom have some problems. The lack of independence of the CRC and the high turnover of its commissioners and directors is worrisome.16 On the other hand, while lots of attention has been paid to mobile markets, hardwired markets have been left virtually unregulated. In Colombia, like in the rest of the world, services tend to converge, and providers tend to

15Other MVNOs are: Uff! Móvil and Móvil Éxito who offer mobile voice and Internet services over Tigo’s network. Metrotel and Virgin Mobile have wholesale agreements with Movistar. Since Tigo merged with UNE-EPM it has been able to offer converged services across fixed and mobile networks. 16The Information and Communication Technology Minister is the Chairman of the regulatory agency, which cannot meet without his presence. At the same time, the government owns 30% of the second largest operator in the country. In addition the Superintendent for Trade and Commerce, the other consumer and competition authority, is appointed by the president.

54 oﬀer voice, data and video all together. The regulators act as if they are not aware of that and still regulate hardwired and mobile services as if they were completely independent services. In particular they need to consider bundling when deﬁning markets.

2.5. Data

To estimate the model, I use data from two sources. The ﬁrst source is data from CRC’s Form 5. From the data in Form 5, I get market shares and product characteristics. The second source of information is the Gran Encuesta Integrada de Hogares (GEIH) which is a household survey in the style of the Current Population Survey. From the GEIH I get demographics describing households in each market.

2.5.1. The CRC’s Form 5 data

Firms providing any telecommunications services, be it mobile or hardwired, fill out Form 5 every quarter. The firms report information about the number of subscriptions to individual and bundled services in each city. There are 5 basic services reported by the firms: Internet, cable TV, phone, cell phone and mobile Internet. Also, firms must also report detailed information about the product attributes. For instance, when a bundle includes TV, the firms report if at least one channel is High Definition (HD). They also report whether the subscription includes at least one premium channel, as well as the technology used to deliver the service.17 I do not observe, however, how many or which channels are in the bundle. This raises some concerns for the estimation, as it is likely that

17Premium channels are channels like HBO, AMC or Sundance. TV can be delivered through 3 means: IPTV, cable and satellite

55 more expensive bundles include more premium channels. I describe below how I address these concerns. When the bundle includes broadband, firms report the download and upload speeds advertised. The information about Internet service also describes the technology used for its delivery.18 Form 5 does not account for rented modems. That means that I do not know whether a household is renting a modem or using its own. This should not be a big problem giving the high prevalence of modem rentals around the world. In spite of the fact that I do not have data about how often Colombians rent modems I do know that Americans do it more than 90% of the time.19 Given that Colombians are, in general, less tech savvy than Americans, it is not a wild assumption to say that Colombians cannot be bothered setting up their own modems and over 90% of them prefer to rent them. In the data there is a minute number of residential Internet plans with less than 4 subscribers and with prices well over 6 standard deviations above the mean . The technology used for the delivery of those plans was clear channel (fiber). All other Internet plans delivered via that technology are business plans. I believe there was a mistake when reporting those plans and they ended up coded as residential when they were in fact business. Therefore, I drop them from the dataset. In any case their market shares were infinitesimal. For cell phone and mobile Internet I observe the amount of data usage included. Cell phone plans specify the number of minutes and messages that consumers can use. Form 5 also records the price of additional messages and minutes, in spite of the fact that the actual consumption is not observed nor is the overage. Finally, I observe which units are used to determine charges.20

18The last mile connection can be done via ADSL, DSL, coaxial, ﬁber, wimax, etc. 19http://archive.is/qdoVR 20In some plans charges are calculated by the minute, whereas in others the unit used to compute the charge are seconds.

56 Table 12: Bundle characteristics 2015Q1 2015Q2 2015Q3 2015Q4 Pricea 26.39 29.72 28.31 27.55 (31.13) (38.27) (31.88) (32.10) Download speedb 2.85 3.12 3.43 3.40 (1.92) (2.81) (3.11) (3.26) Premium channelc 0.31 0.33 0.34 0.39 (0.30) (0.32) (0.32) (0.34) HD channels 0.26 0.24 0.27 0.23 (0.19) (0.18) (0.17) (0.15) Datad (Mobile) 3.42 3.68 3.76 3.52 (2.90) (2.87) (2.89) (3.11) Minutese (Cell) 301.01 409.12 410.36 452.31 (100.08) (102.89) (132.38) (137.55) Datad (Cell) 2.68 2.89 3.17 3.20 (4.14) (5.17) (5.28) (4.18) Minutes (Phone) 435.22 482.337 473.21 461.46 (212.17) (283.47) (473.21) (258.81) Observations 28,934 28,965 28,919 28,727 (a) price in 2015 dollars; (b) advertised download speed of bundles with Internet; (c) for bundles with TV (channels like HBO, Sundance, etc.); (d) usage allowance for the plan; (e) number of minutes included with the plan.

57 Table 13: Price and market share of bundles Mean Median Std Min Max Pricea 26.39 25.72 38.31 2.75 619.33 Shareb 1.03 0.96 3.21 0.01 8.59 (a) price in 2015 dollars; (b) percent of market; 115,545 observations.

Providers also report in which market they offer the bundle. For each market they indicate the bundle’s price and the number of households subscribed to it.21 Figure 10 in the appendix shows all the details included in Form 5. Table 12 shows summary statistics for the characteristics of bundles included in the data. Notwithstanding the short period of time, there is variation from quarter to quarter. For instance, download speeds are always higher in the last quarter. Most of the increase in speed is due to firms retiring old plans with low speeds in fringe markets. Table 13 provides basic summary statistics about prices and market shares. There is a huge deal of heterogeneity in the data, that comes from two sources. First, the firms providing selling the services are quite heterogeneous. Some of the firms are giants that operate in almost every market offering numerous plans. For instance, the largest mobile carrier in the data reports over 30 million subscribers every quarter. In contrast, there are small local firms like RIS with 42 subscribers in a quarter and a single plan. The other source of heterogeneity are the plans themselves. Landline standalone plans, for instance, are very cheap. On the contrary, bundles including high-speed Internet and premium channels tend to be expensive. Table 14 presents summary statistics for firms. The most remarkable feature in the table is the large standard errors. Firms are vastly different, which means that the ranges of all the variables presented are enormous. For instance, the

21They also report the price of the individual components in the bundle when there are more than one

58 Table 14: Summary statistics for ﬁrms 2015Q1 2015Q2 2015Q3 2015Q4 Firm level Subscribers 2,214,828 2,211,547 2,211,441 2,213,830 (6,791,197) (6,673,909) (6,543,127) (6,346,187) # plans 123.84 128.35 124.22 118.38 (387.35) (418.88) (347.85) (393.41) Observations 82 82 81 80

Firm-market level Subscribers 649,018 654,520 641,785 674,676 (3,423,785) (3,367,082) (3,040,004) (3,040,004) # plans 54.28 59.61 44.89 57.61 (126.48) (144.84) (109.28) (156.59) Price 15.59 16.00 15.34 17.39 (20.40) (26.24) (24.30) (22.63) Observations 3,521 3,405 4,186 4,402 (a) price in 2015 dollars; (b) advertised download speed of bundles with Internet; (c) for bundles with TV (channels like HBO, Sundance, etc.); (d) usage allowance for the plan; (e) number of minutes included with the plan. Standard errors in parenthesis. number of subscribers ranges from a couple of millions for a large mobile carrier, to a couple of dozens for a small ISP. After controlling for number of plans there is not much variation from quarter to quarter or from market to market. However, there is considerable variation across plans. Figure 5 shows the correlation between the presence of standalone goods in bundles.22 Blue indicates a positive correlation whereas a red indicates a negative correlation. For instance, mobile Internet and cell phone exhibit a strong positive correlation, meaning that households tend to buy bundles that include both. Analogously, Cable TV and Internet, and land-line phone and mobile Internet show negative correlations.

22Table 36 in the appendix shows the numbers behind the heat map.

59 Cell

Cable TV

value 1.0

0.5

Mobile 0.0

−0.5

−1.0

Phone

Internet

Internet Phone Mobile Cable TV Cell

Figure 5: Correlations between standalone goods

2.5.2. Cities, strata and markets

Colombian cities are divided into smaller areas and each area is labeled with a stratum. A number from 1 to 6 identifies each stratum. In the beginning, policy makers created strata to cross-subsidize public utilities (water, electricity and sewage). The idea is straightforward. Higher strata (typically 5 and 6) pay higher prices than they would without stratification. Lower strata (typically 1 and 2) pay lower prices. The extra revenue generated by the higher strata covers the losses of revenue from lower strata. Consequently, the strata were designed to be highly correlated with income.23 Providers of hardwired services quickly realized that they could use strata for price discrimination. Since the stratum number is tied to the dwelling, households cannot engage in arbitrage. A person living in a stratum 3 house cannot buy cheaper cable TV to resell it to a stratum 6 household for a profit. The strata system is perfect for third-degree price discrimination, and firms know it and act accordingly. For that reason, the relevant markets are the

23Formally strata are not supposed to be a measure of income. They are supposed to be a measure of the surrounding amenities. Despite that, strata are an aggregated measure of income as rich people live in areas with better amenities. The summary statistics presented below show that strata increases with income.

60 strata and not the cities. Firms treat each strata within a city as an independent market. Form 5 reflects that and requires firms to report the number of subscribers for each stratum within a city. For many years Telecom, a state owned company, held the monopoly on long distance calls. Similarly, within every region, a public monopoly was responsible for the provision of telephone services. Since the 90’s there has been a wave of privatizations. As a result most of the companies are now private or mixed. The telecommunications sector in Colombia nowadays, looks like that of most countries. There are a few big, dominant firms with an almost national presence. Competing with them, there are numerous small firms whose presence are mostly local. Another feature in determining the present state of the sector has been technological convergence. Today, most firms provide more than one service.

2.5.3. The GEIH data

The other source of data, is the Gran Encuesta Integrada de Hogares (GEIH). The GEIH is a household survey similar to the CPS conducted by the Cen- sus Bureau and the Bureau of Labor Statistics in the US. It samples households from 24 populated geographical regions.24 Households report information about their labor force and living standards. From each one of the 24 geographic areas I get a sub-sample of households and their characteristics. Moreover, I take sub-samples from each stratum in each city. As mentioned earlier, stratiﬁcation divides Colombian cities into smaller areas. These areas are heterogeneous in many dimensions. By sampling from each stratum in each city I get a more precise distribution of households characteristics. This enhances the precision of the estimates on preferences.

24The 24 areas are 13 metropolitan statistical areas and 11 cities. Well over 80% of the country’s population live in those 24 areas.

61 Table 15: Household characteristics by stratum 1 2 3 4 5 6 Schoolinga 6.94 8.28 9.72 11.75 11.85 13.86 (2.45) (2.49) (2.96) (3.42) (3.27) (3.42) HH ageb 0.68 0.61 0.62 0.59 0.56 0.65 (0.47) (0.49) (0.49) (0.49) (0.50) (0.48) Family size 5.34 4.76 3.65 3.86 3.59 3.37 (3.45) (2.27) (1.48) (1.62) (1.53) (1.52) Incomec 386.26 534.17 556.37 933.36 1084.72 1957.3 (366.42) (363.72) (523.72) (997.25) (1549.46) (2277.01) Observations 720 720 720 720 390 390 (a) average number of years of schooling for members within the household; (b) head of the household is between 25 and 45 year old; (c) monthly income in 2015 dollars.

