How Proximity to Transit and Walkability Influence Business Creation and Performance

ECONOMIC DEVELOPMENT FOR THE 21st CENTURY: HOW PROXIMITY TO TRANSIT AND WALKABILITY INFLUENCE BUSINESS CREATION AND PERFORMANCE

Kevin N. Credit

A DISSERTATION

Submitted to Michigan State University in partial fulfillment of the requirements for the degree of

Geography – Doctor of Philosophy

2018

ABSTRACT

ECONOMIC DEVELOPMENT FOR THE 21st CENTURY: HOW PROXIMITY TO TRANSIT AND WALKABILITY INFLUENCE BUSINESS CREATION AND PERFORMANCE

Kevin N. Credit

Interest in rail transit and walkable built environments has increased in recent years as

North American cities have invested large sums on these interventions and scholars have undertaken a vast research program to examine their impact on travel behavior, development, and property values. The economic impact of these “active” transportation interventions is particularly important to understand, given their large fixed cost and the desire of policy-makers to use them as economic development and urban revitalization tools. At the same time, new business creation is an important engine of sustainable local development, necessitating a more detailed understanding of the direct links between infrastructure investment and new business creation.

To better understand this process, this dissertation examines the economic impact of rail transit systems and walkable built environments using several quantitative approaches, including quasi-experimental time series regression, spatial econometric approaches, and hierarchical linear modeling. The first chapter evaluates the impact of Phoenix’s light rail system, which opened in 2008, on new firm formation in specific industries. Findings show that the transit adjacency is worth an 88% increase in knowledge sector new starts, a 40% increase in service sector new starts, and a 28% increase in retail new starts at the time the system opened, when compared with automobile-accessible control areas. However, the light rail also appears to suffer from a ‘novelty factor’ – after the initial increase in new establishment activity in adjacent block

groups, the effect diminishes at the rate of 8%, 6%, and 7% per year, respectively. The results also provide insight into the spatial extent of light rail impacts to new business formation, with areas 1 mile from stations observing 21% fewer retail new business starts and 12% fewer knowledge sector new starts than areas within ¼-mile of stations.

The second chapter evaluates the relationship between transit station proximity and new business creation in five regions with varying levels of maturity in rail transit development and/or entrepreneurial ecosystems – Boston, San Jose, Austin, Cleveland, and Philadelphia. It tests a variety of spatial econometric models to find the best specification and compares the results to the kinds of non-spatial models currently used in the literature. This provides a better understanding of the role of various forms of spatial dependence in the transit—new business creation relationship and shows that existing models may overstate the impact of transit on new business creation. In addition, the paper teases out differences between regions, rail modes, and business types that can be usefully applied to a variety of urban contexts.

Finally, the third chapter examines the importance of place-making in economic development by evaluating the relationship between specific urban design features – based on

Jacobs’ “four generators of diversity” (1961) and Ewing and Cervero’s “Five-D’s” (2010) – and business sales volume using a hierarchical linear modeling (HLM) framework in Phoenix and

Boston. The results indicate that specific features of walkable built environments are positively associated with business performance. However, the relationship between walkable built environments and business performance varies considerably depending on the type of business and city-level context being studied, indicating that significant nuance must be used when considering place-based economic interventions.

To my wife, Meghan – who was my partner through all the long hours, the many moves, and the last-minute freak-outs – and to my parents, Peter and Martha, who have always believed in me.

ACKNOWLEDGEMENTS

Thanks to Elizabeth Mack for providing considerable assistance with helping to refine the analysis, provide suggestions and support, and revise numerous drafts of this dissertation.

Thanks also to the other members of my dissertation committee – Ashton Shortridge, Zenia

Kotval, and Igor Vojnovic – for helping to guide the scope, methods, and generalizability of this work. The blind peer reviewers who made comments and suggestions for each of these papers also helped to make the final versions considerably better than the initial drafts. Thanks as well to Julia Koschinsky and Emily Talen for sharing their unique block group dataset, and thanks also to the Ewing Marion Kauffman Foundation (Grant #20130782) and the MSU Graduate

School Dissertation Completion fellowship for supporting this research. And, finally, thanks to

Lisa-Marie Pierre, Jonah White, and Kelsey Nyland for helping to edit revisions and provide me with a sounding board for new ideas.

TABLE OF CONTENTS

LIST OF TABLES ...... ix

LIST OF FIGURES ...... x

INTRODUCTION ...... 1 Economic Development in the 21st Century ...... 1 Research Context: Transportation and Economic Development ...... 5 Transportation in Location Theory ...... 6 Economic Impacts of Transit ...... 8 Economic Impacts of Walkable Built Environments...... 10 Under-Explored Issues: Agglomeration, Spatio-Temporal Relationships, and Methods ...... 10 Research Goals...... 19

CHAPTER 1. TRANSIT-ORIENTED ECONOMIC DEVELOPMENT: THE IMPACT OF LIGHT RAIL ON NEW BUSINESS STARTS IN THE PHOENIX, AZ REGION...... 23 Introduction ...... 23 Theoretical Context ...... 25 Conceptual Framework ...... 27 Study Area and Data ...... 31 Methods...... 33 Treatment and Control Area Selection ...... 34 Specification of the AITS Model ...... 36 Results ...... 38 Socioeconomic Context of Light Rail Station Locations ...... 38 Descriptive Statistics ...... 41 AITS Model Results ...... 42 Discussion and Conclusion ...... 46

CHAPTER 2. TRANSITIVE PROPERTIES: A SPATIAL ECONOMETRIC ANALYSIS OF NEW BUSINESS CREATION AROUND TRANSIT...... 51 Introduction ...... 51 Literature ...... 53 Data and Study Areas...... 57 Data ...... 57 Study Areas ...... 60 Methods...... 61 Modeling Spatial Count Data...... 64 Exposure Variables for New Business Creation ...... 67 Spatial Regression Specification ...... 68 Results ...... 76 Descriptive Statistics and Spatial Patterns ...... 76 Regression Results ...... 76 Conclusions and Discussion ...... 83

vii

CHAPTER 3. PLACE-MAKING AND PERFORMANCE: THE IMPACT OF WALKABLE BUILT ENVIRONMENTS ON BUSINESS PERFORMANCE IN PHOENIX AND BOSTON...... 86 Introduction ...... 86 Perspectives on Good Urban Design ...... 88 Benefits of Walkable Urban Design ...... 89 Economic Value of Walkable Urban Design ...... 90 Conceptual Framework ...... 90 Study Area ...... 93 Data ...... 94 Business Data ...... 94 Tract-Level Data ...... 96 Methodology ...... 99 Null Model ...... 100 Random Coefficients Model ...... 102 Random Slopes Model ...... 103 Results ...... 104 Model Results ...... 107 Discussion and Conclusion ...... 111

CONCLUSION ...... 116 Limitations ...... 116 Physical Determinism ...... 117 Situation Within Cultural, Political, and Economic Contexts .... 119 NETS Data Limitations...... 124 Generalizability and Implications for Theory and Practice ...... 126 Avenues for Future Research ...... 130 Summary ...... 134

APPENDICES ...... 136 APPENDIX A: Chapter 1 Supplementary Materials ...... 137 APPENDIX B: Chapter 2 Supplementary Materials ...... 144 APPENDIX C: Chapter 3 Supplementary Materials ...... 152

REFERENCES ...... 154

viii

LIST OF TABLES

Table 1. Theorized tendencies for transportation modes to produce given agglomerative economies, and the business types most likely to benefit...... 17

Table 2. Poisson AITS model results and incidence rate ratios (IRR) for three dependent variables of interest...... 43

Table 3. Data sources and descriptions...... 58

Table 4. Rate-smoothing methods and exposure variables considered for testing. The four “best” method/variable combinations selected for comparison (for each of the four dependent variables of interest) are marked in bold...... 69

Table 5. Comparison of model results for the four “best” method/variable combinations for each of the four dependent variables of interest...... 71

Table 6. Table summarizing pooled and region-specific Spatial Durbin model significant total effects for mode and business type...... 78

Table 7. Table summarizing differences between Spatial Durbin and Poisson model results, i.e., Type I and Type II Errors...... 82

Table 8. Activity-, form-, and image-based components of place-making...... 93

Table 9. List of variables considered for use in models...... 95

Table 10. Intraclass correlation coefficients for Phoenix and Boston...... 101

Table 11. Additional variance explained by random slopes model for Phoenix and Boston. ... 108

Table 12. Final model results for Phoenix and Boston...... 110

Table 13. Summary of model results...... 113

LIST OF FIGURES

Figure 1. Diagram showing theoretical connections between transportation investment and economic performance and value creation...... 12

Figure 2. Diagram showing hierarchy of different transportation mode types...... 16

Figure 3. Conceptual framework showing how transit infrastructure influences individual business location decisions and clustering...... 29

Figure 4. Map showing six treatment and control areas considered for analysis...... 35

Figure 5. Map showing SES index score (in 2000) distributions for Maricopa County and light rail-adjacent block groups...... 40

Figure 6. Map showing percent change incidence rate ratios for DPOSTIMP variable for varying definitions of the treatment boundary – ¼, ½, and 1 mile – in central Phoenix...... 46

Figure 7. Analysis framework for testing rate smoothing and spatial econometric approaches. 64

Figure 8. Comparison of AIC for each tested pooled model type (for each business type dependent variable). Lowest AIC shown in red circle...... 75

Figure 9. Census blocks in top 25% for Log10(expected new knowledge business density) and rail transit buffers for all modes and regions...... 80

Figure 10. Conceptual framework for components of urban place-making (Montgomery, 1998)...... 91

Figure 11. Maps showing spatial ‘hot spots’ and distribution of average sales volume per employee by tract in Phoenix and Boston...... 106

Figure 12. City of Phoenix Zoning Map from December 29, 2017 (Phoenix 2017a)...... 120

Figure 13. New Knowledge, Retail, and Services Business Creation in TOD Zoning Overlay Districts in the Phoenix region...... 122

INTRODUCTION

Economic Development in the 21st Century

Since the mid-1970s, the global economy has largely stopped being driven by vertically- integrated, monolithic manufacturing firms and the associated political institutions and cultural consumption norms that supported them, known as Fordism (Harvey, 1989; Jessop, 1992; Amin,

1994; Freeman & Perez, 1988; Piore & Sabel, 1984). Instead, the economy today is characterized primarily by innovation, speed, and flexibility, which has led to a greater focus on entrepreneurship, creativity, diversity, and flexible specialization – both for individual firms and for regions (Sabel, 1994; Florida, 2002; Florida, 2004; Storper & Scott, 2009; Wennekers &

Thurik, 1999; Birch, 1987; Acs & Audretsch, 1990; Acs & Audretsch, 2003). In addition to the changing structure of the economy, post-Fordism has brought with it new political norms (fiscal austerity and neoliberalism), sped-up business cycle fluctuations, and all of the destabilizing effects of globalization, like increased out-sourcing, shifting production centers, and ‘foot-loose’ business location decision-making (Harvey, 1989; Freeman & Perez, 1988; Amin, 1994).

While this change to a globalized, post-industrial economy began over forty years ago

(Harvey, 1989), many cities are still struggling to make the transition. Much of the difficulty has to do with costly fixed infrastructure investments, such as freeways and suburban industrial parks, which cannot be easily transitioned to match changes in social and economic behavior

(Harvey, 1989), as well as the defunding of many federal programs for economic development

(Eisinger, 1988). This neoliberal turn in public finance has also resulted in calls for more fiscal discipline, smaller budgets, and lower taxes, all of which fuel the local politics of the “growth machine”, resulting in patterns of investment and urban policies that primarily benefit and

1 protect wealthy residents at the expense of social services and reinvestment in low income neighborhoods (Molotch, 1976). In this way, the continuing reduction of funding for local governments has created a fiscal crisis that is ‘solved’ by focusing on unsustainable (often suburban) growth models that are fueled primarily by exogenous new investment.

Exacerbating this fiscal crisis, most economic development policies used by cities are stuck in the past tradition of “smokestack chasing”, still focused on the recruitment of large manufacturing sector base employers (Cobb, 1993; Eisinger, 1988; Graham & Marvin, 2001;

Leigh & Blakely, 2013; Malecki, 1984). Unfortunately, the widespread use of incentive-oriented economic development policies, like tax abatements, is a zero-sum game in which resources are shuffled from one location to another – an unsustainable solution for local economic development. Many local governments ultimately lose out on their bid to attract a footloose manufacturing plant, wasting resources that could have been applied elsewhere; or, perhaps worse, they win the bid and commit vast sums of tax dollars to a company that will ultimately export most of the revenue to locations outside of the local economy (to executives in the headquarters, investors, or financial institutions), making it difficult for the local government to ever truly recoup its economic development investment

Given the changing nature of economic activity, public finance, and urban development that has occurred over the past forty years, cities need to exploit more sustainable economic development approaches in order to remain economically viable. Entrepreneurship provides a sustainable path to regional growth and innovation in the post-industrial economy, for several reasons (Wennekers & Thurik, 1999; Carree & Thurk, 2003). The ‘flexible specialization’ of post-Fordism has placed small businesses at an economic advantage relative to large businesses because they are more able to adapt to change (Harvey, 1989). At the same time, new business

2 creation strategies are generally less expense than incentive packages targeted at large existing companies and have the benefit of generating ‘home-grown’ economic activity that circulates revenue within the local economy rather than exporting it to another region (White & Kotval,

2012). Entrepreneurship initiatives also have the potential to benefit traditionally disadvantaged populations, since their focus in some ways circumvents the traditional institutionalization of wealth by providing opportunities for residents to become business owners that may not have had the opportunity to work their way up through the corporate or educational hierarchy

(Blackburn & Ram, 2006).

In addition, new businesses are often the physical products of innovation as a result of spinoff activity from parent companies or large research institutions, and thus are an important factor linking new knowledge generation to regional economic growth (Acs & Audretsch, 2003;

Wennekers & Thurik, 1999; Carree & Thurik, 2003; Klepper, 2009; Klepper & Sleeper, 2005).

Regions that have the right mix of cultural, financial, political, human capital, and financial components can foster “entrepreneurial ecosystems,” capturing the home-grown spillover effects from innovation (Mack & Mayer, 2015; Stam, 2015). This leads not only to economic success for individual firms (and the regional benefits those bring), but also to generation of a self- sustaining supportive environment for new business formation.

New business creation is also an important factor in the generation of specialized competitive clusters (Teece, 2007; Asheim, 1996; Florida, 1995; Porter, 2000; Malmberg &

Maskell, 2002). Clusters are formed and augmented initially through new business activity.

Clusters that develop the dynamic of “competition and community”, discovered by Saxenian

(1994) in Silicon Valley, drive demand for better products both vertically and horizontally, which leads to higher levels of innovation and new business creation within the cluster, and

3 eventually increased regional competitiveness and economic growth (Porter, 2000; Rutten &

Boekema, 2007; Malmberg & Maskell, 2002). Mature clusters also then (iteratively) provide a set of attractive features for new business creation, and have been shown to increase the rate of new firm formation and survival (Delgado, Porter, & Stern, 2010).Entrepreneurship in the high technology sector, in particular, has been the emphasis of much economic development research, given the fact that high technology industries have the potential to generate regional economic growth (Leigh & Blakely, 2013; Malecki, 1984; Malecki, 1991; DeVol & Wong, 1999; Chapple et al. 2004).

Regional economic diversity (in addition to specialization) is equally reliant on new business formation (Jacobs, 1967). As industries mature, large vertically-integrated firms emerge to take advantage of economies of scale. This increases efficiency but also puts the region that relies on those firms at the risk of “lock-in”, where the ability to innovate is lost as businesses and institutions become overly reliant on the dominant industry (Boschma & Lambooy, 1999;

Boschma & Frenken, 2006). New business creation is vital to diversifying the economic base of a region, because new businesses provide competition and innovation within existing industries.

This spurs the industrial evolution and creativity that ultimately contribute to the long-term sustainability of the economy: higher levels of entrepreneurship make it more likely that a given region invents the new technology that changes the structure of the economy, rather than lagging behind it (Schumpeter, 1934; Jacobs, 1967; Saviotti & Pyka, 2004; Frenken, Van Oort, &

Verburg, 2007). In addition to this benefit through related variety, entrepreneurship in unrelated sectors also makes a region more resilient to macro-scale shocks by spreading risk among regional investments in different kinds of industries, known as the “portfolio” approach to regional diversity (Montgomery, 1994; Frenken, Van Oort, & Verburg, 2007).

Entrepreneurship also creates the kind of mixed land use diversity that promotes urban vitality and activates urban public space (Jacobs, 1961). New (often small) businesses fill old buildings and can revitalize urban places by offering a greater mix of activities within an area, generating street traffic, increasing physical activity, and giving people options for community venues where they can socially interact, i.e., “third places” (Jacobs, 1961; Gehl, 2010;

Innovation District, 2015; McCormack, Giles-Corti, & Bulsara, 2008; Krizek, 2003a; Ewing &

Cervero, 2010; Oldenburg, 1999). A greater variety of businesses also adds to a region’s unique

‘sense of place’ by offering non-standardized environments (generated by the minds of many individual actors) and catering directly to unique local sensibilities (Jacobs, 1961; Relph, 1976;

Robertson, 1999; Kunstler, 1994; Walljasper, 2007; Alexander, 1977). This emphasis on new businesses also has the potential to create a positive economic development feedback loop in cities, where entrepreneurship leads to unique places with a diversity of land uses, which in turn attract greater numbers of entrepreneurs.

Research Context: Transportation and Economic Development

At the same time that interest in more sustainable economic development activities (such as supporting homegrown entrepreneurship) has begun to increase, “active” transportation infrastructure systems such as rail transit systems and walkable built environments have become the focus of much recent urban sustainability discourse (Frank et al. 2007; Besser & Dannenberg,

2005; Edwards, 2008; Lindstrom, 2008), which has led to massive direct investment from cities.

As of 2014, over 700 local municipalities have enacted “Complete Streets” policies to guide walkable transportation design, while hundreds of miles of light rail lines in 16 new regional systems have also been completed, at an expense of over 25 billion dollars (Smart Growth

America, 2014; Freemark, 2014). 5

Given the attention and public expenditure directed at these investments – and the possibility they offer for increasing environmental and social sustainability (Calthorpe, 1993;

Dittmar & Ohland, 2004; Curtis, Renne, & Bertolini, 2009; Newman & Kenworthy, 2013;

Camagni, Gibelli, & Rigamonti, 2002; Garrett & Taylor, 1999; Holzer, 1991; Kawabata & Shen,

2007) – the purpose of this dissertation is to investigate the economic development impacts of these active transportation systems. This helps to provide not only a sense of the (indirect) return on transportation investment, but also to assess the feasibility of using active transportation investments as sustainable economic development and revitalization tools.

The economic impact of transportation systems – and the dynamic interplay between urban development, economic development, and transportation accessibility – has been widely studied in economic geography, regional science, and urban planning. To contextualize the theoretical and empirical context for this dissertation, and the gaps that it fills, the theoretical basis for understanding the role of transportation in fostering economic activity must be examined in some detail. The next three sections illuminate the chronological development of this literature, first as it relates to transportation systems in general, and then more specifically as it relates to the active transportation features of interest in this dissertation. Then, the final section sets out a novel theoretical framework for understanding how specific transportation modes influence new business creation in specific sectors.

Transportation in Location Theory

At the most basic level, it is impossible to make or sell goods without some connection to inputs or customers. For this reason, it has long been acknowledged that transportation systems are important factors for economic and urban development (Hoyt, 1939; Harris & Ullman, 1945;

Jacobs, 1967). Transportation costs play a central role in classical location theory, which

6 attempts to explain the distribution and spatial pattern of various economic activities based on purely economic considerations. In von Thünen’s original theory of agricultural land use allocation, transportation cost (per acre-product) for various goods determines the concentric distribution of land uses; heavier or bulkier products will be grown closest to the market (such as milk), while lighter products or those that require a large areal extent (like wheat) will be located furthest from the market (Beckmann, 1968). Alonso (1960) extends this concept to urban land uses, positing that uses or residents that require a shorter commute – such as commercial businesses that require face-to-face contact or low-income households with limited transportation budgets – will locate closer to the center, while uses and residents that require more land (and can afford a longer commute) will locate further from the center. For both von Thünen and

Alonso, the nature of real transportation networks found in the world can skew the idealized concentric ring pattern, because locations near transportation networks have increased accessibility to the center.

Weber’s (1929) theory of industrial location is also based primarily on the transportation costs of industrial inputs and outputs. For Weber, the optimal location for a manufacturing plant minimizes the cost-distance from each of its input sources and final market. Since cost-distances are determined primarily by the weight of the inputs and the final product, the plant will tend to be located closer to its heavier inputs if the final product weighs less than the sum of its inputs

(e.g., steel), and located closer to the market if the final product weighs more than the sum of its inputs, or if one of its inputs is ubiquitous (e.g., beer). Of course, the location of specific transportation networks alter the distance calculations from the idealized “triangle” depicted in

Weber’s work, since goods cannot be transported in straight lines directly from point to point.

While Weber is concerned with the pattern of producers, Central Place Theorists Lösch

(1967) and Christaller (1966) examine the pattern of demand for various products in the form of market areas. The size of a market for a given product is determined by two related transportation factors: the threshold – the number of potential consumers within a given distance from the retailer (or producer) – and the range, which is the distance a consumer is willing to travel to purchase a given product. For Lösch, new retailers will emerge to fill the gaps in market area until the space is exhausted by a regular system of market areas, which are optimally hexagonal in shape. Since each product (and thus business) has different range/threshold characteristics, there is a different pattern of hexagons for each type of product; the market areas for smaller range businesses (such as gas stations) nest within those of larger range businesses

(such as designer clothing stores). The places where every type of good is sold are most

“central”, also and the biggest settlements. Beyond explaining the distribution of business types and cities, Central Place Theory also develops a theory for understanding why transportation networks are built where they are: to connect settlements at the same level in the central place hierarchy.

These core theories – and their numerous empirical and theoretical extensions – show that transportation is intimately related to the location of economic activity and urban development, and also provide the foundational frameworks for thinking about where businesses locate and why (Beckmann, 1968).

Economic Impacts of Transit

While classical location theorists are interested in explaining the spatial distribution of economic activity – in which transportation accessibility plays a key role – they do not explicitly

8 evaluate the economic development impacts of transportation systems. Empirical research applying these theoretical tenets of location theory, i.e., that transportation investments increase surrounding accessibility and thus property values, to transit systems in specific regions began in earnest in the 1970s and 1980s, (Knight & Trygg, 1977; Cervero, 1984; Cervero, 1994). This work generally confirms the idea that station proximity increases property values and thus development intensity, primarily when the service itself is frequent and reliable, when the system is planned in conjunction with supportive land use policies, and is built in areas with high existing economic growth and development potential (Knight & Trygg, 1977; Cervero, 1984;

Cervero, 1994; Damm et al. 1980; Landis, Guhathakurta, & Zhang, 1994; Knaap, Ding, &

Hopkins, 2001; Cervero, 2004; Agostnini & Palmucci, 2008; Weinberger, 2001; Weinstein &

Clower, 2003; Golub, Guhathakurta, & Sollapuram, 2012). While property value premiums are higher for commercial land uses and commuter rail lines (Mohammad et al. 2013), light rail systems have also been shown to confer property value benefits (Weinberger, 2001; Weinstein &

Clower, 2003; Golub, Guhathakurta, & Sollapuram, 2012).

However, some call this relationship into question, positing instead that the economic benefits observed due to transit systems are simply a re-focusing of development that would have occurred somewhere else (e.g., the suburbs), or that it is predominantly due to regional economic conditions (Giuliano, 2004; Vessali, 1996; Schuetz, 2014). Others have also argued that transit systems are not the primary driver of increasing walkability, density, or business agglomeration

(Quinn, 2006; Bollinger & Ihlanfeldt, 1997; Chatman, 2013).

Despite this division in the literature, a recent meta-analysis has shown that transit generally confers property value benefits to adjacent areas (Mohammad et al. 2013). Research that evaluates the association between transit proximity and other economic variables of interest,

9 such as new business creation, also confirms this positive association (Mejia-Dorantes, Paez, &

Vassallo, 2012; Song et al. 2012; Chatman, Noland, & Klein, 2016). However, given the small number of studies evaluating these economic benefits (e.g., new business creation) additional research is certainly needed.

Economic Impacts of Walkable Built Environments

As shown in the section above, there is an extensive existing literature on the economic impacts of transit. Empirical research on the economic impacts of other active transportation systems – specifically, walkable built environments – is much less common, because most of research on walkable built environments focuses on their propensity for increasing walking travel behavior (Ewing & Cervero, 2010; Cervero & Kockelman, 1997; Boarnet & Crane, 2001;

Zhang, 2004). Several studies evaluate the association between walkable built environments and property values, generally finding premiums in walkable neighborhoods (Leinberger & Alfonzo,

2012; Pivo & Fisher, 2011; Li et al. 2014). However, research that looks at other economic benefits such as employment, new business creation, or business performance, has appeared to this point mostly in the “gray” literature as municipal white papers or other reports for specific localities (Ozdil, 2006; NYCDOT, 2013; Hass-Klau, 1993). Certainly, these relationships require a more rigorous analysis so that planners and researchers can better understand the economic impact of walkable built environments.

Under-Explored Issues: Agglomeration, Spatio-Temporal Relationships, and Methods

As this discussion demonstrates, while transportation investments have always played a central role in economic theories of city formation (Hoyt, 1939; Harris & Ullman, 1945; Jacobs,

1967), much of the literature on the economic impacts of active transportation investments

10 focuses on property value premiums in walkable and transit-adjacent neighborhoods (Leinberger

& Alfonzo, 2012; Cervero & Landis, 1997; Golub, Guhathakurta, & Sollapuram, 2012; Pivo &

Fisher, 2011; Knaap, Ding, & Hopkins, 2001). While property value increases are important indicators of economic value, and relevant to municipal governments interested in raising property tax revenues, it stands to reason that property value increases are an economic byproduct of the true mechanism through which transportation increases economic growth: agglomeration economies (Marshall, 1890; Hoover, 1937; Jacobs, 1967). While all transportation systems create some level of agglomerative potential, different modes have unique features that are more or less conducive to specific agglomeration benefits, and different kinds of businesses also respond differently to various agglomeration benefits (Chatman & Noland, 2011).