Table 15 shows descriptive statistics for the sample of households. All 24 geographic areas (MSA and cities) have at least 4 strata. Strata 5 and 6 are only present in the main 13 Metropolitan Areas. I sample 30 households from every strata in each area. The simulation step uses those to compute the inner integral in (6). A noteworthy feature of the table is how correlated income (and schooling and family size) and strata are. That is why strata are so useful to exert price discrimination.

2.6. Empirical model

The effect of a merger with complements on consumer welfare depends on the distribution of tastes for bundled and standalone goods. The merged firm can use mixed bundling to price-discriminate, increasing the price of individual goods and reducing the price of the bundles. If many consumers like individual goods, price discrimination can reduce welfare. To be able to understand a merger with complements, it is necessary, first, to know whether the products in the data are complements or substitutes. I draw from the framework devised by Gentzkow (2007) to get estimates of the consumer’s willingness to pay for joint consumption. On the demand side, I estimate a random coefficient

62 discrete choice model for telecommunications services. I write a random utility function that captures the extra utility of joint consumption. Since my goal is to learn about the effect of mergers with complements I need a supply. For the supply side, I assume that the firms engage in price competition in a Bertrand- Nash fashion and that the costs of the merging firms do not change with the merger. Assuming no cost synergies, helps me concentrate exclusively on the strategic aspect of the merger.

2.6.1. Empirical demand

Suppose that in a given market there are F ﬁrms indexed by f ∈ {1, 2, 3, ..., F }.

Let g ∈ {1, 2, 3, ..., Gf } index the standalone services provided by a ﬁrm. With those services, the ﬁrm can sell bundles indexed by b ∈ {1, 2, 3, ..., B(Gf )}. Let any combination of the same basic goods be a type of bundle.25 Let j ∈ {0, 1, 2, 3, ..., 2max{G1,G2,...,GF }}, with zero indexing the outside option, denote those types.26

2.6.2. Utility

The utility that a consumer i derives from subscribing to a service g is given by

X ¯ u¯ig = (pg + ∇pg)α ¯i + xgkβigk + ξg (4) k

where pg is the price, ∇pg is a discount when the good is bundle with other goods (by deﬁnition ∇pg = 0 for singleton bundles), xg1, ..., xgK are the observed non-price characteristics and ξg is the preference for the unobserved

25For simplicity I use the term bundle to denote singleton bundles and the outside option as well as bundles with more goods. 26For instance, in the application below, there will be 5 standalone services: phone, Internet, cable TV, cell phone and mobile Internet. Suppose two bundles containing TV and Phone but diﬀerentiated as one has HD channels. The type of the two bundles is the same.

63 ¯ attributes of the good. The α¯i and ,βigk are the consumer’s preferences for price and characteristic k respectively. I allow these preferences to vary by consumer. In particular, I specify them as follows:

P o u α¯i = α + r zirαr + νipα (5) ¯ P o u βigk = βgk + r zirβgkr + νigkβgk

I decompose a consumer’s preference for attribute xgk into a mean that is constant to all consumers, and a deviation from that mean that depends on the individual’s observed (zir) and unobserved (νigk) characteristics. We can now write the utility of any given bundle b which is of type j as the sum of the utilities of the standalone services g in the bundle, plus a term Γj that is the same for all bundles of type j:

  εib if b = 0 uib = (6) P  g∈b u¯ig + Γj(b) + εib if b > 0

where εib are residual terms assumed to follow a type 1 extreme distribution.

By assumption Γj(b) = 0 for singleton bundles. Putting together equations (1), (2) and (3) we get the utility for bundles

X o X o X u X u uib = δb+ (pg + ∇pg)g zirαr+ xgkzirβgkr+ (pg + ∇pg) νipα + xgkνigkβgk+εib g∈j,r g∈j,k,r g∈j g∈j,k (7) where

X X X δb = (pg + ∇pg) α + xgkβgk + Γj(b) + ξg (8) g∈j g∈j,k

64 As is evident from (4) and (5), every individual’s utility for bundle b has three components. A component that is common across households consuming

P o P o the bundle (δb); a term ( g∈j,r (pg + ∇pg) zirαr + g∈j,k,r xgkzirβgkr) that allows households with diﬀerent observed characteristics to value characteristics and price diﬀerently; and a term not observed by the econometrician that helps to rationalize why households with the same observed characteristics would

P u substitute more strongly between certain bundles ( g∈j (pg + ∇pg) νipα +

P u g∈j,k xgkνigkβgk). Given the distributional assumption on εib, the market share of households choosing bundle b is given by

hP i exp g∈b u¯ig + Γj(b) sb = h idFzdFν (9) ˆ ˆ PF PB(Gf ) P ν z 1 + f=1 l=1 exp g∈l u¯ig + Γl(b)

where Fz and Fνare the distributions of observed and unobserved characteristics of households, F is the number of firms in the market and B(Gf ) is the number of bundles offered by firm f. The estimation consists of finding the parameters that make the predicted shares in (6) close enough to the shares in the data. The task is challenging in challenging in no less than two ways that are worth specifying. The first challenge comes from the choice of distribution for the errors. Under logit errors, there is no analytical closed form solution to the integrals in (6). Second, the goods have attributes that are not included in the data. Those unobserved characteristics could allow the firm charging higher prices. Not accounting for unobservables, may lead to underestimating the effect of price. I tackle the first challenge by computing the integrals by simulation.

65 2.6.3. Instruments

The second challenge is tackled by using instrumental variables. In particular, I construct a GMM estimator, as suggested in Berry (1994). Formally, I need instruments that are orthogonal to the unobserved characteristics of the bundles. In particular, let H be a matrix with the instruments and ω(θ∗) be the error term, written as an implicit function of the parameters. Then, the moment conditions are given by

E[Hω(θ*)] = 0 (10)

For the instruments to be valid, there must be a correlation between them and the prices. Also, the instruments have to be uncorrelated with the value consumers assign to the bundle. For instance, the prices of the same bundle in two different markets are correlated. This is so because the marginal costs affect the price determination. If one is willing to assume that the demand shocks are uncorrelated across markets, then, the price of a bundle in a market is a good instrument for the price of the same bundle in a different market. This is the idea behind the so called Hausman instruments used, among others, by (Nevo, 2001). Similarly, the price of a bundle is correlated with the characteristics of competing bundles. The reason is that rational firms decide their pricing strategies based on their competitors’ attributes. Hence, when pricing a given bundle the firm considers the characteristics of competing bundles. For the characteristics of other bundles to be valid instruments, one needs to assume that the characteristics are decided exogenously or at least they are predetermined. This is the kind of instruments that Berry et al. (1995) advocate for.

66 2.6.4. Identiﬁcation of complementarity

Suppose, for simplicity, that there are only 2 goods: A and B. The goal is to know whether the goods are complements or substitutes. For that, we need to estimate ΓAB, the parameter describing the substitution pattern. Moreover,

ΓAB > 0 means that the goods are complements whereas ΓAB < 0 means that the goods are substitutes. The idea behind the identiﬁcation of the parameter is the following: the estimation should rationalize large shares of the bundled goods (sAB) relative to the shares of standalone goods (sA and sB) with a positive estimate for ΓAB. However, as mentioned before, that is not enough since the preferences for A and B could be correlated. To separate correlated preferences from real complementarity it is necessary to have exclusion restrictions. In particular, any variable entering the utility of, say, good A but not the utility of B nor ΓAB. If there is such a variable then

0 0 observing two values of it (x, x ) adds three new pieces of information (sA(x ),

0 0 sB(x ) and sAB(x )) to help identify ΓAB. For instance, suppose that often people subscribe to both goods (sAB is high relative to sA and sB). If a high value of ΓAB is the reason for it, the goods are complements, and increasing the utility of good A by altering the value of x should also increase the share of good B. If instead, ΓAB is zero and the observed high share of the bundle is the result of mere correlation, the share of good B should not change with diﬀerent values of x. For a more concrete example, in light of the application underneath, suppose A is TV and B is mobile Internet. A large proportion of households subscribe to both goods. There could be two reasons for that. One reason is that the preferences for both services are correlated. This could, for instance, be due to people wanting to follow sports. They watch soccer matches at home using cable TV but when they are away from home they keep track of the score using

67 mobile Internet. If that is the case, the true Γ is zero and correlated preferences explain the relatively high share of bundled TV and mobile Internet. In this case the presence of HD channels should increase the share of TV and bundles with TV, leaving the share of mobile Internet unchanged. High Definition channels, alter the utility of subscribing to TV but have no effect on the utility of using mobile Internet. Thus, the presence of HD channels aids the identification of complementarity between TV and mobile Internet.

2.7. Supply

On the supply side, I assume that the firms decide prices as in a static Bertrand game. The firms may sell several differentiated products. Before the game starts, the firms decide the attributes of their many products. Contingent on its rivals’ attributes, each firm decides a vector of prices that maximizes its profits. I do not impose any restrictions on the marginal costs. Effectively, each bundle will have its own marginal. The FOC for this model of competition are well known and given by

p = mc + Ω(p)−1s(p) (11)

where p and mc are vectors containing the prices and marginal costs and s(p) is a vector of market shares. The dimension of these three vectors is equal to the number of products in the market, say I . The Ω matrix is the product of two matrices. First a matrix O (known as the ownership structure matrix) of dimensions I ×I , whose (m, n) element is 1 only if the same ﬁrm produces bundles m and n and zeros otherwise. Second a matrix ∇s(p) with the price derivatives of the shares. The Bertrand-Nash equilibrium is described by the vector p that solves the non-linear system of equations described in (8).

68 2.8. Counterfactuals

After obtaining the estimated preferences it is possible to carry out different simulations. The simulations help understand the price effect of mergers with complements. As described before, the firms decide their pricing strategies as (static) multi-product Bertrand competitors. The solution concept is Bertrand-Nash and the FOC in equation (8) describe the equilibrium. In this setting, simulating a merger is equivalent to changing zeros into ones and ones into zeros in the ownership matrix. As an example, suppose that we want to simulate the merger of the firms producing bundles m and n. To achieve that, all that is necessary is to change the zero in (m,n) into a one. Note that the first order conditions depend on the marginal costs. Although the marginal costs are not observed I can get estimates for them. After obtaining estimates for the preferences, the only unknowns in (8) are the marginal costs. Plugging the baseline prices, and the estimated shares allows us to retrieve the estimated marginal costs mcd . Under the assumption that the marginal costs, do not change after the merger, a simulated equilibrium can be obtained by solving the system of non-linear equations

post h post post i−1 post p = mcd + O ∇s(p ) s(p ) (12)

where ppost is the vector containing the post-merger prices and Opost is the new ownership structure matrix. With the simulated prices, the eﬀect that the new prices have on a consumer by calculating the compensating valuation

PB P post ! 1 b=0 exp g∈b u¯ig(p ) + Γj(b) CVi = ln PB P pre (13) α¯i b=0 exp g∈b u¯ig(p ) + Γj(b) where b = 0 is the outside option and b = B is the number of bundles available in consumer i´s market. For each market I calculate the total com-

69 pensating valuation as market size times average compensating variation. The ¯ Pns average compensating variation is given by CV = (1/ns) i=1 CVi.

2.9. Results

In this section I present the results of the estimation in two parts. First, I show the estimates for tastes on bundle characteristics. Second, I report and discuss the estimates for substitution patterns.