Figure 1 shows the theorized agglomeration mechanisms through which transportation investments increase economic productivity, performance, and, eventually, property values. The direct impact of a new transportation investment is an increase in accessibility for parcels of land adjacent to it (Hanson & Guiliano, 2004). Immediately, that land becomes more connected to the rest of the region and to other regions, and the cost of moving goods, services, people, and information decreases. Of course, the degree or type of accessibility increase is determined by a multitude of factors specific to the mode: airports only provide additional access to other regions, while transit stations must be supplemented by well-designed pedestrian environments in order to reach their full potential for use (Newman & Kenworthy, 2013). These nuances of accessibility are one important factor in determining how different kinds of transportation investments impact different kinds of business development.

The other important factor to business development, productivity, and performance is the nature of the specific agglomeration mechanisms provided by the increase in accessibility. As

Figure 1 shows, there are six primary agglomeration benefits facilitated through accessibility increases: market access, social networks, freight access, labor market access, information, and the potential for complimentary spillover development (Marshall, 1890; Hoover, 1937; Jacobs,

1967; Gordon & McCann, 2000; McCann & Sheppard, 2003; Porter, 2000; Chatman & Noland,

2011).

Figure 1. Diagram showing theoretical connections between transportation investment and economic performance and value creation.

It is important to note that the accessibility increase only creates an agglomerative potential in surrounding land parcels by better connecting them to other parts of the region (or other regions); additional factors like specific site, situation, distance to the central business district (CBD), land use regulations and plans, specific nature of the business, and many other factors will of course also influence the degree to which agglomeration benefits actually accrue

12 to a piece of land. The specific factors of the site, region, and business will also influence whether the benefits are primarily classified as localization or urbanization economies – while the six factors identified here could certainly all be considered urbanization benefits, they also have the potential to increase localization benefits as well (e.g., by better connecting specialized industrial suppliers with producers in the region, decrease input costs).

In any case, accessibility increases from transportation investments have the potential to increase the visibility of a business, which is especially important to retail and service firms who rely on walk-in or pass-through traffic, as well as access to customers generally by better connecting the parcel to the entire urban market (or other markets). Transportation also fosters social network creation and the maintenance of “weak ties” by lowering the cost of face-to-face communication (Gordon & McCann, 2000; McCann & Sheppard, 2003; Storper & Scott, 2009).

This is particularly important for economic development, because the social connections between producers, suppliers, customers, and support organizations have been shown to be a key component for fostering new business creation and the development of entrepreneurial ecosystems (Spigel, 2015; Mack & Mayer, 2015), as well as for developing a unique culture conducive of competition and cooperation that increases regional competitiveness (Saxenian,

1994; Porter, 2000). The movement of factor inputs and outputs is also facilitated by accessibility increases, which lowers the cost of production. Similarly, labor market access for a parcel is increased by new transportation investment, lowering the cost of finding new employees.

One of the most critical benefits created by accessibility increases is an increase in information availability (Marshall, 1890; Hoover, 1937). This benefits businesses not only by providing easier access to knowledge of new business opportunities or competitive strategies,

(Stam, 2007; Figueiredo, Guimaraes, & Woodward, 2002), but also by fostering informal interactions and idea exchange – even among employees in different industries – that can lead to regional learning, innovation, and eventually entrepreneurship (Saxenian, 1994; Jacobs, 1967;

Wennekers & Thurik, 1999). These impacts to entrepreneurship – which are also supported by the presence of the other agglomerative benefits identified here – are critically important for creating sustainable regional economies, because new businesses drive technological change, create employment, and protect the region from macro-scale economic shocks (Jacobs, 1967;

Wennekers & Thurik, 1999; Birch, 1987).

Finally, increased accessibility creates a spillover effect for neighboring parcels of land that has the potential to create a positive feedback loop, increasing other specific agglomeration benefits even more. For instance, a new transit station not only increases the agglomerative potential of the parcels immediately adjacent to it, likely leading to some new business creation or relocation there, it also creates secondary benefits on nearby parcels; if those develop into complimentary businesses, additional localization or urbanization economies could be created based on the burgeoning cluster of new business activity focused on the transit station.

As Figure 1 shows, the economic potential created by a transportation investment is substantial, and the quantifiable result of the presence of these agglomeration factors is an increase in property values for the newly-accessible land. While property value increases are an important indicator of increased accessibility, by focusing research only on the end result of this complex process, previous research has tended to oversimplify and pass over the specific mechanisms by which transportation investments increase property values. The complex nature of the agglomerative potential created by new transportation modes deserves additional research,

14 particularly because the impact of individual modes on individual businesses can vary quite substantially.

In addition to this lack of understanding of the specific mechanisms by which transportation systems create agglomerative potential, there is likewise a lack of research on the economic connection between urban transportation modes and specific types of businesses, i.e., what industries gain most from what kinds of transportation investments. Transportation modes can be broken down (for economic purposes) in several ways based on their cargo type and speed. As Figure 2 shows, at the most basic level, a given transportation mode can carry either humans or freight (Hanson & Guiliano, 2004). Passenger modes can also be classified by their speed – “rapid” modes like highways, streets, and transit have higher mobility and are generally able to connect larger portions of the region, while “slow” modes like biking and walking allow users to interact with their environment at a slower speed and a human scale, allowing for chance encounters, increased physical activity, and enjoyment (Vojnovic et al. 2013).

Connections (shown in Table 1) can then be made from these broad classes of transportation modes to each of the agglomeration benefits outlined above, which are themselves connected to specific types of businesses which might theoretically be predisposed to benefit from specific types of agglomeration benefits. Market area benefits are most likely to be created by rapid transit modes, since these extend regionally rather than at a neighborhood scale

(Rosenthal & Strange, 2001); the businesses most likely to benefit from these economies are customer-oriented firms like retailers, services, entertainment, and healthcare. These are businesses for which access is very important; thus, well-planned highway, street, and transit development are most likely to benefit or encourage these types of businesses. Social networks are fostered through all kinds of interpersonal interaction, sometimes extending regionally (as in

15 the case of clients for advanced service firms or business network organizations) as well as locally (Nelson, 2005). The businesses most likely to benefit from this kind of face-to-face communication are those require a lot of face-to-face communication and interaction to do business: advanced service firms and consultants, high-tech businesses, and educational institutions.

Figure 2. Diagram showing hierarchy of different transportation mode types.

As might be expected, freight access is only theoretically benefited by modes that carry cargo, which serves firms in sectors like manufacturing, wholesale, and large-scale retail most prominently. Labor market access requires regional-scale access to people (Rosenthal & Strange,

2001) but is likely most useful for businesses without specialized hiring needs, such as manufacturing, wholesale, retail, personal services, food, and entertainment. Industries that are 16 more specialized or require higher levels of human capital will most likely conduct targeted searches for employees that are less dependent on transportation access to a large pool of candidates. Finally, information economies are most directly fostered by “slow” transportation modes that favor chance encounters, informal interaction, casual socialization, reflection, and personal experience of situations and conditions. Previous research on information benefits has shown that they are present only over relatively short distances (Rosenthal & Strange, 2001), which makes sense due to the fact that they are related to personal exchanges and knowledge of local conditions. And, while information economies can benefit businesses in all sectors, they are particularly useful for businesses in advanced services, high-tech, and education.

Table 1. Theorized tendencies for transportation modes to produce given agglomerative economies, and the business types most likely to benefit.

Transportation Mode Agglomeration Mechanism Business Types Rapid Market Access Large- and Small-Scale Retail, Personal Services, Food, Entertainment, Healthcare All Passenger Social Networks Advanced Services, High-Tech, Education Freight Freight Access Manufacturing, Wholesale, Large- Scale Retail Rapid Labor Market Access Manufacturing, Wholesale, Large- and Small-Scale Retail, Personal Services, Food, and Entertainment Slow Information Advanced Services, High-Tech, Education

Of course, additional empirical research – currently lacking in the literature – is needed to confirm, complicate or reject the broad connections sketched out here, specifically when it comes to non-auto transportation systems. There are also additional modal differences – within transit, for example, between light rail, heavy rail, and commuter rail modes – that need to be

17 examined in more detail, given the vast differences in cost, user perception, and political palatability between them. Regional comparisons are also needed between areas with historically dense urban development patterns and well-developed, mature transit systems (like Boston, MA) and those focused currently on auto-oriented development and highway construction (such as

Phoenix). Different regions also of course have vastly different industrial concentrations, business cultures, internal supply/demand structures, and patterns of investment that will influence the connection between any individual business and a given transportation mode. In addition, fine-grained temporal considerations, such as the “novelty factor” identified by previous researchers that suggests a new transportation investment is most economically important at the time it opens, with declining importance as it continues to operate and the novelty wears off (Mohammad et al. 2013; Golub, Guhathakurta, & Sollapuram, 2012), need to be considered in more detail.

Existing empirical research has only just begun to broach these subjects in a meaningful way. The existing work on the connection between transportation and agglomeration does not explore the fine-grained, individual mechanisms through which specific transportation modes increase agglomerative potential (Chatman & Noland, 2011). And while researchers have begun to study the connection between public transit modes and new business creation in specific industries and regions (Mejia-Dorantes, Paez, & Vassallo, 2012; Song et al. 2012; Chatman,

Noland, & Klein, 2016), additional study areas, business types, regions, temporal variation, and methodological approaches are needed. Significantly, existing studies have not used methods that explicitly take into account either the hierarchical nature of aggregated geographic data

(such as hierarchical linear models) or the spatial structure of connected areal units (such as spatial econometric models). And, while time series approaches have been used, quasi-

18 experimental methods that explicitly test the direct impact of the construction of a new rail system – and also provide more fine-grained detail on pre- and post-construction trends– are currently lacking in the research record.

Research Goals

Given these gaps in the existing literature, the goal of this dissertation is to examine the extent to which investments in transit and walkable built environments influence business formation, location, and performance (in specific industries) by applying various quantitative methodological approaches to modeling individual-level business data and neighborhood characteristics. These approaches include quasi-experimental time-series methods that isolate the treatment effect for a specific intervention (in this case, the construction of a new light rail line in

Phoenix) by taking into account temporal trends in the data, hierarchical linear models (HLM) that explicitly incorporate the nested, scale-dependent nature of geographic data (e.g., individual business points within neighborhoods), and spatial econometric methods that account for the effect of neighboring observations on the variable of interest.

To do this, the project employs detailed data on individual business establishments from the National Establishment Time Series (NETS) that allow for fine-grained spatial and temporal analysis and provide information on industrial classification, sales volume, size, and other important characteristics. These data are combined with supplemental sociodemographic, economic, and built environment data at the neighborhood scale in order to examine the relationships between individual businesses and their surrounding urban context. Given the large nature of these datasets (which provide a census of the business establishment population in a given year), this dissertation is able to provide a more complete picture of the economic

19 processes at work than in many previous studies, and also offers insights into methodological techniques that can be applied to future studies using large datasets and/or combining individual and neighborhood-scale characteristics.

Specifically, the first paper, “Transit-Oriented Economic Development: The impact of light rail on new business starts in the Phoenix, AZ region1,” analyzes the rate of new business formation within walking distance to transit stations in Phoenix. This paper uses an “adjusted interrupted time series” (AITS) Poisson regression methodology to analyze 25 years of new business point data. This unique method establishes ‘control’ and ‘treatment’ areas and uses a long time series of pre-intervention data to establish a baseline trend, providing a relatively definitive assessment of the independent impact of light rail on new business formation. In this paper, the control area – which represents the ‘baseline’ condition for new business activity in the region – is based on automobile accessibility. This provides a direct test of light rail’s effectiveness in new business attraction versus the more prevalent automobile mode.

The second paper, “Transitive Properties: A spatial econometric analysis of new business creation around transit,” builds on this analysis by looking explicitly at the relationship between transit station proximity and new knowledge industry business creation in five established transit and/or high technology regions – Boston, San Jose, Austin, Cleveland, and Philadelphia. To obtain a log-transformable distribution that can be used in spatial econometric models, different smoothing techniques for four types of new knowledge industry businesses (standardized by area, population, or number of existing businesses) are tested. The results of the spatial models used in the paper are also tested against anon-spatial Poisson model in order to ascertain the

1 This paper has been published in Urban Studies: Credit, K. (2017). Transit-Oriented Economic Development: The Impact of Light Rail on New Business Starts in the Phoenix, AZ Region. Urban Studies. DOI: 10.1177/0042098017724119. 20 degree to which non-spatial models overestimate the impact of transit proximity on new business creation.

The third paper, “Place-making and Performance: The impact of walkable built environments on business performance in Phoenix and Boston2,” addresses the influence of walkable urban design and place-making on business performance in Phoenix and Boston. While previous research has assessed the impact of the built environment on transportation mode choice (Ewing & Cervero, 2010; Cervero & Kockleman, 1997; Chatman, 2013), there is relatively little work on the built environment’s impact on business performance. Previous studies have found a connection between walkable urban environments and property values

(Pivo & Fisher, 2011); others have just begun to look at the qualitative relationship between business type and urban design (Spencer, 2015). This paper contributes to this burgeoning field by using a two-level hierarchical linear model to assess the relationship between fine-grained land use factors (e.g., Cervero & Kockelman’s ‘5-D’s’ (1997) and Jacob’s four generators of diversity (1961)) and the performance of existing businesses in Phoenix and Boston, measured by sales volume per employee.