2.9.1. Bundle characteristics

Table 16 reports the mean preference coefficients on bundle characteristics. Column 1 assumes that consumers have homogeneous preferences. The only difference between households is their individual shock, εij. The specification in Column 2 is a normal logit as well. That specification, however, uses the characteristics of competing products as instruments for the price. Similarly, column 3 estimates an IV logit. This time the instruments are the prices of similar products produced by the same firm in other markets. Finally, the last column contains the result for the full model. It allows households to differ in their tastes for characteristics. The estimate for the coefficient on price has the right sign across all specifications. However, a quick look at the price coefficients across columns confirms that the prices are endogenous. The difference in magnitude between column 1 and the others is evidence of unobserved characteristics that drive the price up. Although the sign of the estimate is correct, the endogeneity dampens the estimated effect of price on utility. The inclusion of instruments in columns 2 through 4 more than doubles the magnitude of the estimate. Both types of instruments help curb the problem caused by the correlation between unobserved quality and price. Despite that, I decide to use Haus-

70 man instruments when estimating the full model. There are two fundamental reasons behind doing as such. First, the magnitude of the estimate on price is larger with Hausman instruments than it is with BLP instruments. That indicates that the former does a better job at explaining prices. Second, my previous work suggests that download speed is a strategic decision for ISPs. They decide their prices and characteristics based on their rival’s. Thus, the exogeneity assumption that makes the BLP instruments attractive is not likely to hold. Car manufacturers decide their products characteristics after a long period of research and development. That means that car’s attributes can’t be easily changed. So, competitors take their rival’s attributes as given. ISPs, on the contrary, can rapidly change download and upload speeds in response to their competitors. The signs of all the estimates are as expected. An increase in prices reduces the utility of subscribing to a bundle. Besides, the price elasticities implied by these estimate correspond in magnitude to those estimated by Grzybowski et al. (2014). On the contrary, households value all the other attributes of the bundles. Higher speeds, more usage data and access to premium channels, all increase utility.

2.9.2. Substitution patterns

Table 17 reports the estimates for the substitution patterns, Γj, and their equivalent dollar amount.27 The appendix reports the actual estimates as well as their standard errors. A positive (negative) number means that the goods in question are complements (substitutes). The magnitude describes a consumer’s willingness to pay (required compensation) for their joint consumption. For instance, consumers perceive Internet and cellphone as complements and

27Table 32 in the appendix reports the estimates and standard errors for all the goods. Tables 33, 34 and 35 show the dollar amounts.

71 Table 16: Bundle characteristics OLS BLP Hausman Random Instruments Instruments coeﬃcients Mean Std. Dev. Pricea -0.16** -0.38*** -0.41*** -0.48** 0.03*** (0.08) (0.15) (0.18) (0.22) (0.01) Download speedb 2.23** 1.57** 1.79** 2.01* 2.01*** (1.05) (0.65) (0.84) (1.09) (0.15) Premium channelc -0.06 0.03 0.03 0.09* 1.74** (0.06)* (0.04) (0.03) (0.05) (0.96) HD channels 0.79 1.32** 1.28** 1.25* 1.65 (0.61) (0.66) (0.52) (0.66) (1.89) Datad(Mobile) 1.04** 1.05** 1.03** 1.04** 2.31** (0.48) (0.51) (0.46) (0.50) (1.69) Minutese (Cell) 0.04*** 0.03** 0.03** 0.03** 1.04** (0.01) (0.01) (0.01) (0.01) (0.64) Datad(Cell) 0.47*** 0.40** 0.39** 0.67** 1.71* (0.11) (0.02) (0.197) (0.28) (1.62) Minutese (Phone) -0.03 0.02 0.02 0.02 6.71** (0.24) (0.09) (90.08) (0.45) (5.73) Observations 115,545 115,545 115,545 115,545 115,545 (a) price in 2015 dollars; (b) advertised download speed of bundles with Internet; (c) for bundles with TV (channels like HBO, Sundance, etc.); (d) usage allowance for the plan; (e) number of minutes included with the plan.

72 Table 17: Estimated substitution patterns for selected goods

Γj Dollar value Internet/Phone 0.82 *** $0.92 (0.13) Internet/TV -1.78 * -$2.00 (1.60) Internet/Mobile 1.35 *** $1.51 (0.27) Internet/Cell 2.79 ** $3.13 (1.38) Phone/TV 0.92 *** $1.03 (0.04) Phone/Mobile -2.11 *** -$2.36 (0.57) Phone/Cell 1.02 *** $1.15 (0.02) TV/Mobile 0.02 $0.02 (0.01) TV/Cell 1.73 *** $1.94 (0.01) Mobile/Cell 2.57 * $2.89 (1.22) Internet/Phone/TV 1.17 *** $1.31 (0.26)

73 the extra utility of having access to both is worth over $3 every quarter. Land- line phone and cell phone are perceived as complements as well. Households value joint subscription to the services at about $1 per quarter. Unsurpris- ingly, the goods included in the bundle marketed as Triply-Play (Phone, TV and Internet) are also complements. The main takeaway from the estimated substitution patterns is that Colom- bian households perceive hardwired (Internet, cable TV, and phone) and wireless (mobile Internet and cell phone) services as complements. Of course, some households are engaging in cord-cutting practices by subscribing just to mobile services but the vast majority still get extra utility from subscribing to both type of services. This is an important piece of information that regulators and policy makers seem unaware of. For instance, the Superintendencia de Comercio and the CRC treat mobile and hardwired services as independent when considering market deﬁnitions, when it is clear from these results that they are not independent. Similarly, when the government subsidizes access to hardwired services like phone or Internet, they need to take into account the spillovers created by the complementarity between them and mobile services. A subsidy on telephone service, for instance, is bound to have some direct eﬀects on consumer surplus via the adoption of phone, but also through the additional enjoyment of other mobile services. A correct evaluation of such programs has to take into account these complementarities.

2.10. Counterfactual scenarios

In the next two subsections I simulate two mergers in a bid to understand the eﬀects of mergers with complements. In the ﬁrst simulation ETB, a hardwired services operator, is merged with Avantel, a mobile carrier. The government of Bogotá owns over 84% of ETB and the council of the city decided in May

74 2016, after debating for 12 hours, to sell their controlling share. The simulation presented below helps the parties involved in the sale of ETB. It helps them decide whether Avantel is a desirable buyer. Besides, the simulation is used to see what happens to the equilibrium distribution of prices under a merger with complements. Because the firms do not sell bundles in the baseline scenario, this simulation recovers the effects of the merger without bundling.28 The second simulation splits a consolidated merger. In 2012 two firms producing complementary services merged under the name Claro. The merging firms were a mobile carrier (Comcel) and an ISP (Telmex). I simulate an equilibrium in which the merging firms operate as independent firms, hence removing their ability to bundle. Since the merger, Claro is able to sell bundles of mobile and hardwired services, potentially using its ability to bundle to exert price discrimination. Simulating the split up of Claro, allows me to understand the impact that bundling has on consumer welfare.

2.10.1. A merger with complements

Figure 6 shows the distribution of prices under two alternative market struc- tures. First, a baseline scenario, in which the market structure is the current market structure. Second, a scenario where two ﬁrms producing complementary goods merge. Namely, the counterfactual scenario simulates the merger between the ISP, ETB, and the mobile carrier, Avantel. The plot shows the distribution of prices for the markets aﬀected by the merger. The upper panel shows the distribution of prices for the standalone goods. The lower panel shows the distribution of prices for bundles containing several goods. In both panels, the bold dark line shows the baseline distribution of prices. The lightest line shows the distribution of prices under the merger with

28Firms that already sell bundles will continue to sell bundles in the counterfactual scenario of course.

75 0.075 Standalone

0.050 density

0.025

0.000

0 20 40 60 Price

0.075 Bundles

0.050 density

0.025

0.000

0 20 40 60 Price

Baseline market structure Merger with complements

Figure 6: Distribution of prices after merging ETB and Avantel

76 complements. Summary statistics for each distribution of prices are shown in tables 18 and19. The upper panel shows a slight increase in the prices of standalone goods. Relative to the baseline prices, the merger with complements shifts mass toward higher prices. Effectively, it means that in equilibrium some of the firms increase the prices of the standalone goods slightly. When able to bundle, the merged firm finds it optimal to increase the price of standalone prices. An increase in price reduces the sales of standalone goods for the merged firm, but some consumers affected by the increase in price, will gravitate toward the cheaper bundle. The simulated counterfactual, however is not capable of accounting for that, because the data do not contain market shares for bundled goods produced by ETB and Avantel. Despite not being able to bundle in this scenario, the mean price for the services sold by the merged firm falls by $1.3 (with a standard error of 0.6). In this simulated equilibrium, the incentive for the merged firm to reduce prices is just the internalization of the extra profits created by the complementarity between the goods. The distributions of prices for bundles exhibit a different pattern. The merger with complements reduces the prices of the bundles of competing firms. The merger with complements shifts mass downwards, relative to the baseline. Since the merged entity is cutting the prices, its rivals have to do so too. Mix- and-match bundles with the cheaper standalone goods are now competing with existing bundles and help drive their prices down. Concerning welfare, the merger generates a net gain for consumers. On average, households get in excess of an extra dollar worth of consumers surplus under the new price distribution every quarter. The total gains in consumer

77 Table 18: Price distributions for standalone goods 1st 2nd 3rd Mean Min. Max. Quartile Quartile Quartile Baselinea 7.79 10.40 16.01 13.68 1.03 49.28 Mergerb 8.23 10.72 16.31 13.72 0.95 49.21 (a) Original market structure; (b) ETB and Avantel.

Table 19: Price distributions for bundled goods 1st 2nd 3rd Mean Min. Max. Quartile Quartile Quartile Baselinea 14.16 17.02 25.88 20.28 8.31 153.83 Mergerb 13.33 14.59 22.55 18.49 7.15 153.97 (a) Original market structure; (b) ETB and Avantel. surplus for the 13 largest cities accrue over 7 million.29 These numbers, however, do not account for the use of price discrimination through bundling. Because the FOC in (9) are based on existing bundles, the simulated scenario allows the merged firm to sell both goods and internalize the profits from the joint sale, but not to sell them as a bundle. Given how small the shares of standalone goods are, it is likely that even accounting for price discrimination the effect on consumer surplus will remain positive.

2.10.2. Splitting up a consolidated merger

Now, I move to simulate an equilibrium that can shed some light on the effects of bundling. This simulation breaks up Claro into the original merging firms: Comcel and Telmex. Currently, Claro sells bundles and standalone goods as well. The counterfactual scenario removes the ability to bundle hardwired and mobile services. The distribution of prices resulting after the split-up of the consolidated merger is shown in figure 7. The density plot does not clearly capture the net effect on the prices of standalone goods. Cheap services like

29Table 37 shows the gains by city and for average households in those cities.

78 0.100

0.075 Standalone

0.050 density

0.025

0.000

0 20 40 60 Price

Bundles

0.06

0.04 density

0.02

0.00

0 20 40 60 Price

Baseline market structure Break−up of Claro

Figure 7: Distribution of prices after breaking up Claro

79 Table 20: Price distributions for standalone goods 1st 2nd 3rd Mean Min. Max. Quartile Quartile Quartile Baselinea 7.23 9.72 15.30 12.90 0.05 49.12 Split-upb 6.79 9.40 15.01 12.86 0.03 49.13 (a) Original market structure; (b) Claro split-up..

Table 21: Price distributions for bundled goods 1st 2nd 3rd Mean Min. Max. Quartile Quartile Quartile Baselinea 18.48 28.60 37.62 28.57 5.99 158.18 Split-upb 20.72 28.96 37.45 29.42 6.15 158.22 (a) Original market structure; (b) Claro split-up.. phone are even cheaper after the break-up, what suggest that the merged firm is using price discrimination (setting higher prices for standalone goods), but the distribution also shifts mass towards more expensive goods. Luckily, table 20 is less ambiguous and shows clearly that the first, second and third quartile are all smaller after the split-up. The lower panel tells a more straightforward story. The resulting distribution of prices after splitting up Claro shows a definite increase in prices, in particular for the mid-range bundles. As can be seen in table 21 the split up increases the prices of all bundles. In general, consumers are better off with Claro operating as unified firm. With Claro’s mobile and hardwired divisions working independently, the consumer surplus is reduced by a little more than $11 million.