From a theoretical perspective, the intellectual merit of this work is its exploration of the specific ways in which investments in urban form and transportation – such as walkable neighborhoods and transit systems – influence the formation, clustering, and performance of individual businesses. While previous research has begun to look at the neighborhood-level determinants of new business activity (Renski, 2008; Spencer, 2015; Mack & Credit, 2016), little is known about the micro-foundations of the relationship between urban form and

2 This paper, a collaboration with Elizabeth Mack, has been published in Environment and Planning B: Credit, K., & Mack, E. (2017). Place-Making and Performance: The Impact of Walkable Urban Design on Business Performance in Phoenix and Boston. Environment and Planning – B. DOI: 10.1177/2399808317710466. 21 entrepreneurship. By addressing these gaps, this dissertation contributes to the theory of business location, providing new insights on the micro-foundations of agglomeration and business performance that can be used as a building block for future work. In addition, the methodological techniques employed in this project can be used by future researchers as large individual datasets, with fine-grained spatial and temporal characteristics, become more readily available.

This project also has significant broader impacts to practicing urban planners, elected officials, and entrepreneurs, since both transit development and walkable neighborhood design are popular urban policies currently being undertaken or considered in many North American regions. Given the scale of these kinds of interventions, their long-term nature, and the political claims often made about their economic benefits, there is a direct need for policy-makers to know whether the investments they are making (or plan to make) in urban infrastructure and design achieve the economic goals they desire. Beyond government policy, a better understanding of regional growth processes can also influence the day-to-day lives of city residents, business-owners, and stakeholders, by giving them the relevant information they need to enhance the economic success and stability of their businesses, neighborhoods, and regions.

What follows in the next three chapters is the full text of each of the three individual papers, including figures and tables; all references and appendices are presented together at the end of the entire document. After the three substantive chapters, the conclusion presents the primary limitations of the work, as well as a summation of its generalizability and key implications for both theory and practice.

CHAPTER 1. TRANSIT-ORIENTED ECONOMIC DEVELOPMENT: THE IMPACT OF LIGHT RAIL ON NEW BUSINESS STARTS IN THE PHOENIX, AZ REGION. Citation: Credit, K. (2017). Transit-Oriented Economic Development: The Impact of Light Rail on New Business Starts in the Phoenix, AZ Region. Urban Studies. DOI: 10.1177/0042098017724119.

Introduction

Over the last thirty years in the United States, billions of tax dollars have been spent on the construction of some 650 miles of light rail3 lines in 16 regions, with nearly 150 additional miles planned or under construction (as of 2014) (Freemark, 2014). Most of these new systems – and complementary transit-oriented development (TOD) schemes – have been touted as economic development tools to catalyze infill development and reinvestment in central cities.

But, despite a multitude of studies into the influence of rail transit on station-area property values, much remains unknown about the economic impact of transit investments (Cervero &

Landis, 1997; Cervero, 2004; Agostini & Palmucci, 2008; Golub, Guhathakurta, & Sollapuram,

2012).

The Phoenix metropolitan area provides an important context in which to study the effects of new transit system construction and associated transit-oriented development. It is one of the largest and fastest-growing regions in the United States, having more than doubled in population from 2.1 million in 1990 to almost 4.7 million in 2016 (NHGIS, 1990; United States

Census Bureau, 2016). As such, it is a key representative of the American ‘Sunbelt’ – warm- weather regions that that have attracted the largest gains in population since World War II, but often without investment in fixed transit infrastructure and an over-reliance on sprawling,

3 In this paper, light rail is defined as “an electric railway with a ‘light volume’ traffic capacity compared to heavy rail. Light rail may use shared or exclusive rights-of-way, high or low platform loading and multi-car trains or single cars.” (APTA, 1994, p. 23). As De Bruijn and Veeneman (2009) point out, as a mode light rail is distinguished by the fact that it is a fixed investment and higher-capacity than a bus, but more flexible and lower-capacity than a heavy rail system. The Phoenix Valley Metro light rail system examined in this paper fits this definition exactly. 23 automobile-oriented development patterns (The Economist, 2017). This makes the introduction of rail transit – and its economic impacts – an interesting object of study in regions like Phoenix, where the existing land use and behavioral patterns do not necessarily support transit-oriented development or ridership. And, as additional Sunbelt cities consider transit projects aimed at increasing transit-oriented development and investment, it has become increasingly important to understand their spatial and temporal economic impacts.

Existing research on the economic impacts of transit stations has focused primarily on changes in station-area property values. While this is a useful way to quantify the impact of transportation investments, property values only provide one piece of a complex picture. Impacts to new business activity are also important to consider because new businesses make substantial contributions to regional economic growth and diversity (Wennekers & Thurik, 1999; Frenken,

Van Oort, & Verburg, 2007). Recent research has begun to investigate the link between transit and new business creation (Chatman, Noland, & Klein, 2016; Song, et al. 2012), but significant questions about the temporal and spatial extent of transit's impact on new business starts remain

– for instance, previous studies of transit investment have found a spike in economic activity leading up to the opening of new systems, suggesting a ‘novelty factor’ for transit that wears off after the system becomes a standard part of the regional transportation network (Mohammad et al. 2013; Golub, Guhathakurta, & Sollapuram, 2012) that deserves further empirical investigation.

By using an “adjusted interrupted time series” (AITS) Poisson regression methodology to analyze new business point data in Phoenix, Arizona from 1990-2014, this paper delivers a quasi-experimental examination of transit’s impact on new business formation, providing estimates of the distance decay function for entrepreneurial benefits related to transit

24 infrastructure (Galster et al. 2004). In so doing, this paper answers three important questions.

First, what impact does the construction of a new transit line have on adjacent new business activity in the retail, service, and knowledge industries, when compared to ‘baseline’ business activity in more automobile-centric neighborhoods? Second, does transit-adjacent new business activity peak when the system opens, or continue to increase afterwards? And, third, how do the system’s impacts on new businesses vary as distance from the stations increases?

The results of the analysis show that, since Phoenix’s light rail line has been operational, neighborhoods within walking distance (1-mile) of transit stations have experienced a substantial increase in new business starts in the knowledge, service, and retail sectors when compared with automobile-accessible control areas. However, this increase in the absolute number of businesses erodes over time, suggesting the presence of a ‘novelty factor’ for the system (Mohammad et al.

2013). The results also provide insight into the spatial extent of light rail impacts to new business formation, with areas 1 mile from stations observing 21% fewer retail new business starts and

12% fewer knowledge sector new starts than areas within ¼-mile of stations. Given the spatial context of business location and planning policy decisions, the geographic character of transit’s impact on business formation is particularly important for policy-makers and entrepreneurs to understand and could help guide future station-area planning processes in other regions.

Theoretical Context

Transit-oriented development is a theoretical planning concept designed to help mitigate the negative externalities of suburban sprawl (Calthorpe, 1993; Brueckner, 2000; Newman &

Kenworthy, 2013). The idea, which has now been implemented in several cities in North

America, is based off of Peter Calthorpe’s “pedestrian pocket” (1993) and involves creating small nodes of walkable urbanism that are connected to the rest of the region by transit.

TOD has several attractive features for promoting urban sustainability. By providing complementary development around transit stations, it increases use of the regional transit network, thus decreasing congestion, automobile pollution, traffic deaths, and other negative externalities associated with suburban environments (Brueckner, 2000; Ewing, 1997). TOD also provides new options to meet the unmet demand of consumers who would prefer to live in a walkable community but cannot find housing in areas that meet their financial or employment needs, which – according to some sources – is substantial (ULI, 2015; APA, 2014).

Most importantly for this paper, TOD fosters economic development around new transit stations, which have a built-in market potential that is sometimes not realized due to poor station accessibility (Newman & Kenworthy, 2013). This increased development potential can be used as both an engine for new business formation and an economic revitalization tool for areas that are suffering from extreme disinvestment.

Much of the related literature focuses on implementation and design principles for successful TOD (Dittmar & Ohland, 2004; Curtis, Renne, & Bertolini, 2009). Empirical research on the economic benefits of TOD – which focuses heavily on property-value impacts – is generally mixed in its conclusions. There is some evidence that transit generally has a positive effect on property values and/or local business development, especially when a) the transit service itself is high-quality and b) the project is developed in conjunction with land-use planning that maximizes the accessibility benefits conferred by the investment (Damm et al.

1980; Landis, Guhathakurtra, & Zhang, 1994; Knaap, Ding, & Hopkins, 2001; Cervero, 2004;

Agostnini & Palmucci, 2008; Mohammad et al. 2013). Light rail systems in particular have been shown to confer economic benefits to the value of properties in proximity to stations, generally with a ‘nuisance’ penalty in value for those properties directly adjacent to the tracks

(Weinberger, 2001; Weinstein & Clower, 2003; Golub, Guhathakurta, & Sollapuram, 2012). In addition, recent research shows that commercial properties (or land) tend to be much more highly impacted by rail investments than residential properties, and “the perceived benefit of the rail system at time of announcement is often higher compared to the actual realized benefit after the system stabilizes,” a result that speaks to the novelty factor of new transportation infrastructure (Mohammad et al. 2013).

However, other studies find that the connection between transit and economic development is dependent primarily on the underlying context of the region (Vessali, 1996;

Giuliano, 2004; Schuetz, 2014). Some question the extent to which TOD is able to promote compact development, reduce automobile use, and provide secondary economic benefits (Quinn,

2006; Bollinger & Ihlanfeldt, 1997; Chatman, 2013). Giuliano raises the issue that economic growth observed due to transit development may simply be a re-focusing of development that would have occurred somewhere else (2004). Other work has found a significant difference in the new business-creating impact of transit systems in different types of regions, with the more auto-centric planning policies of Dallas-Ft. Worth perhaps fostering reduced economic impact from rail systems (Chatman, Noland, & Klein, 2016).

So, despite a large research effort into the economic impacts of transit systems, key questions remain as to the spatial impacts of transit infrastructure on new business formation, as well as the presence of a ‘novelty factor’ in transit system use, and the viability of TOD in more auto-oriented regional environments.

Conceptual Framework

While property values capture the economic benefits provided by transit systems, they do not speak to the specific role of transit in fostering new business formation. This underexplored

27 relationship is central to this paper’s thesis, because new businesses benefit cities in several ways that go beyond direct monetary impacts. New businesses serve as conduits for innovative activity and knowledge spillovers (from parent companies or large research institutions). In addition, the knowledge spillovers and innovations that some new firms develop have the potential to diversify the industrial base and provide protection from economic shocks to particular sectors

(Jacobs, 1967; Frenken, Van Oort, & Verburg, 2007). Finally, a variety of new (often small) businesses has the potential to add significantly to the land use diversity and urban vitality that planners have long espoused (Jacobs, 1961). While entrepreneurship is critical to the economic development of cities, its connection to transportation requires further study. Figure 3 shows a variety of economic benefits that transit infrastructure provides to businesses of different types.

Social networks, trust, and “weak” ties fostered by face-to-face contact and relationship building are important components of contemporary business success, especially for firms in industries that require strong interpersonal client relationships, such as producer services (Gordon &

McCann, 2000; Nelson, 2005; Storper & Venables, 2004). Social embeddedness is also particularly important for new, small businesses that rely on local networks and personal contacts to survive (Robinson, Dassie, & Christy, 2004; Stam, 2007).

More traditional agglomeration benefits are also fostered by transit infrastructure. Transit provides increased regional access to labor markets, particularly for businesses without specialized hiring needs, such as retail, personal services, and hospitality (Chatman, Noland, &

Klein, 2016; Chatman & Noland, 2011; Rosenthal & Strange, 2001).

Figure 3. Conceptual framework showing how transit infrastructure influences individual business location decisions and clustering.

By supporting informal interaction, casual socialization after work, chance encounters, and reflection, transit can help foster the knowledge spillovers and information exchange that have been identified as key drivers of innovation in the knowledge economy (Chatman &

Noland, 2011; Asheim, Lawton, & Oughton, 2011; Saxenian, 1994). Transit is particularly well- suited to foster these benefits, as they exist only over relatively short distances (Rosenthal &

Strange, 2001). While knowledge spillovers can benefit businesses in all sectors, they are particularly useful for firms in the knowledge industries, producer services, and education. On the other hand, benefits from a larger market area and higher visibility are most likely to benefit customer-oriented firms like retailers, service providers, and hospitality businesses, for which visibility plays an important role in start-up location decisions (see, e.g., Bradley University,

2008). In an urban context, transit stations function as primary transportation nodes, increasing the visibility of surrounding businesses and providing access to a larger market area and increased demand potential.

Finally, transit plays a role in fostering an urban, car-free lifestyle that may be attractive to the millennial generation (Weissmann, 2012; APA, 2014; ULI, 2015; RSG, 2014). As researchers like Richard Florida have theorized, the new creative economy is driven by highly- skilled workers with specific kinds of neighborhood preferences (2002; 2004). This reverse type of labor sorting has helped fuel a burgeoning “back-to-the-city” movement and increasing demand for urban amenities (Hyra, 2015). While the extent to which this trend constitutes a real shift in urban development remains to be seen (Sturtevant & Jung, 2011), businesses interested in fostering socially-conscious or environmentally-friendly images to recruit younger, talented workers – especially those in competitive creative industries such as high-tech or producer

30 services – may play an increasing factor now and into the future (Orlitzky, Schmidt, & Rynes,

2003; Hammann, Habisch, & Pechlaner, 2009; Florida, 2002).

Given the variety of reasons why new businesses in particular industries might find an advantage from locating near transit stations, this paper explores the transit-entrepreneurship relationship in more detail by studying the rate of new business formation in three sectors particularly influenced by transit’s accessibility benefits. Retail businesses (NAICS 44-45) include stores selling retail goods, from cars to home goods, and grocery stores to office supply stores. Knowledge businesses (NAICS 51-52 and 54-55) include firms in the information sector like software publishers, finance and insurance companies like banks, and professional and scientific services like lawyer’s offices and management consulting firms. Services (NAICS 72 and 81) include hotels, restaurants, bars, and other personal services like dry cleaning and barber shops.

Study Area and Data

The Phoenix metropolitan area is a particularly interesting region in which to study the effects of light rail construction on entrepreneurship, for several reasons. First, its population has grown dramatically since 1990, and is now the twelfth largest metropolitan area in the United

States (NHGIS, 1990; United States Census Bureau, 2016). Second, it exemplifies the large-scale demographic shifts that have taken place in the US since the end of WWII, as millions have left the industrial Midwest and East Coast for warmer cities in the ‘Sunbelt’, like Phoenix, Orlando,

Houston, and Atlanta (The Economist, 2017). In terms of urban form, the hasty growth in these areas has largely been automobile-focused and sprawling, which discourages transit use.

However, it is increasingly important to study the economic impacts of new transit systems in these traditionally automobile-oriented contexts, especially as these areas plan for new rail transit

31 network construction and expansion as they experience renewed post-Great Recession population growth.

The business location data for this paper comes from two separate sources, both based on the Dun and Bradstreet business database. For 1990-2009, the paper uses 2012 National

Establishment Time Series (NETS) data prepared by Walls & Associates; for 2010-2014, the data come from ESRI Business Analyst (Walls & Associates, 2012; ESRI, 2014). Dun and

Bradstreet collect data on every business establishment in the country, providing a near census of the business population. Also, this data provides much greater coverage of very small and self- employed establishments compared with federal data sources, making it an ideal source to study the full pattern of entrepreneurship over time (Walls & Associates, 2012). The decision to combine two different datasets comes from the necessity of measuring impacts after the light rail line opened in 2008.

Between 1990 and 2009, a business is considered a new start if it is a non-branch plant domestic establishment with at least one employee that did not relocate4. Due to the limited nature of the ESRI data, for 2010-2014 a business is included in the analysis only if it survived until 2014 and had at least one employee; relocation, branch-plant, and foreign-ownership information is not included. Despite the discrepancies between the two datasets, the data used are the same (by year) for both the control and treatment areas; thus, any systematic bias in the way that the data are collected would impact both areas equally. The business points are divided by industry and opening year and spatially joined to 2010 Census blocks.

In order to create a stable boundary for the inclusion of secondary Census data from

1990, 2000, and 2010 in the analysis, demographic data from 1990 and 2000 were first joined to

4 For NETS data, the businesses are geocoded to their last location. This makes it necessary to consider only businesses that did not relocate to ensure that the spatial locations of the points are correct. 32 their corresponding Census block geographies. These polygon shapefiles were then converted to points and spatially-joined to the 2010 Census blocks. The counts attached to these points were then summed and divided by their corresponding summed total to create a percentage for the variable that matches the 2010 Census block boundary. Additional control variables available at the point level – such as the number of existing retail, manufacturing, and information businesses in each block in each year – were then spatially joined to the stable 2010 Census geography.

Methods

In order to gauge the economic impact of the light rail over time, this paper uses the

AITS methodology of Galster et al. (2004). This method uses time series data in a regression framework to compare the change in a given variable of interest – in this case, the count of new business starts in specific industries by Census block – between an identified treatment area and a control area, both before and after the point of intervention. In this paper the intervention point is 2008, when the light rail system opened.

Methodologically, AITS is somewhat similar to other difference-in-differences approaches that have been used in previous studies to determine the impact of policy interventions on neighborhood characteristics (Tranel & Handlin, 2006), including in rail transit contexts (Pagliara & Papa, 2011). AITS, however, goes beyond the standard difference-in- differences technique by explicitly accounting for the change in both the absolute value and trend of the dependent variable before and after the intervention. One of the strengths of the method is that it statistically models differences both cross-sectionally (between the treatment and control areas) and longitudinally (between the pre-intervention and post-intervention time periods of the impact area). This provides a robust estimate of the impact of a particular policy intervention,

33 despite the many different forces at work at any given time in complex urban environments

(Galster et al. 2004).

Treatment and Control Area Selection

As with any quasi-experimental study design, the selection of the treatment and control areas deserves care. Rather than selecting an arbitrary boundary for these areas, this paper explicitly tests a range of definitions in order to find the combination with the most similar pre- intervention trend. For the light rail treatment area, network-based buffers using the United

States Census Bureau 2015 TIGER/Line road shapefile are calculated for ¼, ½, and 1 mile

(corresponding to 5, 10, and 20-minute walks, respectively) around each transit station to form three possible treatment area boundaries. Ranges from ¼ to ½ mile have been cited previously in the literature as the standard area of impact around transit stations (Mohammad et al. 2013;

Guerra, Cervero, & Tischler, 2011; Calthorpe, 1993; Zhao et al. 2003); the larger 1-mile boundary is included in order to test whether light rail impacts to business creation might extend beyond this typically-assumed range.

Given this paper’s interest in comparing the impact of the automobile and light rail modes on business formation rates, the control area is similarly-defined by network-based buffers calculated around major intersections (selected based on functional classification5) for 10-minute drives at average speeds of 15 mph, 30 mph, and 45 mph (2.5, 5, and 7.5 miles, respectively), which represent different ways to conceptualize automobile accessibility. Figure 4 shows the six boundary definitions considered for analysis, as well at the location of the light rail stations, light rail line, major intersections, and highways.

5 Major intersections were selected first by identifying intersections between US, State, and Interstate highways within the cities of Phoenix, Mesa, and Tempe. A geographically-representative group of additional intersections between these roads and other major arterials was also added in order to ensure ample coverage. 34

Figure 4. Map showing six treatment and control areas considered for analysis.

In quasi-experimental study design, it is important to select treatment and control groups that have similar pre-intervention trends in the variable of interest to limit selection bias and provide greater confidence that post-intervention differences between the groups are in fact due to the intervention and not some pre-existing differences between the groups. In order to select the most similar combination of treatment and control areas, the average number of new business starts per year (in each of the three industries of interest) was plotted on a graph for each of the six boundaries considered. OLS was then used to obtain an equation for the linear trend of new business starts over time for each of the boundary definitions.

The slope coefficients for each of these equations (which indicate the average linear trend for the given boundary definition) were then compared for each of the nine possible combinations of treatment and control areas, as shown in Appendix A. The boundary combination with the least difference between slope coefficients (i.e., the most similar trend) was the 1-mile treatment area and the 7.5 mile control area, which was then used in the estimation of the AITS model. Appendix A.3 contains graphical depictions of the average trend in new business formation by block for this selected boundary definition for each of the three types of businesses of interest.

Specification of the AITS Model

The specification of the AITS model used in this paper follows the basic equation laid out by Galster et al. (2004) but uses a Poisson regression model to assess the relationship between the variables. The dependent variable, in this case, is the count of new businesses (by block) for each of the three industries of interest: retail, knowledge, and services. Poisson models are preferred in the use of count data, and work on the assumption that the dependent variable has a

Poisson distribution, with the natural log of its expected value modeled linearly (UCLA, 2007;

Faraway, 2006):

(1) log(휇푖) = log(훿) + 푧훾 + 푥훽, where 휇 is the mean of the Poisson distribution of the dependent variable for a given industry of interest, i; 훿 is an exposure variable that standardizes the dependent variable count by a set exposure (rate); 푧 is a vector of dummy and trend variables that control for pre-intervention (and overall) differences between the treatment and control areas and identify post-intervention differences in both the trend and absolute value of the dependent variable;푥 is a vector of relevant non-collinear sociodemographic, economic, and spatial covariates (the independent variables); 훾 and훽 are the regression coefficients for the given explanatory variables 훼 and 푥, respectively. Appendix A.4 shows the variables used in the final AITS specification. The AITS model used in this paper was run in Stata version 13.1 using the “poisson” command, after converting the relevant time-series data from “wide” to “long” format.

Even with the careful selection of treatment and control areas, differences are sure to exist between any experimental groups chosen in complex urban environments. The unique benefit of AITS is that it controls for these inevitable differences using a set of dummy and trend variables. The 퐷퐼푀푃 variable controls for absolute differences in the pre-intervention values of the dependent variable between the control and treatment groups, while the 푇푅퐼푀푃 variable controls for differences in the pre-intervention trend in the dependent variable between the two groups. In addition, the 푇푅퐴퐿퐿 trend variable controls for the overall trend, ensuring that a significant result is not simply an artifact of large-scale trends affecting both areas.

In addition to these controls, other variables (shown in Appendix A.4) are included in the model in order to ensure that other specific differences between the treatment and control areas

37 do not bias the results, including controls for spatial autocorrelation at two scales – regional and local. A dummy variable denoting location in one of the region’s downtown business districts generally controls for regional spatial autocorrelation, while a spatial lag of the average value of the dependent variable in neighboring blocks controls for local spatial autocorrelation, or the fact that high values of new starts in one block may be significantly impacted by high values of new starts in the neighborhood (Ward & Gleditsch, 2008). Additional variables6 control for demographic differences between blocks, business dynamics, general economic characteristics, and the built environment characteristics of each block.

Finally, in order to test whether the intervention is significant or not – to answer the research question of the paper – a dummy variable denoting the absolute value of the dependent variable in only the treatment area for only the post-intervention time period is included

(퐷푃푂푆푇퐼푀푃) in the model, as well as a trend variable for only post-intervention observations in the treatment area (푇푅푃푂푆푇퐼푀푃). Given the careful study design employed in this paper, a significant positive result for 퐷푃푂푆푇퐼푀푃 indicates that the light rail made an impact on the absolute number of new businesses in the treatment area. Similarly, a significant positive result for the 푇푅푃푂푆푇퐼푀푃 variables indicates that the slope of the trend line for new business starts in the treatment area has shifted upward due to the construction of the light rail.

Results

Socioeconomic Context of Light Rail Station Locations

Before describing the model results, it is important to develop a socioeconomic context for the light rail station locations, since transit investments are always linked to a region’s

6 Correlation matrices (available upon request) were used to assess collinearity between a larger set of potential covariates. Confounders were removed from the final specification, leaving those listed here. 38 socioeconomic stratification and political decision-making processes. To do this, an adjusted

Darden-Kamel Composite Index (Darden et al. 2010; Darden & Kamel, 2000) was created at two scales – the regional and station-area – in order to better understand how socioeconomic factors influence planning for the overall path of the light rail system, as well as the specific location of individual stations. The index uses eight relevant indicators (by Census block group) in 2000, one year after the first proposal for the Phoenix light rail line was developed (Golub,

Guhathakurta, & Sollapuram, 2012): per capita income, public assistance income, unemployment rate, homeownership rate, percent Bachelor’s degree attainment or higher, percent of workers in management and professional occupations, median contract rent, and median housing value. The index was calculated by summing the z-score for each of the eight variables, and then dividing the resulting values into quartiles at two scales: 1) based on the entire region’s index scores, and

2) based on the index scores of just the 1-mile light rail treatment area, which generally represents station-adjacent neighborhoods. Each block group falls into one of the following categories: low, medium-low, medium-high, and high socioeconomic status (SES).

Figure 5 includes the number and percentage of block groups containing stations in each of the socioeconomic status categories for both boundaries. Interestingly, at the regional scale, a very small number of stations are located in block groups with high or medium-high SES, supporting the political evidence that the wealthiest areas – including cities like Scottsdale – declined to be connected to the light rail line. It also reflects the obvious correlation between high socioeconomic status and suburban/exurban areas in which light rail planning makes little sense due to reduced population density and sprawling land use patterns.

Figure 5. Map showing SES index score (in 2000) distributions for Maricopa County and light rail-adjacent block groups.

However, if the analysis is confined only to block groups that intersect a 1-mile network buffer of stations – that is, examining the fine-grained spatial decision-making process for individual station locations within the general path of the rail line – over 61% of block groups with stations have medium-high or high SES. This shows that, within the context of fine-grained decision-making, station locations are spatially-biased towards block groups with higher socioeconomic status. They are also driven by a desire to connect large employers, entertainment facilities, and institutions in downtown Phoenix and Tempe. This is important to understand when interpreting the AITS model results, since the location of the light rail line itself – as well as associated entrepreneurship and economic investment – is influenced by regional and local social conditions. Even though the light rail might significantly impact new business starts, these businesses are still likely to be located in relatively-higher SES areas (due to the location of the stations) and thus perhaps less likely to directly impact disadvantaged populations and spur revitalization.

Descriptive Statistics

Descriptive statistics for the variables used in the model are shown in Appendix A.5.

They provide a general overview of the economic context for both the treatment (1-mile) and control (7.5-mile) areas. In general, the blocks closest to the light rail station show slightly higher levels of new business starts in the three industries of interest over the time period, with the biggest average difference in knowledge sector new starts. Economic variables are also found in higher average quantities in the treatment area, including average employment density (10.92 employees per acre to 1.83) and the percentage of existing retail, manufacturing, and information establishments, which makes some sense due to the larger proportion of blocks within downtown business districts (7.9% to .4%). The built environment and demographic variables are relatively

41 similar between the two areas, except that average block length and block size are smaller in the treatment area than in the control area.