2.11. Comments

I estimate demand for telecommunications services using quarterly data from Colombia’s telecommunications sector on subscription and product characteristics. The demand model has two key features: it allows for the demand of

80 bundled and standalone goods and recovers substitution patterns determining whether goods are complements or substitutes. I simulate different scenarios in a bid to understand the effect of mergers with complementary goods, like ISPs and mobile carriers. The estimated substitution patterns identify mobile and hardwired services as complementary. I then simulate a merger between a mobile carrier (Avantel) and a large state-owned ISP that is soon to be sold (ETB). The price effects of the merger are pro-competitive and consumer surplus rises as a result by about $7 million. Simulating the merger between ETB and Avantel precludes the use of bundling as a mechanism to exert price discrimination. To understand how bundling can affect the prices after a merger with complements I proceed retroactively: I split up Claro, a consolidated merger, into the original merging companies: Comcel (mobile) and Telmex (ISP). Breaking up Claro increases the prices of bundles reducing consumers surplus by $11 million. Both simulations confirm that mergers with complementary goods have potentially pro-competitive effects. There are a few courses in which the current model can be enhanced. First, the model does not account for cost synergies between the merging parties. Cost reductions are often cited as being the main motivation behind these mergers and could further increase the welfare gains obtained here. Second, although the data used is far from perfect. With the current data it is only possible to observe bundles sold by the same firm. Consumers could be doing mix-and-match themselves, and subscribing to, say, Internet and Cable TV from two different operators. As a result, the level of complementarity between hardwired and hardwired services is underscored by my model. A data set with information about mix-and-match bundling would improve the assessment of complementarity. Third, this paper remains agnostic about the effect that the

81 merger could have on quality. Quality is likely endogenous and as such will be affected by the merger. For instance in previous work I find that the Claro merger unequivocally increased the average speeds in markets affected by it. Endogenizing quality will certainly make it more difficult both conceptually and computationally, but will enhance our understanding of these phenomena. Lastly, I abstracted away from all the dynamic considerations. There are several ways in which the current work can be improved in future versions. First, the model does not account for cost synergies between the merging parties. Cost reductions are often cited as being the main motivation behind these mergers and could further increase the welfare gains obtained here. Second, in spite of the fact that the data used is wonderful, is far from perfect. With the current data it is only possible to observe bundles sold by the same firm. Consumers could be doing mix-and-match themselves, and subscribing to, say, Internet and Cable TV from two different operators. There- fore, the level of complementarity between hardwired and hardwired services is underscored by my model. A data set with information about mix-and-match bundling would improve the assessment of complementarity. Third, this paper remains agnostic about the effect that the merger could have on quality. Quality is likely endogenous and as such will be affected by the merger. For instance in previous work I find that the Claro merger unequivocally increased the average speeds in markets affected by it. Endogenizing quality will certainly make it more difficult both conceptually and computationally, but will enhance our understanding of these phenomena. Lastly, I abstracted away from any dynamic considerations.

82 3. The Costs of Banning Price Discrimination

under Imperfect Competition: Evidence from

Colombia’s Telecoms

3.1. Introduction

Third degree price discrimination is a common practice. Well-known examples include its use by theaters, diners or car rentals: students pay lower prices than the rest of moviegoers for tickets to the same film; senior citizens get a reduce price on their meals; veterans obtain discounts for renting a vehicle. To treat different types consumers as independent markets, firms must first identify them, often relying on information provided by a third party. Theaters use student identification cards issued by colleges and universities; restaurants ask senior citizens for AARP cards or drivers licenses; and car rental companies require a Veteran Affairs identification card. In this paper, I examine the welfare effects of limiting firms’ access to the information that allows them to price discriminate. If the firm practicing price discrimination is a monopoly, economic theory has clear predictions: price discriminating monopolist makes more profits and some consumers will be better off at the expense of others. The role that information about consumers plays on a monopoly’s profits is straightforward. More knowledge about consumers allows the monopolist to devise more ornate tariffs. More ornate tariffs cannot harm the monopolist because it solves the same profit maximization problem but unconstrained. In other words, the monopolist always retains the option of charging uniform prices. Thus, a monopolist profits are weakly increasing in information about consumers.30

30Some fringe exceptions arise when ﬁrms face commitment problems. But it is well accepted

83 Whether uniform price negatively affects consumers depends on how different the prices are under price discrimination and how many consumers pay each price. But in many industries, like the ones in the examples above, the firms practicing price discrimination are oligopolies. As I discuss below, theory does not provide clear predictions about the welfare effects of banning price discrimination in oligopolistic markets. For instance, not all additional information about consumers is equally useful for the firms. As shown in Armstrong (2010), information about vertical attributes, for instance, is innocuous. Duopolists gain no advantage from learning the consumers’ valuation for the good if consumers don’t care about the identity of the producer. Corts (1998) shows that information about tastes and "choosiness", on the contrary, can be crucial. If information about consumers suggests to one firm that its price to one type of consumer should rise, and suggest to the other firm that its price to the same type of consumer should fall, the outcome could well be that all prices fall in equilibrium. In these cases, a ban on price discrimination could increase the profits for both firms and harm consumers. It is important to empirically analyze the impact of a ban of price discrimination for at least two reasons. First, because the literature is incapable of making a straightforward predic- tion. Second, it is policy relevant in Colombia as authorities are considering removing the stratification system, and should take into account welfare effects beyond basic utilities. In this paper, I use data from Colombia’s telecommunications industry to understand the effect of banning price discrimination. Most of Colombia’s telecom providers practice third degree price discrimination. Telecom providers bundle services, as well, which can be understood as second degree price dis-

that a monopolist is better oﬀ under price discrimination.

84 crimination. For the same plan, they charge different prices to households dwelling in different areas of the city. Colombian authorities are considering policies that will limit the information firms use to identify markets within a city, hindering their ability to exert price discrimination. Using data on prices, market share and characteristics, I estimate a demand model for bundles containing combinations of Internet, cable TV and phone, and use the estimates to recover marginal costs. To understand the effects of a ban on price discrimination, I use the demand and marginal cost estimates to simulate an equilibrium in which firms charge uniform prices. Colombia’s telecom markets are perfect to study competing firms that practice price discrimination. Colombian cities are divided in strata. Households living in the same strata are, to some extent, homogeneous and share characteristics like income, education and family size. Telecom companies treat these strata as their relevant markets, and households in different areas of the city end up paying different prices for identical services. More information about strata and how firms use them is described in detail below.31 As an illustration, table 22 shows a real example from the data, of a plan sold under this kind of price discrimination. During the second quarter of 2015 in Bogotá, Claro sold a broadband plan with a download speed of 20 Mbps. The plan was available to households in strata 3 through 6 at 3 different prices: strata 3 and 4 paid specific prices while strata 5 and 6 paid one price. My goal is to simulate Claro’s optimal response if it had to charge a uniform price to households in strata 3, 4, 5 and 6 for the plan. I am interested in the equilibrium outcome so, in practice, I do the same for every firm practicing price discrimination in all cities. Results indicate that a ban on price discrimination would have almost neg-

31More information about strata and how ﬁrms use them is described in detail below.

85 Table 22: An example of Price Discrimination Price Stratum 3 32.44 Stratum 4 33.88 Stratum 5 39.30 Stratum 6 39.30 Price in 2015 dollars of Claro’s 20 Mbps Internet plan in Bo- gotá, second quarter, 2015. ligible effects on total consumer surplus. The total yearly compensating variation is under $80,000 dollars. Nevertheless, households in lower strata would face higher prices and households in higher strata would face lower prices. Al- most all cities exhibit the same patterns, although the magnitude of the price changes differ. The effects on profits are heterogeneous. Some firms see their profits increase slightly under uniform price. The rest of the article is organized as follows. In the literature section I present the relevant economic theory that explains some of the expected effects of banning price discrimination, as well as empirical papers that relate to this one. In the next section I describe in detail Colombian strata and how firms use them to define the markets where price discrimination will be applied. Then, I present the empirical demand model and the results of estimating it. Then, I show the effects on consumers and firms of banning price discrimination. Finally I discuss the results.

3.2. Literature

The literature on price discrimination and its effects on total welfare is vast. A specially fruitful strain of theory studies the extent to which access to more detailed consumer information affects the firm’s ability to devise more ornate tariffs, its profits and consumer welfare. Armstrong (2010) provides a good

86 survey about that literature. In particular, Corts (1998) obtains conditions under which, competing firms that engage in price discrimination can end up intensifying competition. Similarly, Armstrong and Vickers (2001) find that if markets are sufficiently competitive, firms can always make more profits with discrimination. Dobson and Waterson (2005) show that practicing price discrimination is not always best for a chain-store and they find conditions under which competitive firms raise profits by charging uniform prices. Leslie (2004) estimates a structural model of price discrimination with data from Broadway shows. The estimates allow him to compare welfare under the observed price discrimination and the counterfactual uniform pricing. He finds that firms make more profits under price discrimination while consumer welfare wouldn’t vary much should firms charge uniform prices. Villas-Boas (2009) looks at the effects of banning wholesale price discrimination on business-to- business transactions between supermarkets and wholesalers. Her estimates suggest that welfare increases if wholesalers can’t exert price discrimination. Grennan (2013) studies price discrimination practiced by upstream firms that provide hospitals with medical device markets. According to his findings, if upstream providers practice uniform pricing, hospitals’ profits suffer because it softens competition. Bargaining seems to be the reason why the results are different for hospitals and supermarkets. The results presented below are similar to those found by Gary-Bobo and Larribeau (2004) who estimate a structural model of price discrimination in the French mortgage market and find that the price-elasticity of demand for housing varies with the lender’s characteristics, in particular their occupational status.

87 3.3. Best response asymmetry in Colombia’s telecoms

A key determinant of the sign of the effects of banning price discrimination is how the firms best-response functions relate to one another. If the best- response functions are symmetric, all firms want to raise prices in the same markets should they have the opportunity to practice price discrimination. As a result all firms make more profits, consumer surplus is increased in low valuation markets and decreased in high valuation markets. If the best-response functions are asymmetric, a firm may want to raise prices in markets where other firm wants to lower them, and that can enhance competition. In this scenario, the competition pressure generated by the price discrimination can be such that all prices are lowered and firms profits fall. Although is likely that the vast majority of providers in Colombian telecom markets rank consumers equally, there is anecdotal evidence of enhanced competition due to price discrimination. For many years, DirecTV was marketed at richer households and as a result it has traditionally been perceived as an elitist brand associated with higher strata.32 Since 2010, DirecTV started selling prepaid TV.33 Once quality is controlled for, DirecTV prepaid plans are cheaper than major competitors like Claro or UNE, and as a result, launching prepaid plans aimed at poorer households put competitive pressure on cable operators that were providing services to lower strata.34 One possible example of firms that may exhibit best response asymmetry are ET Palmira and Ca- blevisión. Although competing in the same city and having subscribers in all strata, 85% of ET Palmira’s subscriptions go to strata 2 and 3 whereas 79% of Cablevisión’s go to strata 5 and 6.

32"DirecTV: We used to be seen as an elitist product"(https://archive.is/RKhej). 33"DirecTV goes after strata 1, 2 and 3" (https://archive.is/i75zK) 34https://archive.is/rBRsV

88 3.4. Strata: what are they and how ﬁrms use them

Simply put, strata are numbers that the Colombian government assigns to houses. These numbers typically range from 1 to 6 and correlate well with characteristics like the income of the people dwelling in the house, their level of schooling and family size. The government introduced strata to assist in assigning subsidies for basic utilities like water, electricity and gas. The idea behind the strata was to charge a price below marginal cost to households in lower strata and a price above marginal cost to households in higher strata. The extra revenue generated by overcharging households in higher strata should cover the losses generated by undercharging households in lower strata. The first attempt at regulating the price of utilities comes with the creation of the Bureau for Utility Pricing whose objective was to make sure that households payed according to their purchase power.35 The bureau chose the value of the dwelling where the household resided as a proxy for the household’s purchase power. At first the Bureau used the last price the house was traded for as proxy for its value. Soon, it was evident that this method had flaws.36 Despite its flaws, the Bureau assigned subsidies in this manner for over fifteen years. After that and until 1990, it was the providers who had to appraise every case and decide which households should receive subsidies. This new method of assigning strata imposed a huge burden on the providers, which were already struggling to recover their costs. The 1991 constitution passed on the responsibility of defining the strata to the municipal governments. In the current approach municipal governments must first identify homogeneous cadastral areas. Then, the government collects information about the characteristics of the dwellings in those areas. Some of the characteristics an-

35Junta Nacional de Tarifas de Servicios Públicos ( Law 3069 of 1968) 36For instance, rich households owning large houses that had not been sold for over 300 years were being subsidized because of the low nominal price of their estates.

89 alyzed are kind of materials used on facades, the number of restrooms and the number bathrooms, and surrounding amenities. The predominant characteristics of an area determine how it ranks with respect to other areas within a city. Finally, all the houses within the same homogeneous area are assigned the same stratum. As a result of the methodology used to rank cadastral areas the strata are highly correlated with the characteristics of the households that reside within them. For instance, households living in stratum 6 areas tend to be richer, more educated and smaller than households living in stratum 1. Because they are highly correlated with the household’s characteristics, Ca- ble TV and Internet service providers in Colombia have in strata an almost ideal tool to exert price discrimination. When a household wants to subscribe to a cable TV plan, for instance, it needs to produce a utility bill. Utility bills tell the operator the address where the service would be delivered and the strata associated with the address. The operator can then charge a price that is speciﬁc to people living in that strata. Note that because the strata is tied to the address where the service is provided, poor households can’t buy a plan and re-sell it to richer households for a proﬁt.

3.5. Data

To estimate the model, I use two sets of data: administrative data collected by the CRC through Form 5 and data from a current population survey. The data in Form 5 contains market shares and product characteristics. The current population survey is the Gran Encuesta Integrada de Hogares (GEIH) which is a household survey in the style of the Current Population Survey. From the GEIH I get demographics describing households in each stratum.

90 3.5.1. Form 5

All firms providing telecommunications services fill out Form 5 every quarter. Providers in all cities report information about the number of subscriptions to standalone and bundled services in each stratum. I use data from 3 basic services reported by the firms: Internet, cable TV and phone. In addition to reporting the number of subscribers, the firms report the attribute of such plans as well. When a subscription includes TV, the firms report whether at least one channel is High Definition (HD), whether the plan includes at least one premium channel, and the technology used to deliver the service.37 I do not observe, however, how many or which channels are included with the TV component of the bundle. This raises some concerns for the estimation, as it is likely that more expensive bundles include more or better premium channels. This concern, however, is no different then the usual concern about mean utility that, although observed by the firm and the consumer, is unobserved by the econometrician, generating an endogeneity bias. I describe below such concerns are addressed. When the bundle includes broadband, firms report the download and upload speeds advertised. The information about Internet service also describes the technology used for its delivery.38 Form 5 does not account for rented modems.

37Premium channels are channels like HBO, AMC, Playboy or Sundance. TV can be delivered through 3 means: IPTV, cable and satellite. The utility specification does not include the technology used to deliver the TV service for 2 reasons. First, because average households are unlikely to have strong preferences about one or the other. Second, because even if households could prefer a type of thechnology, usually only one is available per house and households are limited by the only technology available. 38The last mile connection can be done via ADSL, DSL, coaxial, fiber, wimax, etc. Same as witih TV. I don’t include the means used to deliver broadband in the utility. From the consumers standpoint if they subscribe to a 10 Mbps broadband plan, they should get 10 Mbps regardless of if it is delivered via coper or optic fiber. Also, most of the times households can choose as availability of either in the area is determined on a technical level.

91 That means that I do not know whether a household is renting a modem or using its own. This should not be a big problem giving the high prevalence of modem rentals around the world. In spite of the fact that I do not have data about how often Colombians rent modems I do know that Americans do it more than 90% of the time.39 Given that Colombians are, in general, less tech savvy than Americans, it is not a wild assumption to say that Colombians cannot be bothered setting up their own modems and over 90% of them prefer to rent them.

3.5.2. Summary statistics

Table 23 shows the average prices for all the bundles in the data. Internet tends to be the most expensive of the singleton bundles while phone is the cheapest. The most abundant and expensive of bundles is, unsurprisingly, the bundle containing all the services, which is often referred to as Triple Play. As the table sums up plans oﬀered in diverse cities to even more diverse types of customers, there is a huge deal of heterogeneity, as the large standard deviations attest.

Table 23: Price of bundles Obs. Mean SD Internet 1,590 16.32 8.17 Phone 1,888 9.25 5.63 TV 2,748 10.89 3.67 Internet-phone 3,778 21.93 10.31 Internet-TV 1,843 21.95 7.99 Phone-TV 1,175 12.81 5.45 Internet-Phone-TV 5,299 23.47 9.76 Price in 2015 dollars of bundles oﬀered by all providers and all cities.

Naturally, the price of plans vary by strata. As is clear from table 24 house-

39http://archive.is/qdoVR

92 Table 24: Price of bundles by stratum in 2015 dollars Stratum 1 2 3 4 5 6 Internet 14.31 14.17 16.35 19.12 25.00 27.76 (6.03) (5.75) (7.01) (10.14) (11.73) (10.95) Phone 8.24 8.53 9.17 10.15 11.60 12.14 (5.28) (5.62) (5.33) (5.47) (5.61) (6.08) Internet-phone 20.59 20.07 21.32 24.22 28.21 31.72 (8.75) (9.29) (9.29) (10.97) (12.77) (14.14) TV 10.15 10.20 10.30 12.88 15.06 16.45 (2.73) (2.84) (3.12) (3.91) (5.25) (7.39) Internet-TV 19.90 20.11 21.64 23.75 27.38 32.25 (4.67) (4.91) (7.05) (10.06) (12.22) (14.43) Phone-TV 10.73 10.70 13.66 14.18 16.23 16.21 (4.21) (3.43) (5.01) (6.08) (8.04) (8.27) Internet-Phone-TV 20.17 19.95 23.33 25.20 30.27 32.37 (6.85) (6.64) (9.18) (9.82) (12.11) (13.51) Price of average bundle by stratum in 2015 dollars. The numbers in parenthesis are the standard errors. holds in higher strata subscribe to more expensive plans. For instance, the average price of a standalone Internet subscription in stratum 6, is almost twice the average price of a standalone Internet plan in stratum 1. Of course, this is not evidence of price discrimination, as the table does not control for the attributes of those plans. However, as pointed out in de Mercados (2015) cable TV operators usually exert price discrimination -most commonly a price for strata 1,2 and 3, and another price for strata 4, 5, and 6- although in some municipalities the prices are uniform. This may reflect the use of different strategies depending on the profile of the consumers in each city. Table 25 shows how the average characteristics of Internet and Phone plans in the data vary by stratum. Internet plans sold to higher strata tend to be faster both in their download and upload dimensions. For instance, the average download speed households can get in stratum 6 (6.64 Mbps) is almost three times that households in strata 1 can (2.27 Mbps). In contrast, the number of

93 Table 25: Internet and Phone characteristics by stratum (Means and S.D.) Download Upload Minutes (Mbps) (Mbps) Stratum 1 2.27 0.70 308.31 (2.48) (1.10) (1034.47) Stratum 2 2.71 0.76 295.84 (2.89) (1.00) (1051.10) Stratum 3 3.45 0.90 288.94 (3.99) (1.25) (1172.92) Stratum 4 4.75 1.13 269.46 (5.61) (1.43) (1258.47) Stratum 5 5.76 1.24 309.77 (7.28) (1.45) (1272.41) Stratum 6 6.64 1.42 265.68 (8.06) (1.78) (1144.55) Average speeds of bundles containing broadband and phone. Download and upload speed in megabits per second. Minutes refers to number of long distance minutes included with phone subscription. minutes included in phone plans for stratum 6 households seems to be slightly lower than that for plans for stratum 1. This last statement should be taken with caution given the large variation on minutes. In table 26 the means and standard deviations account for the number of subscribers to each plan. Because a large percentage of households falls in lower strata, this table should be a more accurate depiction of the actual average characteristics Colombia households subscribe to. After weighting for number of subscriptions, the gap between characteristics of the average plan available to lower and higher strata is even larger. The average speed to which households in stratum 6 subscribe (7.60 Mbps) is more than 5 times than that of the plans chosen by households in stratum 1 (1.78 Mbps). This persitent gap in adoption of fast Internet has spurred the government to design policies to encourage poorer families to subscribe faster broadband. Similar patterns arise for the characteristics of TV. In general TV plans are

94 Table 26: Bundle characteristics by stratum (Weighted Means and S.D.) Download Upload Minutes (Mbps) (Mbps) Stratum 1 1.78 0.49 306.71 (2.15) (1.11) (1022.52) Stratum 2 2.33 0.68 286.34 (2.45) (1.08) (1046.27) Stratum 3 3.75 1.04 281.24 (3.93) (1.28) (1171.89) Stratum 4 5.27 1.12 264.88 (6.27) (1.46) (1268.46) Stratum 5 6.02 1.00 312.85 (7.87) (1.29) (1285.72) Stratum 6 7.60 1.01 263.71 (10.18) (1.62) (1146.68) Average speeds of bundles containing broadband and phone weighted by number of subscribers. Download and upload speed in megabits per second. Minutes refers to number of long distance minutes included with phone subscription. better in higher strata. For instance, as seen on table 27, only a third of plans offered in strata 1 have at least one High Definition channel, whereas over a half plans offered in the highest stratum have them. While only 3% of cable TV plans sold to strata 1 have premium channels just over 12% of those offered in the highest strata have them. The incidence of Video on Demand in the highest strata almost doubles that of the lowest.

95 Table 27: Characteristics TV (Percent of plans with) HD Premium VoD Stratum 1 0.33 0.03 0.06 Stratum 2 0.45 0.06 0.07 Stratum 3 0.49 0.11 0.09 Stratum 4 0.55 0.13 0.12 Stratum 5 0.56 0.12 0.11 Stratum 6 0.55 0.12 0.13 HD: proportion of TV subscriptions with at least one channel in high-deﬁnition. Premium: proportion of TV subscriptions with at least one premium channel. VoD: proportion of TV subscriptions that include access to a library of titles on demand.