Descriptive statistics for the pre- and post-intervention characteristics of the 1-mile treatment area are shown in Appendix A.6. Population, employment density, the percentage of existing businesses in the retail and information sectors, the percentage of population that is black non-Hispanic, and the percentage of the population aged 19-64 all increased slightly in the post-intervention time period; interestingly, only the percentage of existing manufacturing businesses decreased. On the other hand, there are slightly fewer new starts in each industry (on average) in the post-intervention time period than in the pre-intervention time period. However, it is not possible from looking at these raw averages to know whether these drops were higher or lower than the baseline trend in the auto-oriented control areas, or what the post-intervention impact is after controlling for important intervening factors, such as location in a downtown business district. In order to obtain the direct statistical impact of the light rail while controlling for such differences and trends, the AITS model must be used.

AITS Model Results

The results of the AITS regression, incidence rate ratios (IRR), and percent change values

(1 – IRR) in the dependent variable for a one-unit increase in a given covariate are shown in

Table 2. IRR are calculated as 푒 raised to the power of the given coefficient value, and are reported as numeric figures in Table 1, while 1-IRR values are reported as percentages

(Rodriguez, 2007; UCLA, 2007). All of the variables included in Table 2 are statistically- significant at a p-value of .05 or less.

Table 2. Poisson AITS model results and incidence rate ratios (IRR) for three dependent variables of interest.

Dependent variable # Retail New Starts # Knowledge New Starts # Service New Starts Coef. Std. Err. IRR 1 - IRR Coef. Std. Err. IRR 1 - IRR Coef. Std. Err. IRR 1 - IRR AITS variables (α) DIMP 0.535 0.053 1.707 71% 1.512 0.033 4.537 354% 0.753 0.049 2.124 112% DPOSTIMP 0.250 0.093 1.284 28% 0.630 0.046 1.879 88% 0.335 0.073 1.398 40% TRIMP -0.018 0.004 0.982 -2% -0.065 0.003 0.937 -6% -0.019 0.004 0.981 -2% TRPOSTIMP -0.073 0.026 0.929 -7% -0.079 0.013 0.924 -8% -0.058 0.019 0.943 -6% TRALL 0.064 0.001 1.067 7% 0.127 0.001 1.135 14% 0.075 0.001 1.078 8% TRPOSTALL -0.273 0.005 0.761 -24% -0.304 0.003 0.738 -26% -0.224 0.004 0.799 -20% Covariates (x) AVLAG7 0.839 0.010 2.315 131% 0.317 0.002 1.373 37% 0.964 0.010 2.622 162% BUS_D 1.381 0.029 3.978 298% 1.365 0.018 3.916 292% 1.303 0.027 3.682 268% INT_DEN -0.085 0.008 0.918 -8% -0.023 0.004 0.978 -2% -0.085 0.007 0.919 -8% BLOCK_L 0.000 0.000 1.000 -0.01% 0.000 0.000 1.000 -0.01% -0.0001 0.000 1.000 -0.01% BLK 0.619 0.155 1.857 86% -0.398 0.125 0.671 -33% 0.876 0.131 2.401 140% MID 0.375 0.030 1.455 46% 1.251 0.017 3.494 249% 0.815 0.026 2.260 126% EMPD 0.001 0.000 1.001 0.1% 0.001 0.000 1.001 0.1% 0.001 0.000 1.001 0.1% RETP 2.045 0.038 7.726 673% 0.304 0.040 1.355 36% 1.003 0.048 2.726 173% MANP 0.234 0.108 1.264 26% -0.243 0.078 0.784 -22% INFP 0.729 0.138 2.073 107% - - - - 0.354 0.139 1.424 42% Constant -5.927 0.018 0.003 -100% -6.141 0.014 0.002 -100% -6.034 0.018 0.002 -100% Offset (δ) ACRES 1 1 1 N 1040250 1040250 1040250 Psuedo-R² 0.085 0.126 0.089 All results reported here are significant at the .05 level (p ≤ .05). INFP not included in model for knowledge new starts due to collinearity.

7Values here reported for the spatial lag associated with each equation’s dependent variable: first retail, then knowledge, then services. 43

The 퐷푃푂푆푇퐼푀푃 variable is significant and positive for all three dependent variables, indicating that the treatment area shows a significantly higher number of new business starts per acre in each block than the control area. The light rail line increases new starts in the knowledge sector most significantly, with location in the treatment area from 2008-2014 worth an 88% increase in new knowledge business starts compared to the control area. There are also 40% more new starts in the service sector and 28% more new starts in the retail sector in the treatment area from 2008-2014, indicating that the light rail fostered new business activity in all three sectors of interest.

At the same time, this advantage is on a downward trend, indicated by the results for the

푇푅푃푂푆푇퐼푀푃 variable in all three regression equations. Each year after 2008, the number of knowledge new starts decreased by 8% in the treatment area compared to the control area.

Similarly, the treatment area’s advantage in service and retail new starts decreased by 6% and

7%, respectively, in the post-intervention time period. In simple terms, these results show that the light rail had a significant initial effect on new business starts in all three industries at the time it opened, but this impact has eroded consistently as time has passed since its opening.

As for the control variables, a few interesting relationships stand out. A 10% increase in black population relates to an 8.6% increase in retail new starts and a 14% increase in service new starts, but a 3.3% decrease in knowledge sector new starts, perhaps indicating the segregated nature of new business location decisions in Phoenix (Bolin, Grineski, & Collins, 2005). Also, as expected, the business district dummy variable is significant and positive for each of the three dependent variables and shows a substantially high coefficient value. Location in a downtown business district over this time period is worth over 250% additional new starts per block; however, the variable’s inclusion in the model does not alter the significance of the 퐷푃푂푆푇퐼푀푃

44 and 푇푅푃푂푆푇퐼푀푃factors. Therefore, this analysis shows that the light rail’s effect on new business formation functions independently of the positive entrepreneurial environment of downtown districts.

The average number of new starts of a block’s contiguous neighbors also significantly impacts each of the three dependent variables, indicating that the number of new business starts in these sectors is positively related to the presence of new business starts in neighboring blocks.

Interestingly, service and retail new starts exhibit substantially greater spillover effects than knowledge new starts. More importantly, the inclusion of these variables controls for the impact of these spillovers on the variables of interest, 퐷푃푂푆푇퐼푀푃 and 푇푅푃푂푆푇퐼푀푃.

These results are also robust to the definition of the treatment area boundary: AITS models run using alternative definitions of the treatment area – the ¼, ½, and 1-mile boundaries8 coupled with the same 7.5-mile control area – show similar signs and patterns of significance, with some interesting spatial variation in coefficient magnitude. While these additional treatment and control area combinations do not demonstrate the ideal minimization of difference in pre- intervention trend (described above), they do provide some insight into the spatial extent of light rail impacts on new business starts. Figure 6 shows the statistically-significant incidence rate ratio percent change values for 퐷푃푂푆푇퐼푀푃 for each of the three dependent variables of interest in these alternative models. Interestingly, the impact of the light rail on the absolute number of new business starts is greatest within the ¼ mile boundary and decreases with distance from the stations. This diminishing spatial gradient is greatest for retail new starts least significant for service new starts.

8 Full regression results for these models available upon request. 45

Figure 6. Map showing percent change incidence rate ratios for DPOSTIMP variable for varying definitions of the treatment boundary – ¼, ½, and 1 mile – in central Phoenix.

Discussion and Conclusion

The results of this analysis speak to three important factors regarding the planning and construction of the Valley Metro light rail system in Phoenix and associated TOD that can be applied in other fast-growing metropolitan areas with new fixed rail transit systems. First, when compared to automobile-adjacent areas, proximity to the light rail significantly increases new business starts in the three industries of interest: retail, services, and knowledge. Even when controlling for a variety of factors – including location in a downtown business district – adjacency to light rail stations is worth about 88% additional new starts in the knowledge sector,

40% new starts in the service sector, and 28% new starts in the retail sector over the time that the line has been open. This suggests that TOD planning efforts are economically worthwhile, given

46 the fact that new businesses have taken advantage of economic benefits in station-adjacent blocks. This provides some evidence to support the use of light rail systems as new business catalyzation or revitalization tools, even in historically auto-centric regions.

However, the second key finding of this study is that the automobile-oriented development areas have been making up the gap in new business creation at the rate of 8%, 6%, and 7% per year (in the knowledge, service, and retail sectors, respectively) since the light rail opened. This interesting finding follows the pattern of rail system impact discussed in previous research. There appears to be a much higher anticipated value or perceived benefit to a new rail transit system that catalyzes economic activity leading up to, or shortly following, its opening, which then dissipates as the ‘novelty factor’ wears off (Mohammad et al. 2013). There is not enough data yet to know whether, in the long run, the diminishing returns (in terms of new business starts) from the light rail might eventually erase the absolute advantage it conferred.

Certainly, as the system stabilizes, its agglomeration and visibility benefits may still foster a significantly higher number of new business starts than would have otherwise been the case, but simply at a lower level than was observed immediately after the line opened. This effect is observed in relation to the auto-oriented control areas and is not likely to indicate the ultimate demise of the economic well-being of the station areas.

The idea that perception is bigger than reality when it comes to the economic development potential of fixed-route transit systems is particularly illuminating for the urban transportation planning community, raising the obvious question of how to prevent or mitigate these drop-offs in new business activity. One solution suggested by the case of Phoenix – where the system serves a relatively small proportion of regional travel – is to ensure that new transit systems are truly competitive with other modes from a transportation service standpoint.

Ultimately, any economic benefits stem from the system’s usefulness as a mode of transportation, which requires a high quality of service, dependability, speed, low cost, and a wide range of potential destinations. If the rail system does not compete well in these facets, then it is likely that any economic benefits observed at its construction – investments dependent on the pure novelty of the mode rather than its practical benefit – will diminish over time. Thus planners concerned in maintaining the initial economic benefits from the system’s construction must work to ensure increasing service quality and usefulness of the line.

Finally, the third key result of this paper is that impacts to new business creation diminish with distance from the stations but are still significant for all three of the studied types at 1 mile

(network distance) from stations. In addition, the three types of businesses are impacted differently by distance from light rail stations: the benefits drop most quickly with distance for retail businesses (21%), while knowledge (12%) and service (0%) businesses do not drop as quickly. Taken together, these results suggest that TOD schemes may be most effective at fostering business formation in the knowledge sector, and that the closer they are targeted to the station, the more impactful they will be. This spatially-explicit information is also useful to station-area planners and entrepreneurs concerned with capitalizing on the economic benefits of the new system, as the results show that locations within ¼-mile of stations experience greater levels of new business investment. Planners looking to target station areas for special TOD zoning districts or other development-supportive policies (e.g., increased density) should look for opportunities as close to stations as possible and transition on a gradient outward, as far away as a mile.

Of course, there are several limitations to the approach used in this study: it does not directly consider the individual impacts of zoning, construction, and transit service on business

48 formation or trends across different metropolitan areas. Also, while this study measures the general social context for light rail planning in Phoenix, it does not provide a holistic account the multifarious ways in which social conditions influence the relationship between transit system development and new business activity. In addition, the statistical approach employed in this paper inevitably relies on a set of specific assumptions that simplify economic decision-making behavior and offer only quantitative results. Future extensions of this work could control for the impact of land use regulations, provide comparisons with other regions investing in new transit systems, elaborate on the social context for regional economic development, and provide additional explanation and ‘ground-truthing’ of these results, perhaps through a follow-up qualitative analysis of station-area land use characteristics and entrepreneurs. Finally, a useful theoretical extension of this paper could test the explicit impact of transit on fostering business and social ties and/or creative class workers, as suggested by the paper’s conceptual framework.

Methodologically, this paper provides a new application of the quasi-experimental AITS approach and neighborhood-scale spatial analysis, both of which have been missing from previous analyses of the economic impact of public transportation systems. The adjusted- interrupted nature of the analysis provides significant clarity on the direction of the relationship that is not possible in cross-sectional analyses. In addition, by employing point-level NETS data and explicitly considering a range of spatial scales, this paper develops a robust method that examines transit impacts to new business formation in fine spatial detail.

As for theoretical implications, the results of this paper confirm that there are indeed economic benefits for new firms due specifically to the construction of light rail transit and TOD.

This is an important addition to existing knowledge about the economic impacts of transit systems, because it shows that transit stations can, in fact, provide micro-agglomeration benefits

49 to businesses, and that these benefits vary by sector and diminish with distance. The long-term effects of these benefits, the specific mechanisms through which they operate, and their effect on larger economic factors like productivity, innovation, and economic diversification remain to be studied. However, this paper provides strong evidence for the idea that transit systems foster business development in adjacent neighborhoods. This information can be used by planners, policy makers, and entrepreneurs to maximize economic benefits around transit and to create sustainable transit-oriented economic development.

CHAPTER 2. TRANSITIVE PROPERTIES: A SPATIAL ECONOMETRIC ANALYSIS OF NEW BUSINESS CREATION AROUND TRANSIT.

Introduction

Interest in rail transit has increased in recent years, as concerns about sustainability, traffic congestion, and sprawl have come to the forefront of urban research (Ewing & Cervero,

2010; Richardson & Bae, 2016; Batty, Besussi, & Chin, 2003; Squires, 2002). North American cities have invested large sums on new transit systems and scholars have undertaken a vast research program to examine their impact on travel behavior, development, and property values in detail (Freemark, 2014; Ewing & Cervero, 2010; Mohammad et al. 2013). The economic impact of transit systems is particularly important to understand, given their large fixed cost and the desire of policy-makers to use transit systems as economic development and urban revitalization tools.