3.5.3. Household Survey

An important source of information for this article is the Gran Encuesta In- tegrada de Hogares (GEIH). The GEIH is a household survey similar to the CPS the Census Bureau and the Bureau of Labor Statistics conduct in the US. It samples households from 24 populated geographical regions.40 The households report occupational information about their working members and their living standards. From each of the 13 largest cities and each strata on them I draw 80 households and their characteristics. Obtaining 80 households from the highest stratas was impossible as the survey sample doesn’t include many of them. Sampling from each stratum in each city permits a finer control on households characteristics enhancing the precision of the estimates on preferences obtained below. Table 28 shows descriptive statistics for the sample of households. A noteworthy feature of the table is how correlated income (as well as schooling and family size) and strata are. The high correlation of households characteristics and strata is the reason why telecom providers find strata so useful to define, within a city, the markets in which exert price

40The 24 areas are 13 metropolitan statistical areas and 11 cities. Well over 80% of the country’s population live in those 24 areas.

96 Table 28: Household characteristics by stratum Stratum 1 2 3 4 5 6 Schooling 6.94 8.28 9.72 11.75 11.85 13.86 (2.45) (2.49) (2.96) (3.42) (3.27) (3.42) HH age 0.68 0.61 0.62 0.59 0.56 0.65 (0.47) (0.49) (0.49) (0.49) (0.50) (0.48) Family size 5.34 4.76 3.65 3.86 3.59 3.37 (3.45) (2.27) (1.48) (1.62) (1.53) (1.52) Income 386.26 534.17 556.37 933.36 1,084.72 1,957.3 (366.42) (363.72) (523.72) (997.25) (1,549.46)(2,277.01) Observations 1,040 1,040 1,040 1,040 1,040 1,040 Schooling: average number of years of schooling for members within the household. HH age: head of the household is between 25 and 45 year old. Family size: number of members of the household. Income: monthly income in 2015 dollars. discrimination.

3.6. Empirical demand

A given city is divided into up to 6 strata indexed by d ∈ {1, 2, 3...}. Operating in the city are F firms indexed by f ∈ {1, 2, 3, ..., F }. A firm can provide any subset of the following standalone services: land-line phone, cable TV and broadband Internet. The standalone services are indexed by g ∈ {1, 2, 3}. Furthermore, the firms can bundle any subset of the standalone services they provide. These bundles are indexed by b ∈ {1, 2, 3, ..., Bf }, with Bf denoting the number of bundles sold by the firm. One famous bundle, for instance, is the one containing Internet, phone and TV. It is known in telecom parlance as Triple Play. From here on,the term bundle will include singleton bundles as well. Finally, the firms may sell the same bundle to different strata at different prices.

97 3.6.1. Utility

The utility a consumer i living in stratum d derives from subscribing to a standalone service g is given by

X ¯ u¯idg = (pdg + ∇pdg)α ¯id + βidgkxgk + ξdg (14) k

where pdg is the price charged to consumers in stratum d for good g and

∇pdg is any discount applied to the good when bundled with other goods

(by definition ∇pdg = 0 for singleton bundles). xg1, ..., xgK are the service’s observed non-price characteristics. The term ξdg captures the utility derived from characteristics of the service that although observed by the consumer and the firm, are not unobserved by the econometrician. Examples of ξdg particular to the telecom industry are coupons, waved installation fees and ¯ modem rentals. The α¯id and, βidgk are the preferences a consumer i living in a house in stratum d has for price and characteristic k respectively. I allow these preferences to vary by consumer. These preferences are specified as follows:

P o u α¯id = αd + r αdrzidr + αd νidp (15) ¯ P o u βidgk = βdgk + r βdgkrzidr + βdgkνidgk

A consumer’s preference for price pdg (attribute xgk) is modeled as deviation from a mean preference αd (βdgk). The responses of individual consumers to changes in the good’s attributes will diﬀer. How diﬀerent the response of an individual consumer is with respect to the average consumer, will depend on the consumer’s observed (zidr) and unobserved characteristics (νidgk). The utility a consumer gets from subscribing to a bundle is written as the sum of the utilities of standalone goods in the bundle. If we let b = 0 be the outside option, then the utility of any bundle is given by

98   εidb if b = 0 uidb = (16) P  g∈b u¯idg + εidb if b > 0

where the εidb are residual terms assumed to follow a type 1 extreme value distribution. Putting together equations 14, 15 and 16 we get the utility for bundles

X o X o X u X u uidb = δdb+ αdr (pdg + ∇pdg) zidr+ βdgkrxgkzidr+ αd (pdg + ∇pdg) νidp+ βdgkxgkνidgk+εidb g∈b,r g∈b,k,r g∈b g∈b,k (17) where

X X X δdb = αd (pdg + ∇pdg) + βdgkxgk + ξdg (18) g∈b g∈b,k g∈b

Note that the utility an individual in stratum d derives from consuming bundle b consists of three parts. First, a component that is the same for all households in the stratum consuming the bundle (δdb) and that only depends on the bundles characteristics, both observed and unobserved. Second, a term

P o P o ( g∈b,r αdr (pdg + ∇pdg) zidr + g∈b,k,r βdgkrxgkzidr) that allows for households with diﬀerent observed characteristics to have diﬀerent tastes for characteristics and price. And third, a term not observed by the econometrician that helps rationalize why households with the same observed characteristics would

P u have diﬀerent attitudes toward the same bundle ( g∈b αd (pdg + ∇pdg) νidp + P u g∈b,k βdgkxgkνidgk). Given the distributional assumptions, the market share of households choosing bundle b is obtained by integrating out the εidb to get

99 hP i exp g∈b u¯idg sdb = h idFzdFν (19) ˆ ˆ PF PBf P ν z 1 + f=1 l=1 exp g∈l u¯idg

where Fz and Fν are the distributions of observed and unobserved characteristics of households, F is the number of firms in the market and Bf is the number of bundles offered by firm f. Described succinctly, the estimation algorithm finds the parameters that make the predicted shares in 19 close enough, under some metric, to the shares observed in the data. The task is not trivial and there are at least two challenges worth mentioning. The first challenge comes from the choice of distribution for the errors. Under logit errors, the integrals in 19 have no analytical closed form solution. The lack of analytical solution to the integrals is solved by simulating the integrals. Second, the unobserved attributes of the goods, ξdg in equation 14, are not included in the data. Because those attributes are known to the firm and the consumers they may be correlated.

That is higher values of ξdg are associated with more expensive bundles. This fact may lead to underestimate the coeﬃcient on price.

3.6.2. Instruments

The second challenge is tackled using instrumental variables. In particular, I construct a GMM estimator, as suggested in Berry (1994). Formally, one requires instruments that are orthogonal to the unobserved characteristics of the bundles. In particular, let H be a matrix of instruments and ω(θ∗) be the error term, written as an implicit function of the parameters. Then, the moment conditions are given by

100 E[Hω(θ*)] = 0 (20)

For the instruments to be valid, they must be correlated with prices. Besides, the instruments should be uncorrelated with the unobserved value consumers assign to the bundle. For instance, the prices of the same bundle in two different markets are correlated, because the marginal costs of providing them influence the price determination. If one is willing to assume that the demand shocks are uncorrelated across markets, then, the price of a bundle in a market is a good instrument for the price of the same bundle in a different market. This is the idea behind the so called Hausman instruments used, among others, by Nevo (2001). Similarly, the price of a bundle is correlated with the characteristics of competing bundles. The reason is that rational firms decide their pricing strategies based on their competitors’ attributes. Hence, when pricing a given bundle the firm considers the characteristics of competing bundles. For the characteristics of other bundles to be valid instruments, one needs to assume that the characteristics are decided exogenously or at least they are predetermined. This is the kind of instruments that Berry et al. (1995) advocate for. I present here the results obtained with Hausman instruments. First, because my previous work suggests that the characteristics of telecom bundles may be endogenous which would render the BLP instruments invalid. Second, the results obtained under BLP, not presented here, are similar in magnitude but Hausman instruments seem to do a better job at reducing the endogeneity bias.

101 3.6.3. Demand Estimates

Figure 8 shows the estimates for the marginal utilities. The plot presents the change in an individual’s utility caused by a marginal changes in price, download and upload speeds, access to premium channels, video on demand and long distance minutes. Each panel depicts a representative consumers response to a bundle’s characteristic by stratum. The coeﬃcients on price have the appropriate negative sign across all strata and are statistically signiﬁcant. Moreover, the sensitivity to a price increase is highest for households in stratum 1 (0.92) and lowest for households in stratum 6 (0.71).41 A negative relation between price sensitivity and strata is expected. Strata exhibit a positive correlation with income and for households in stratum 1 expenditure on telecom services tends to represent a higher portion of their income. Because the income of a representative household raises with stratum, the sensitivity to increases in prices goes down as strata increase. The estimates for marginal utility of download and upload speeds are positive across all strata. Unlike in the case of price, the magnitude of these parameters does not exhibit a correlation with strata. Households dwelling in middle strata show the highest valuation for additional speeds (0.33 for stratum 4). As expected, households in all strata show a higher valuation for download speeds than for upload speeds. Households in all strata like premium channels and video on demand. Ev- erything else constant, access to premium channels has a stronger impact on utility than access to video on demand. The marginal valuation for these characteristics declines for higher strata. The value that households in strata 1 place on having access to premium channels is about 30% higher than households in strata 6.

41Note that the numbers on the y-axis are negative.

102 Finally, the estimated marginal utility of long distance is presented. The coeﬃcient on preference for extra minutes is decreasing in strata. Thus, lower strata have a stronger preference for long distance minutes than do higher strata. This is probably because long distance minutes and cellphones are substitutes and households in higher strata tend to have better cellphone plans.

103 Figure 8: Demand estimates 104 Table 29 present the estimates of consumer demand heterogeneity terms by stratum. In general, there is very little heterogeneity in the price sensitivity coefficient within strata. Lower strata, specifically strata 1 an 2, are more heterogeneous in their preferences for download speed. The coefficients on Video on Demand and Minutes are, by far, those with higher heterogeneity across strata.

Table 29: Demand estimates: Standard deviations Price Download Upload Premium HD VoD Minutes Stratum 1 0.03 1.23 3.27 1.72 1.65 2.8 6.59 0.01 0.09 2.07 0.67 0.23 0.48 1.94 Stratum 2 0.03 2.24 2.49 1.9 1.57 3.4 5.48 0.02 0.19 1.54 0.54 0.89 0.45 1.64 Stratum 3 0.05 2.26 1.11 2.17 1.49 2.83 4.89 0.03 0.06 0.97 1.41 0.35 0.42 1.67 Stratum 4 0.04 1.48 2.78 1.97 1.77 2.87 6.78 0.02 0.98 1.76 1.43 0.49 0.49 1.46 Stratum 5 0.08 1.79 1.97 1.78 1.60 2.77 6.41 0.05 0.77 1.11 2.14 0.78 0.41 1.48 Stratum 6 0.11 2.27 1.14 1.22 1.43 3.01 6.31 0.03 1.71 2.13 1.46 0.41 0.46 1.79

3.6.4. Marginal costs

With the estimates obtained above and the assumption that firms play a static Bertrand-Nash game in each strata, I proceed to obtain the marginal costs. The marginal costs are obtained from the first order conditions of firms deciding the prices of differentiated products to maximize profits. As an example, suppose a city with just 2 strata d ∈ {1, 2}. A firm sells identical products in each stratum. Let p1 and p2 be the prices the firm charges for the product in strata 1 and in strata 2 respectively. Furthermore, let p−1 and p−2 be the prices of all other goods sold in strata 1 and 2. Under price discrimination the firm’s first order conditions to maximize profits are given by

105 ∂π(p1, p2; p−1, p−2) ∂s1(p1; p−1) = s1(p1; p−1) + (p1 − mc) = 0 ∂p1 ∂p1

∂π(p1, p2; p−1, p−2) ∂s2(p2; p−2) = s2(p2; p−2) + (p2 − mc) = 0 ∂p2 ∂p2

From these first order conditions one can obtain the marginal costs. Note that after obtaining estimates for the demand, the only unknowns left in these first order conditions are the marginal costs. The demand estimates and data on prices allow us to write an estimate for the market shares. To obtain estimates for marginal cost, it suffices with plugging all the known data and solve for mc. I find that the marginal costs vary slightly by stratum even after controlling for bundle characteristics. Marginal costs are slightly higher in higher strata. This is probably due to the facts that higher strata are less densely populated, which is a big determinant of last mile cost in telecommunications.

3.7. Uniform pricing

In this section I investigate what would happen to welfare if firms do not have access to the information conveyed by the strata numbers. In particular, I aim to measure the change in consumer surplus and firm profits when, for plan sold under price discrimination, firms charge the same price to all households in a city regardless of their strata. To simulate a world in which firms charge a uniform price, I rely on the estimates for the demand obtained above. With the estimated demand parameters, I reoptimize prices under uniform pricing. To make matters more concrete, I continue with the example above, in which a firm was selling its good to two strata under price discrimination. Next, I obtain the first order conditions when the firm chooses a single price. If the

ﬁrm is forced to charge a single price to both strata, i.e p1 = p2 = p, the proﬁt

106 function looks like

π(p; p−1, p−2) = s1(p; p−1)(p − mc) + s2(p; p−2)(p − mc)

and the new FOC for the ﬁrm is

2 2 X X ∂sl(p; p−l) sl(p; p−l) + (p − mc) = 0 (21) l=1 l=1 ∂p The system of equations defined by first order conditions like 21 for all firms in a city define the new Bertrand-Nash equilibrium under uniform pricing.

3.7.1. Results

The next two subsections describe the main highlights of the new equilibrium under uniform prices. First, the effects of uniform prices on consumers are discussed. Second, I calculate the change in profits for the firms.

3.7.2. Consumers

Table 30 shows the change in consumer surplus caused by forcing firms to charge uniform prices. The first column shows the monthly average compensating variation for households in each city. Column two shows the total compensating variation by city. A key feature of these results is how heterogeneous the effects are. While uniform prices benefit average households in cities like Bogotá or Medellin, it harms households in cities like Barranquilla or Bucara- manga. Incidentally, uniform pricing appears to benefit average households in larger cities but harms average households in smaller cities. The net effect of a regime of uniform prices for the consumers of the 13 cities is under -$80,000 per year. Given that the revenue of the industry is 16 billion dollars, the net effect is negligible.

107 Table 30: Monthly compensating variation Average Total household Bogotá 0.01 -2,475.11 Medellín 0.02 3,657.18 Cali 0.02 2,003.26 Barranquilla -0.03 -1,613.70 Cartagena 0.05 1,572.35 Bucaramanga - 0.03 - 1,769.77 Cúcuta -0.02 831.22 Pereira -0.03 -28,966.67 Manizales -0.02 -1,426.16 Ibagué 0.03 493.89 Armenia -0.02 -490.94 Montería -0.02 -14,400.94 Villavicencio 0.03 1,107.91 Compensating variations in 2015 dollars. Average household

Next, I consider how the eﬀect of the ban varies across strata. Figure 9 shows the variation in consumer surplus by strata. As expected households in higher strata, except those in Pereira, beneﬁt from a uniform price policy whereas households in lower strata are harmed, almost in all cities. This is so, because under price discrimination, the former tend to pay higher prices than the latter.

108 Figure 9: Compensating Variation

Manizales Cartagena 3

Bogotá

Armenia

Ibagué

Bogotá 2 Villavicencio Manizales Medellín

Barranquilla Barranquilla Cúcuta Cartagena Manizales Armenia Cali Villavicencio Medellín Barranquilla 109 Cartagena Cali Villavicencio 1 Cali Barranquilla Bogotá Cali Ibagué Bucaramanga Bogotá Villavicencio Cúcuta Manizales Cúcuta Armenia Cartagena

Compensating Variation Cali Bucaramanga Villavicencio Medellín Cartagena Montería Cúcuta Ibagué Manizales Medellín Armenia Montería 0 Bucaramanga Bogotá Ibagué Pereira Cúcuta Bucaramanga Montería Pereira Ibagué Medellín Bucaramanga Cartagena Armenia Medellín Manizales Ibagué Bucaramanga Cúcuta Montería Villavicencio Barranquilla Pereira Armenia Barranquilla Pereira Cali Montería −1 Montería Bogotá

Pereira

1 2 3 4 5 6 Strata Because Colombia’s government has made huge eﬀorts to foster broadband adoption among poorer households in recent years, I look at the eﬀect that uniform pricing has on download speeds. I compare average download speeds, weighted by number of subscribers, under price discrimination and under uniform pricing. The average speeds poorer households (strata 1 and 2) subscribe to, in the new equilibrium, falls slightly. Although uniform pricing is neutral in terms of total consumer surplus, authorities should take into account that households may substitute faster plans for cheaper slower ones, thus undoing some of the achievements of previous policies.

Table 31: Bundle characteristics by stratum (Weighted Means) Download Speed Baseline Uniform Stratum 1 1.84 1.36 Stratum 2 2.26 2.02 Stratum 3 3.75 4.01 Stratum 4 5.23 5.58 Stratum 5 6.03 6.31 Stratum 6 7.58 7.94 Average download speeds in Mbps. Baseline refers to the predicted equilibrium with price discrimination. Uniform refers to the equilibrium in which ﬁrms charge uniform prices.

3.7.3. Firms

Aggregate revenues increase as a result of forcing the firms to charge uniform prices. Firms make almost 8 million dollars of additional revenue a year when they don’t price discriminate. In an industry with 16 billion dollars of annual revenue, that represents a 0.05% increase. However, the effects on individual firms are heterogeneous. Some firms seem to benefit from a ban on price discrimination whereas profits decrease for others. Moreover, the effects seem

110 to vary by city which would imply that the nature of competition, i.e. whether best responses exhibit symmetry or asymmetry, is not the same in all cities. The increase in aggregate profits begs the question, why would firms choose to practice price discrimination if they could make more profits under uniform prices? Corts (1998) provides a possible explanation, according to which, "firms find themselves in a prisoner’s dilemma: price discrimination is a dominant strategy that results in lower equilibrium profits for the firm". Another explanation, pertaining to the specifics of Colombia’s telecommunications industry, has to do with the fact that most of these firms started as Local Ex- change Carriers. In the past LEC were obligated to charge different prices in each strata, so the current price discrimination may be due to inertia. Finally, an important source of revenue for ISP and cable operators around the world is the rental fees on modems and top boxes.42 Perhaps, ISP and cable operators make more profits by selling subscriptions to more households under price discrimination because then they can rent modems and top boxes to those additional households.

3.8. Comments

This paper explores the effects of banning price discrimination in the context of Colombia’s wired telecommunication services. The results suggest that total consumer welfare would decrease by about $80,000 if firms have to stop practicing price discrimination. In most cities, however, a uniform price regime would result in a transfer of surplus from poor households to rich households. A collateral effect of removing the firms ability to price discriminate is that

42Centurylink charges $10 dollars a month for renting a modem on a $68 broadband subscription, which means that 15% of the revenue coming from the modem, which in addition has zero marginal cost. A low estimate for Comcast revenues coming from renting modems and topboxes is north of $675 million per quarter (https://goo.gl/iRGfwM).

111 households in lower strata substitute away from faster broadband plans toward cheaper but slower plans. Given that the Colombian government has pushed many initiatives to foster adoption of faster Internet, a policy removing strata has the potential to undo some of the progress made. The results also suggests that, as an aggregate, firms would benefit from charging uniform prices. Total profits of the industry would increase by about 8 million dollars per year, despite some individual firms making less profits. There are several explanation as to why firms are engaged in a suboptimal equilibrium. One possible explanation is that firms rank markets differently. Where one firm wants to raise prices other firm wants to lower them. Under this scenario price discrimination is enhancing competition. Banning price discrimination prevents a prisoner’s dilemma and firms reach an equilibrium with higher profits. Another reason could be inertia, as most of these firms were obligated to practice price discrimination in the past. Finally, firms may prefer to serve more households as they get revenue from renting modems and top box, and the data does not account for those revenues.

112 4. Conclusions and Future Directions

4.1. Conclusions revisited

Contrary to people’s fears the 2012 merger between Comcel and Telmex had pro-competitive effects, at least regarding broadband provision. Markets where Telmex provided broadband saw their median download speeds rose by 20% (.64Mbps) relative to markets unaffected by the merger. The increase in the median speed was the result of both, the merged and its rivals, increasing their speeds. Chapter 2 explored one of the possible reasons for the apparent pro-competitive effects of the merger. In chapter 2 I estimated a demand system for telecommunications services using quarterly data from Colombia’s telecommunications sector on subscription and product characteristics. The demand model had two key features: it allowed consumers to choose bundled and standalone goods, and recovered substitution patterns. I used the estimated demand, joint with an assumption of Bertrand competition to simulate different scenarios aiming at understanding the effects of mergers of firms producing complementary goods. In chapter 3 I explored the effects of banning price discrimination in the context of Colombia’s wired telecommunication services. Price discrimination between competing oligopolists can potentially enhance competition. In this case, forcing firms to charge a uniform prices generates a net increase in consumer surplus. The measure of removing strata, overall, hurts the poor and benefits the rich. However, the surplus gained by the rich offsets the surplus lost by the poor. A side effect of removing the firms ability to price discriminate within a city, is that low income strata substitute away from faster broadband plans toward cheaper but slower plans. Given that the Colombian

113 government has pushed many initiatives to foster adoption of faster Internet, a policy removing strata has the potential to undo some of the progress made.

4.2. Future research directions

There are a few courses in which the model presented in chapter 2 can be enhanced. First, when specifying the model, I did not allow for cost synergies between the merging parties. Cost reductions are often the motivation behind mergers and could further increase the welfare gains caused by the merger. Sec- ond, the data used only contains bundles sold by the same firm. Consumers may be doing mix-and-match themselves, and subscribing to, say, Internet and Cable TV from two different operators. For this reason, the level of complementarity between hardwired and hardwired estimated is likely to underestimate the actual one. A better assessment of complementarity can be obtained from a data set with information about mix-and-match bundling. Third, I fo- cused on price effects of the merger and totally disregard any effects on quality. Quality is likely to be determined endogenously and as such be affected by the merger. For instance, chapter 1 shows that the Claro merger unequivocally increased the average speeds in markets affected by it. Endogenizing quality will certainly make it more difficult both conceptually and computationally, but will enhance our understanding of these phenomena. Lastly, I abstracted away from all the dynamic considerations. In an industry in which dynamic pricing is a staple, and where a consumer’s history determines the price she pays, static oligopoly may be a strong assumption. A natural improvement on the work presented here would have to models the dynamics of the industry.