Existing research has identified a strong but nuanced link between rail transit and increasing property values (Ewing & Cervero, 2010; Mohammad et al. 2013; Golub,

Guhathakurta, & Sollapuram, 2012); however, the impact of transit on new business formation has been relatively less explored. Given the fact that transit increases accessibility, it has the potential to foster a number of benefits to businesses that locate near it: increased face-to-face contact, “weak” business ties with clients, labor market accessibility, higher consumer visibility

(for retail products), knowledge spillovers, and recruitment for younger workers who enjoy car- free living (Credit, 2017; Chatman & Noland, 2011). In fact, some researchers have suggested that certain kinds of knowledge and high technology businesses may be moving toward a transit- centric location model in order to take advantage of these economic benefits (and to avoid the negative externalities of increasing traffic congestion) (Weisbrod, Duncan, & Moses, 2014).

Studies that have begun to look at the link between rail transit and new business creation thus far have focused primarily on the construction of new transit systems in cities like Phoenix,

Dallas, and Portland (Credit, 2017; Chatman, Noland, & Klein, 2016). Less is known about the relationship between transit proximity and new business creation in regions with more established rail transit networks. The power of transit stations to attract new business investment well after construction is critical to understand, since transit systems represent large long-term fixed costs, and previous research on new transit construction has found that transit-related investment often peaks in the run-up to a new system’s construction, declining afterwards as the system becomes a common part of the regional transportation network (Mohammad et al. 2013;

Golub, Guhathakurta, & Sollapuram, 2012; Credit, 2017).

Thus the goal of this paper is to examine the relationship between transit station proximity and new business creation in five regions with varying levels of maturity in rail transit development and/or entrepreneurial ecosystems – Boston, San Jose, Austin, Cleveland, and

Philadelphia – using point resolution data. The relationship between proximity to transit – location within ¼ or ½ mile of stations – and new business starts is evaluated using a suite of spatial econometric models at the Census block level, with comparisons of effect made between each metro, mode (light rail, heavy rail, or commuter rail), and business type (all knowledge businesses, high technology, retail/services/food, and producer services).

The results of the analysis indicate that, using a spatially-responsive estimate of the expected density of new businesses per acre as the dependent variable, proximity to transit stations is significantly related to new business creation in a variety of contexts. As indicated by calculating the percentage of positive and significant rail coefficients across all regional models

(out of all possible rail coefficients), new businesses in Philadelphia and Boston – established

52 transit regions with more extensive networks, supportive land use, and generally better service – show the largest association with rail transit. Retail, services, and food businesses are most commonly significantly associated with transit proximity, while the commuter rail mode is most consistently associated with adjacent new business activity of any type. In addition, the use of a spatial Durbin model (compared to a non-spatial Poisson model with spatial lags of the dependent variable) changes the significance of transit proximity variables in 55% of instances, indicating that the use of non-spatial models may, in many cases, overestimate the positive impact of transit proximity on new business creation.

Literature

An extensive literature, stretching back to the 1970s, has examined the economic impacts of rail transit (Knight & Trygg, 1977; Cervero, 1984; Cervero, 1994). This research has generally found that property values increase with proximity to transit stations, especially if the service is good and the land use planning around the stations is done in a way that cohesively integrates new development with the transit station (Damm et al. 1980; Landis, Guhathakurta, & Zhang,

1994; Knaap, Ding, & Hopkins, 2001; Cervero, 2004; Agostnini & Palmucci, 2008; Weinberger,

2001; Weinstein & Clower, 2003; Golub, Guhathakurta, & Sollapuram, 2012). Generally, property value benefits are larger for commercial properties and commuter rail systems

(Mohammad et al. 2013). However, a significant vein of research challenge these assertions, arguing that transit systems simply reflect or refocus the economic growth of the region (Vessali,

1996; Giuliano, 2004; Schuetz, 2014), or that transit does not play a primary causative role in fostering walkability, dense development, and supplemental economic activity (Quinn, 2006;

Bollinger & Ihlanfeldt, 1997; Chatman, 2013). In addition, previous research has found evidence of a so-called ‘novelty factor’ for new transit systems - property values are often highest at the

53 time a new system opens (or right before opening), indicating that perception may be somewhat larger than economic reality for transit (Mohammad et al. 2013; Golub, Guhathakurta, &

Sollapuram, 2012).

As the impact of rail transit on property values has been the focus of the bulk of research on transit’s economic impacts, the relationship between transit and new business creation is comparatively understudied. Existing research has generally taken one of two approaches: cross- sectional multinomial logit analysis where the dependent variable is the type of business (or, in some cases, the type of industrial agglomeration) (Mejia-Dorantes, Paez, & Vassallo, 2012; Song et al. 2012), or time-series Poisson or negative binomial analysis where the dependent variable is the aggregated count of new businesses in a particular industry in a spatial unit (Credit, 2017;

Chatman, Noland, & Klein, 2016). In both cases, a measure of transit accessibility is used as the independent variable of interest. Given the relatively small body of existing literature, each of the relevant studies is examined here in some detail.

Mejia-Dorantes, Paez, and Vassallo (2012) show that proximity to a new Metrosur

(heavy rail) transit station in the Madrid, Spain region is a significant predictor of business activity for all industries, particularly for accommodation/food service and retail businesses.

Interestingly, proximity to the regional commuter rail service decreases the probability of businesses locating nearby, due perhaps to the negative externalities of an above-ground rail line.

In an analysis of regional Seoul, South Korea, Song et al. (2012) find industry-specific effects for subway accessibility – a village’s concentration in the construction, transportation, sales, accommodation/food service, finance, real estate, or other service industries are all significantly

(positively) related to subway accessibility, either directly in the village itself or through a neighboring village.

Similarly, Credit (2017) and Chatman, Noland, and Klein (2016) find that proximity to transit stations is generally a strong predictor of new business activity in the retail, service, information, and finance/insurance industries. Both papers use time series data to evaluate the post-construction impact of light rail systems in US regions that only recently built new rail transit systems: Phoenix (Credit, 2017), Portland, and Dallas (Chatman, Noland, & Klein, 2016).

In general, these studies also find that the impact of transit on new business creation declines as distance from the stations increases (from ¼ mile to 1 mile). By using a quasi-experimental study design, Credit (2017) also finds that the impact of transit proximity on new business creation is highest at the time of the system’s opening, declining steadily as time from opening increases.

This corroborates previous property value research that shows a similar ‘novelty factor’ for station-area property values (Mohammad et al. 2013; Golub, Guhathakurta, & Sollapuram,

2012).

From a methodological standpoint, advances have been made in recent years in the specification and estimation of spatial econometric models. The spatial models most commonly found in the literature – the spatial autoregressive lag (SAR) and the spatial error model (SEM) – are each designed to isolate a specific form of spatial interaction, either of which, if present, violates the independence assumption of ordinary least squares (OLS) regression (Anselin &

Rey, 2014). This is accomplished mathematically by inserting a spatial weights matrix directly into the estimation of the model (Elhorst, 2014; 2017). The SAR model contains an endogenous interaction effect (spatial dependence in the dependent variable), while the SEM accounts for spatial interaction contained in the error term (Anselin & Rey, 2014). While these models improve on non-spatial regression methods and generally do not pose theoretical issues with econometric estimation, recent work – aided by the development of new estimators – has argued

55 that these models are too simplistic (Elhorst, 2014; 2017). Due to this, focus in spatial econometrics has begun to shift to developing and using models that account for multiple types of spatial interaction: the SAC (containing both endogenous and error term interactions)

(Kelejian & Prucha, 1998; LeSage & Pace, 2009), spatial Durbin (SDM) (containing both endogenous and exogenous interaction effects, i.e., dependence from the neighboring independent variables that affects the dependent variable) (Anselin, 1988; LeSage & Pace,

2009), and spatial Durbin error models (SDEM) (containing endogenous, exogenous, and error effects) (Elhorst, 2014). While there remains a debate regarding the presence of identification issues in these more complex models (Anselin & Rey, 2014), if correctly specified, they have the potential to provide a richer estimation of spatial interaction effects than the older SAR and SEM approaches.

Given the state of existing research, several significant questions remain unanswered.

First, the presence of a novelty factor – and whether this influences the relationship between new business creation and transit stations in established transit regions – needs to be explored further.

Second, while previous research has examined the industry-specific impacts of transit proximity, these analyses have generally been done at fairly aggregated industrial scales. Given the theoretical links between transit accessibility and agglomeration benefits such as knowledge spillovers (through informal interactions), face-to-face contact and the construction of social trust, and the recruitment and marketing of businesses to a younger demographic (Chatman &

Noland, 2011; Credit, 2017), it has been hypothesized that knowledge-intensive businesses, such as high technology and producer services firms, might benefit particularly from proximity to transit (Weisbrod, Duncan, & Moses 2014). Despite the importance of these sectors to regional growth (O'hUallachain & Reid, 1991; DeVol & Wong, 1999; Chapple et al. 2004), no study has

56 yet looked at these fine-grained industrial categories, or compared their effects to more traditional business types, such as retail and services. Similarly, the explicit differences in effect between rail modes – light rail, heavy rail, and commuter rail – have not been explored in depth.

And finally, while existing studies have included some spatial variables in their modeling approaches to control for the effects of spatial autocorrelation, none of the existing papers have used an explicitly spatial econometric approach. While this is understandable given the fact that the models employed so far have been non-linear (thus incompatible with some assumptions of

OLS regression inherent for common spatial econometric models), the use of spatial models to explore the link between new business creation and transit proximity remains an important gap in the literature.

Data and Study Areas

Data

Three primary types of data are used in this analysis: information on transit systems, individual business data, and socio-demographic covariates provided by the Census. Table 3 describes the data and each of their sources in detail. The data on transit station locations by mode comes primarily from each region’s respective transportation authorities; in some cases it was provided as a geographic shapefile, and in others it had to be manually digitized (with corroboration from Open Street Map and other sources). Classification of the modes is based primarily on definitions provided by the American Public Transit Agency (APTA) (1994): light rail trains or trolleys are electrically-powered and often run in the street right-of-way; heavy rail transit is generally separated form traffic and designed for larger passenger capacities than light rail; and commuter rail is also grade separated but designed specifically to link a central business district with outlying residential areas.

Table 3. Data sources and descriptions.

Region/ Category Description Date * Source Scale

Type Description Digitized: Austin Commuter Rail Capital MetroRail 2010 https://www.capmetro.org/schedmap/?svc=2&f1=550 &s=0&d=N Commuter Rail Massachusetts Bay Transportation Authority (MBTA) Commuter Rail 2014 Downloaded: http://www.mass.gov/anf/research-and- Boston Heavy Rail MBTA Subway (Red, Orange, and Blue Lines) 2014 tech/it-serv-and-support/application-serv/office-of- Light Rail MBTA Trolley (Green Line) 2014 geographic-information-massgis/datalayers/mbta.html Heavy Rail Regional Transit Authority (RTA) Rapid Transit (Red Line) 2016 Digitized: Cleveland http://www.riderta.com/sites/default/files/gtfs/latest/g Light Rail RTA Rapid Transit (Blue, Green, and Waterfront Lines) 2016 oogle_transit.zip Southeastern Pennsylvania Transportation Authority (SEPTA) Regional Downloaded: Commuter Rail Rail 2009 https://www.arcgis.com/home/item.html?id=1cb6bfb

Transit 987c44859a2fffa7384cc5cd2 SEPTA High Speed Rail 2009 Downloaded: Philadelphia Heavy Rail Port Authority Transit Corporation (PATCO) Speedline 2015 https://njgin.state.nj.us/NJ_NJGINExplorer/DataDow nloads.jsp Downloaded: Light Rail SEPTA Trolley (10, 11, 13, 15, 34, 36, 101, and 102 Lines) 2009 https://www.arcgis.com/home/item.html?id=1cb6bfb 987c44859a2fffa7384cc5cd2 Downloaded: Commuter Rail Caltrain and Altamont Corridor Express Commuter Rail 2013 http://www.dot.ca.gov/hq/tsip/gis/datalibrary/Metadat San Jose a/RR_Commuter_13.html Downloaded: https://data.vta.org/Transit- Light Rail Valley Transportation Authority (VTA) Light Rail 2015 Operations/Stops-January-2015/iy3q-7kq5 All knowledge NAICS codes 51, 52, 54, 55 2011 NAICS codes 3254, 3341, 3342, 3344, 3345, 3364, 5112, 5161, 5179, High tech 2011 5181, 5182, 5413, 5415, and 5417 Business Producer services NAICS codes 5411, 5412, 5414, 5416, 5418, 5419, 5511 2011 National Establishment Time Series (NETS) NAICS codes 4431, 4451, 4452, 4453, 4461, 4481, 4482, 4483, 4511, Retail/services 4512, 4522, 4531, 4532, 4533, 4539, 8114, 8121, 8123, 8129, 8133, 2011 8134, 8139, 7211, 7213, 7223, 7224, 7225 Total population; racial diversity of block (computed using Herfindahl Block-Level 2010 Decennial Census Socio- index); size of block in acres demographic Percent Bachelor's degree attainment or higher; percent population 29 2008- Tract-Level American Community Survey years old or younger; average number of vehicles per household 2012 *For transit data, this year specifies the date the file used in the analysis was created; however, transit stations used in the analysis existed in 2010.

In order to analyze the impact of transit proximity on new business creation, ¼ and ½ mile buffers were calculated around each of the digitized transit stations. For the purposes of this analysis, a census block is considered “within” a given buffer if its centroid falls within the buffer.