114 115 A. Appendix

A.1. Chapter 1

116

Figure 10: Distribution of prices under diﬀerent scenarios Table 32: Estimated substitution patterns for goods included in a given bundle

Γj Std. Error Internet/Phone 0.82 *** 0.13 Internet/TV -1.78* 1.60 Internet/Mobile 1.35 *** 0.27 Internet/Cell 2.79 ** 1.38 Phone/TV 0.92 *** 0.04 Phone/Mobile -2.11*** 0.57 Phone/Cell 1.02 *** 0.02 TV/Mobile 0.02 0.01 TV/Cell 1.73 *** 0.01 Mobile/Cell 2.57 * 1.22 Internet/Phone/TV 1.17 *** 0.26 Internet/Phone/Mobile 0.95 *** 0.13 Internet/Phone/Cell 0.47 * 0.22 Internet/TV/Mobile 1.29 *** 0.32 Internet/TV/Cell 0.33 *** 0.01 Internet/Mobile/Cell 0.09 * 0.05 Phone/TV/Mobile 0.78 1.14 Phone/TV/Cell -0.42* 0.26 Phone/Mobile/Cell 1.25 1.55 TV/Mobile/Cell 2.18 3.12 Internet/Phone/TV/Mobile 0.10 *** 0.02 Internet/Phone/TV/Cell 0.35 *** 0.12 Internet/Phone/Mobile/Cell -0.13 0.25 Internet/TV/Mobile/Cell 0.01 *** 0.01 Phone/TV/Mobile/Cell 1.78 3.611 Internet/Phone/TV/Mobile/Cell 0.22 ** 0.01

Table 33: Substitution patterns (Bundles of 2) Phone Cell TV Mobilea Internet $0.92 $3.13 -$2.00 $1.51 Phone $1.15 $1.03 -$2.36 Cell $1.94 $2.89 TV $0.02 (a) Mobile Internet refers to the use of a dongle on a computer to access Internet via GSM or LTE networks.

117 Table 34: Substitution patterns (Bundles of 3) Internet Phone TV Phone/TV $1.31 Phone/Mobile $1.07 Phone/Cell $0.53 TV/Mobile $1.45 $0.88 TV/Cell $0.37 -$0.47 Mobile/Cell $0.10 $1.40 $2.44 (a) Mobile Internet refers to the use of a dongle on a computer to access Internet via GSM or LTE networks.

Table 35: Substitution patterns (Bundles of 4 and 5) Mobile Cell Mobile/Cell Mobile/ /TV /TV /TV Cell Internet/Phone $2.11 $2.39 $2.14 -$0.24 Internet/TV $2.01 Phone/TV $2.00 (a) Mobile Internet refers to the use of a dongle on a computer to access Internet via GSM or LTE networks.

Table 36: Correlations between standalone goods Internet Phone Mobile Cable TV Cell Internet 1.00 0.51 0.22 -0.22 0.63 Phone 0.51 1.00 -0.41 0.41 0.46 Mobile 0.22 -0.41 1.00 0.34 0.74 Cable TV -0.22 0.41 0.34 1.00 0.41 Cell 0.63 0.46 0.74 0.41 1.00

118 Table 37: Compensating valuations after merging ETB and Avantel Household Total compensating Standard Error compensating variation Variation Bogotá 1.58 0.46 3,867,067 Medellín 1.17 0.22 1,091,445 Cali 1.16 0.21 696,189 Barranquilla 1.13 0.27 324,449 Cartagena 1.19 0.23 251,622 Bucaramanga 1.17 0.25 147,021 Cúcuta 1.18 0.23 167,042 Pereira 1.15 0.22 103,117 Manizales 1.13 0.22 96,789 Ibagué 1.16 0.23 135,097 Armenia 1.15 0.21 85,729 Montería 1.15 0.26 82,854 Villavicencio 1.32 0.45 119,148 The compensating variation is given in 2015 dollars. Bootstrap standard errors.

Table 38: Compensating valuations after splitting up Claro Household Total compensating Standard Error compensating variation Variation Bogotá 2.31 0.71 5,653,750 Medellín 1.92 0.48 1,791,089 Cali 1.93 0.42 1,158,314 Barranquilla 1.84 0.5 528,306 Cartagena 1.93 0.45 408,093 Bucaramanga 1.95 0.47 245,035 Cúcuta 1.97 0.46 278,875 Pereira 1.93 0.47 173,057 Manizales 1.90 0.45 162,743 Ibagué 1.89 0.5 220,115 Armenia 1.86 0.42 138,657 Montería 1.94 0.48 139,771 Villavicencio 2.03 0.51 183,235 The compensating variation is given in 2015 dollars. Bootstrap standard errors.

119 Table 40: Compensating variations S1 S2 S3 S4 S5 S6 Bogotá -0.24 0.00 0.61 0.55 0.67 0.71 Medellín 0.01 -0.10 -0.09 0.20 0.39 0.62 Cali -0.34 0.15 0.40 0.39 0.33 0.63 Barranquilla -0.36 -0.39 0.42 0.70 0.62 0.59 Cartagena -0.04 0.02 0.54 0.57 0.56 0.81 Bucaramanga -0.17 -0.09 -0.07 0.05 0.1 0.35 Cúcuta -0.17 -0.14 0.02 0.37 0.23 0.37 Pereira -0.69 -0.76 -0.38 -0.33 -0.03 -0.06 Manizales -0.13 0.00 0.57 0.58 0.63 0.67 Ibagué -0.22 -0.14 -0.14 0.16 0.25 0.49 Armenia -0.17 -0.09 -0.03 0.30 0.51 0.66 Montería -0.41 -0.50 -0.28 -0.10 0.05 0.01 Villavicencio -0.15 0.37 0.52 0.58 0.67 0.58

A.2. Chapter 3

A.2.1. Demand estimates

Table 39: Demand estimates (Means) Price Download Upload Premium HD VoD Minutes Stratum 1 -0.92 0.32 0.07 3.12 1.07 2.01 0.05 0.08 0.09 0.07 1.20 0.03 1.03 0.01 Stratum 2 -0.84 0.31 0.09 2.67 1.02 1.97 0.05 0.05 0.07 0.07 1.17 0.05 1.07 0.02 Stratum 3 -0.78 0.31 0.12 2.09 0.55 1.04 0.03 0.06 0.09 0.08 1.15 0.06 1.03 0.01 Stratum 4 -0.73 0.33 0.09 2.12 1.01 1.06 0.02 0.15 0.11 0.05 1.13 0.09 1.02 0.01 Stratum 5 -0.72 0.27 0.10 2.11 1.23 1.02 0.02 0.23 0.13 0.06 1.22 0.11 0.97 0.01 Stratum 6 -0.71 0.30 0.10 2.13 1.08 1.01 0.02 0.52 0.22 0.04 1.46 0.34 0.84 0.02

120 A.2.2. Counterfactual prices and consumer surplus

Table 41: Counterfactual prices Uniform Prices Change Average Minimum Maximum Average Minimum Maximum Stratum 1 19.85 9.83 47.81 1.21 0.01 5.55 Stratum 2 20.14 9.74 55.93 1.73 -0.262 8.62 Stratum 3 20.06 9.06 52.69 0.87 -3.06 8.06 Stratum 4 18.953 8.83 46.29 -0.97 -3.65 0.06 Stratum 5 19.15 8.48 75.76 -1.87 -6.85 0.42 Stratum 6 21.11 7.52 70.80 -3.14 -13.93 0.67 Uniform Prices are the prices of a simulated equilibrium in which ﬁrms don’t exert price discrimination across strata within a city. Change refers to the diﬀerence between the simulated prices under uniform rices and the simulated prices under price discrimination.

121 Figure 11: Price histogram

Histogram of PricesMc$price 6000 5000 122 4000 Frequency 3000 2000 1000 0

0 20 40 60 80

PricesMc$price

Blue: Baseline prices. Pink: Uniform prices. References

Aarnikoivu, S. and Winter, J. (2006). Mobile broadband wireless access. In- novation in Telecommunications, page 65.

Anderson, S. P., Loertscher, S., and Schneider, Y. (2010). The abc of complementary products mergers. Economics Letters, 106(3):212–215.

Andersson, K., Foros, Ø., and Steen, F. (2009). Text and voice: complements, substitutes or both? Industrial and Corporate Change, 18(6):1231–1247.

Armstrong, M. (2010). Recent developments in the economics of price discrimination. In Blundell, R., Newey, W. K., and Persson, T., editors, Advances in economics and econometrics: Theory and applications, ninth world congress, Volume III. Cambridge University Press.

Armstrong, M. and Vickers, J. (2001). Competitive price discrimination. rand Journal of economics, pages 579–605.

Banerjee, A. and Ros, A. J. (2004). Patterns in global ﬁxed and mobile telecommunications development: a cluster analysis. Telecommunications Policy, 28(2):107–132.

Berry, S., Levinsohn, J., and Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica: Journal of the Econometric Society, pages 841– 890.

Berry, S. T. (1994). Estimating discrete-choice models of product diﬀerentia- tion. The RAND Journal of Economics, pages 242–262.

Chen, Y. (2001). On vertical mergers and their competitive eﬀects. RAND Journal of Economics, pages 667–685.

123 Choi, J. P. (2008). Mergers with bundling in complementary markets. The Journal of Industrial Economics, 56(3):553–577.

Corts, K. S. (1998). Third-degree price discrimination in oligopoly: all-out competition and strategic commitment. The RAND Journal of Economics, pages 306–323. de Mercados, R. (2015). Mercados Audivisuales en un Entorno Convergente. Technical report, Comision de Regulacion de Comunicaciones-CRC.

Dobson, P. W. and Waterson, M. (2005). Chain-store pricing across local markets. Journal of Economics & Management Strategy, 14(1):93–119.

Gary-Bobo, R. J. and Larribeau, S. (2004). A structural econometric model of price discrimination in the french mortgage lending industry. International Journal of Industrial Organization, 22(1):101–134.

Gentzkow, M. (2007). Valuing new goods in a model with complementarity: Online newspapers. The American Economic Review, 97(3):713–744.

Grennan, M. (2013). Price discrimination and bargaining: Empirical evidence from medical devices. The American Economic Review, 103(1):145–177.

Grzybowski, L., Nitsche, R., Verboven, F., and Wiethaus, L. (2014). Market deﬁnition for broadband internet in slovakia–are ﬁxed and mobile technolo- gies in the same market? Information Economics and Policy, 28:39–56.

Grzybowski, L. and Pereira, P. (2007). Merger simulation in mobile telephony in portugal. Review of Industrial Organization, 31(3):205–220.

Hausman, J. A., Sidak, J. G., and Singer, H. J. (2001). Residential demand for broadband telecommunications and consumer access to unaﬃliated internet content providers. Yale J. on Reg., 18:129.

124 Howell, B. (2012). Competition and regulatory implications of a vodafone- telstraclear merger: ﬁrst thoughts.

Leslie, P. (2004). Price discrimination in broadway theater. RAND Journal of Economics, pages 520–541.

Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica, 69(2):307–342.

Ordover, J. A., Saloner, G., and Salop, S. C. (1990). Equilibrium vertical foreclosure. The American Economic Review, pages 127–142.

Perry, M. K. and Porter, R. H. (1985). Oligopoly and the incentive for horizontal merger. The American Economic Review, 75(1):219–227.

Prieger, J. E. (2015). Multimarket contact and strategic entry decisions. In- ternational Journal of Business Environment, 7(4):396–414.

Ribeiro, R. and Vareda, J. (2010). Crowding out or complementarity in the telecommunications market. Economics Letters, 106(3):212–215.

Salant, S. W., Switzer, S., and Reynolds, R. J. (1983). Losses from horizontal merger: the eﬀects of an exogenous change in industry structure on cournot- nash equilibrium. The Quarterly Journal of Economics, pages 185–199.

Salinger, M. A. (1988). Vertical mergers and market foreclosure. The Quarterly Journal of Economics, pages 345–356.

Srinuan, C., Srinuan, P., and Bohlin, E. (2010). The Mobile Broadband and Fixed Broadband Battle in Swedish market: Exploring complementary or substitution. Technical report.

125 Villas-Boas, S. B. (2009). An empirical investigation of the welfare eﬀects of banning wholesale price discrimination. The RAND Journal of Economics, 40(1):20–46.

126