The business data used in this paper comes from the National Establishment Time Series

(NETS), which provides information on industrial classification (by NAICS code), location, and year of opening (among other features). The underlying business data for NETS is furnished by the Dun and Bradstreet database and geocoded by Walls & Associates, providing a near census of business activity in a metropolitan area given year (Walls & Associates, 2012). For this paper, businesses that started in 2011 in four industries of interest – all knowledge, high tech9, producer services, and retail/personal services – were delineated and spatially joined (using ArcMap v.10.3) to the Census blocks in which they are located. Aggregated counts of new businesses10 in each industry serve as the dependent variable. 2011 new businesses were selected in order to provide at least a one-year lag between neighborhood characteristics (the Census covariates described below) and the opening of a new business, since the decision to open a business in a given location is most likely based on the observed characteristics of the neighborhood at least one year before the business actually opens. Due to the fact that NETS geocodes a business’ location to its last place of establishment, no business points that relocated were included in the analysis.

9 Hecker provides an often-cited definition of high-technology employment based on NAICS industries with high concentrations of high-tech employment (2005). The fourteen industries identified as “Level 1” are (by NAICS code): 3254, 3341, 3342, 3344, 3345, 3364, 5112, 5161, 5179, 5181, 5182, 5413, 5415, and 5417 (Hecker 2005). For the purposes of this paper, new businesses in these sectors are classified as high-tech start-ups. 10 Transformed, as described in the Methods section below, for use in the final model. 59

Socio-demographic data for 2010 were collected at two scales: the Census block and the

Census tract. Basic demographic information – including racial classification and total population – is provided at the block level from the Decennial Census. A Herfindahl Index was used to create a measure of racial diversity (where values closer to 0 indicate higher levels of diversity) at this scale. Unfortunately, the remaining socio-economic characteristics of interest are only provided at the tract level by the 2008-2012 American Community Survey (ACS), so these variables were joined to their nested Census blocks.

Study Areas

In order to ascertain the relationship between transit proximity and new high technology businesses in mature transit regions, five study areas with fixed rail transit systems operating in

2010 were selected to cover a full range of characteristics: regions with established entrepreneurial ecosystems but a relative lack of transit prominence, like San Jose, CA and

Austin, TX; regions with established transit systems but little historical entrepreneurial activity, like Cleveland, OH and Philadelphia, PA; and a region with both prominent entrepreneurial activity and a mature transit system with a supportive built environment, Boston, MA (Saxenian,

1994; Chapple et al. 2004; Richman, 2015). This choice of regions also allows for considerable modal variation: four regions (San Jose, Cleveland, Philadelphia, and Boston) have light rail systems, three (Cleveland, Philadelphia, and Boston) have heavy rail systems, and four (San

Jose, Austin, Philadelphia, and Boston) have commuter rail systems.

As for the technical delineation of these regions, the Census blocks for all counties in each Metropolitan Statistical Area (MSA) containing rail transit stations were used. This allows the paper to employ a regional approach (rather than biasing the sample towards transit use by selecting, e.g., only neighborhoods around transit stations) while limiting some

60 suburban/exurban bias that could crop up if the entire MSA definition were used, given that many of the transit systems examined here extend only into the inner-ring suburbs of their respective regions. It makes little sense to include areas that are truly transit-inaccessible – such as exurban or rural portions of an MSA – in this analysis, since businesses in these areas are highly unlikely to consider transit as a part of their location calculus.

Methods

The model specification for this paper involves some interesting trade-offs – using the individual business points themselves preserves the finest grain of spatial accuracy in relation to transit stations, but unfortunately it does not answer the research question at hand. If the unit of analysis were the businesses themselves, the paper would involve studying how characteristics of businesses influence location – for any given individual business – close to transit; rather, this paper is interested in what (if anything) influences higher numbers of businesses to be located near transit stations because this focus on neighborhood and regional determinants of new business creation is important for helping to guide local economic development and planning efforts (Renski, 2008; Malecki, 1984; Mack & Credit, 2016). To do this, a unit of aggregation must be selected: however, if the aggregation unit is too large (e.g., Census tracts), location within ¼ or ½ mile of a transit station loses its descriptive power. That is why this paper uses the smallest possible aggregation unit, Census blocks, in order to maintain as much spatial variation as possible in measuring proximity to transit. Covariate characteristics that can be measured at the block level are included at the block level (such as race, population, and median age), while other measures only available at the tract level are duplicated for nested blocks.

Of course, the choice of such a small aggregation unit comes with its own challenges.

Given the very fine spatial scale of blocks, in a given year (2011) there are a vast number of zero

61 new business counts. This presents a problem for statistical modeling similar to the “small numbers problem” common in the spatial epidemiology literature – for very rare events (e.g., cancer deaths), the spatial distribution observed in any one timeframe may not reflect the true underlying probability of the event occurring (Lawson et al. 2016). This has the potential to bias regression models constructed based on the observed data, since the very rare observed events will drive, perhaps unfairly, the model results. In these cases, smoothing functions are often employed in order to approximate the “true” underlying probability of an event occurring at any given location by adjusting expected event counts in some way (Kafadar, 1996; Clayton &

Kaldor, 1987; Lawson et al. 2016).

The simplest smoothing function is simply to calculate the raw rate,

푥푖 푧푖 = (2) 푦푖

where 푥푖 is the count of events of interest at location i (in this case, the count of new high technology businesses) and 푦푖 is the exposure variable (in this case, size of the block in acres).

This operation simply standardizes the raw count of events by some underlying exposure rate. Of course, the underlying probability of an event occurring at i may very likely be influenced by its neighbors. In this case, a “spatial rate11” can be calculated, which uses information from neighboring observations to create a smoothed rate (Kafadar, 1996):

푎푣푒푅{푥푖} 푧푖̃ = (3) 푎푣푒푅{푦푖}

where 푎푣푒 is an averaging function and푅is the neighborhood of observations around either 푥푖 or푦푖. In this case, 푎푣푒 is simply the average value for the queen-contiguous neighbors

11 This is the terminology used in GeoDa v.1.8.16.4 1, which was used to calculate all of the smoothed variables used in the paper. 62 of a given block, computed using a spatial weights matrix. An even more detailed smoothing procedure is the Spatial Empirical Bayes (SEB) estimator, which “shrinks” the expected value in a given location towards the mean of its neighborhood’s rate – the “prior” distribution in this case – based on the size of the variance of the raw rate at a given location, thus limiting the impact of observations with extremely high variance (Clayton & Kaldor, 1987; Anselin, Kim, &

Syabri, 2004):

휋̃푖 = 푤푖푧푖 + (1 − 푤푖)휃, (4)

where

∅ 푤푖 = 휃 . (5) ∅+( ) 푦푖

Here, 휃 is the mean and ∅ is the variance of the prior distribution (in this case, the queen- contiguous neighbors of location i). As Anselin, Kim, and Syabri (2004) point out, when the

휃 exposure variable is large, nears 0, which means that 푤푖 moves toward 1, pushing nearly all of 푦푖 the weight to the raw rate (푧푖). Thus the SEB estimate uses both the mean and the variance of the spatial neighborhood to weight a given observation, producing a more-informative expected rate estimate. As described in the section “Exposure Variables for New Business Creation” below, each of these smoothing techniques were calculated for each business type, producing expected rates of new business creation, which were then used as the dependent variable in a series of linear models comparing the relative performance of each smoothing technique in order to find the best method for the final model specification. Figure 7 shows this process diagrammatically: first, different spatial and non-spatial smoothing techniques are tested for each exposure variable.

Then, the diagnostics of OLS models run using the best-performing smoothers for each exposure

63 variable are compared to find the most consistent form of the dependent variable for each business type. These dependent variables are used in a suite of pooled spatial econometric models to find the best-fit spatial model form; once the final model specification is selected, region-specific models are run and compared in order to compare and contrast results across regions.

Figure 7. Analysis framework for testing rate smoothing and spatial econometric approaches.

Modeling Spatial Count Data

In standard econometric practice, generalized linear models (GLM) such as Poisson or negative binomial models are chosen to estimate count data, given the fact that the dependent

64 variable is modeled as the expected outcome of a Poisson distribution, which is non-negative and has a mean equal to its variance12 (Hilbe, 2014; Faraway, 2006).

However, the use of models that directly incorporate a spatial dependence structure into the model specification in the form of a spatial weights matrix is suggested where data are spatially autocorrelated, as is the case with new businesses (Anselin & Rey, 2014). There are three main approaches to estimating GLM models that take into account underlying spatial dependence in the data. The simplest is to include spatially-lagged versions of the dependent and/or independent variables in a non-spatial Poisson or negative binomial model in order to control for spatial dependence. Since these models do not directly incorporate spatial dependence into the covariance structure of the data, they are the least favored.

Spatial Generalized Linear Mixed Models (GLMM), on the other hand, add a random effects parameter to the linear predictor of the Poisson model, which is assumed to be normally- distributed and whose covariance is calculated using a spatial weight matrix (Liu, Davidson, &

Apanasovich, 2007). There is considerable complexity, however, in estimating the model: maximum likelihood estimation results in an n-dimensional integral (where n is the number of observations) that cannot be reduced due to the spatial correlation between the observations, which is difficult to integrate with any large n. Bayesian hierarchical spatial models that employ

Markov Chain Monte Carlo methods to simulate samples of the posterior distribution of the regression parameters and the covariance structure of the Gaussian random field are more widely used, but the impact of data composition on the model, long computation times, and the fact that basic statistical properties of these models are still not fully understood continue to be issues for

12 The negative binomial model is a special case of the Poisson model (with an additional parameter) that is used to model overdispersion in a Poisson distribution, i.e., when the variance of the distribution is greater than its mean (Hilbe 2014). 65 implementation (Diggle & Ribeiro, 2007; Diggle, Tawn, & Moyeed, 1998; Christensen, Roberts,

& Sköld, 2006; De Oliveira, 2013; Jing & De Oliveira, 2015). Other composite likelihood approaches reduce the computational effort for estimation (Liu, Davidson, & Apanasovich,

2007), but even so, spatial GLMM approaches remain challenging to use for many spatial econometric applications.

The third approach is spatial filtering, which uses the eigevnvectors of the data’s spatial relationships (via a spatial weights matrix) to systematically remove residual dependence in the model (Tiefelsdorf & Griffith, 2007; Wang, Kockelman, & Wang, 2013). Since this approach produces a set of spatial eignevectors that can then be added to any model type – including

Poisson or negative binomial models – this approach can be used to eliminate spatial dependence in models of count data. However, since the number of eigenvectors calculated equals the number of observations – from which the top vectors are identified to include in the final model

– the spatial filtering approach can involve lengthy computation times (Wang, Kockelman, &

Wang, 2013).

Given the fact that some computational issues persist with each of these approaches to modeling spatial count data, this paper presents an alternative method for analyzing business count data in a traditional spatial econometric framework13 by using rate smoothing approaches

(mentioned above) to help alleviate the problems of non-normality and small numbers. Log10- transforming the expected counts provided by a ‘best-fit’ smoothing technique then provides a normally-distributed variable that can be used for standard spatial econometric model estimation.

13 It is important to note that this paper does not attempt to argue that the method used here is statistically preferable to either the spatial GLMM or the spatial filtering approaches for modeling spatial count data – a direct comparison of these methods is beyond the scope of this paper. Rather, this study simply presents an alternative method for modeling business data within traditional spatial econometric frameworks. 66

In order to choose a suitable smoothing function, however, an exposure variable for new business creation must be chosen. While various theoretical arguments could be made to support the use of area, population, the number of existing businesses, or the number of all new businesses in a block as a suitable background rate – for instance, businesses driven by consumer demand might logically employ population as an exposure measure – this paper is interested in empirically testing these measures in order to see which performs best in a set modeling framework.

Exposure Variables for New Business Creation

In order to empirically evaluate the logical choices of background exposure variables to construct a suitable dependent variable for use in a spatial regression, the best-performing method from each of the four possible exposure variables is chosen – along with the log transformed raw count - to construct four regressions in order to evaluate the sensitivity of the results to changes in the exposure variable. The full range of possible exposure variables and rate smoothing techniques are shown in Table 4, along with some of their basic characteristics.

For each business type, the smoothing technique14 with the greatest number of non-zero observations and/or transformed normality for each of the four possible exposure variables are chosen for testing. Log10-transformed version of these four variables then become the dependent variables in four OLS regressions, the comparative diagnostics of which are shown in Table 5.

Each of the models is compared based on a range of diagnostics, including adjusted R2 and residual standard error (for overall model fit), multicollinearity condition, Breusch-Pagan and

Koenker-Bassett tests (for heteroskedasticity), and comparison of the signs and significances of

14A first-order queen spatial eights matrix was used to calculate the spatial rate and SEB smoothing rates. 67 the coefficients with a Poisson regression run with the same specification and the exposure variable added as an offset.

As Table 5 indicates, spatial rate smoothing using area as the exposure variable is shown to be the best-performing method for knowledge, retail/services/food, and producer services businesses. For high technology businesses, the SEB smoothing technique with area as the exposure variable performed best, perhaps due to the very small number of observations for that business type. Beyond the quantitative diagnostic advantage of the models that use area as the exposure variable, the area-smoothed dependent variables in these models also simply make greater qualitative sense: the expected value of new business density is an easier concept to pin down in reality than “the expected proportion of new businesses to population” in a block. All of this suggests that area provides the most stable exposure rate from which to measure new businesses, which has useful implications for future spatial econometric research on new business creation.

Spatial Regression Specification

With the proper exposure variable and smoothing technique for each type of new business chosen, the specification for the full spatial regression models can be determined. Given the large number of spatial econometric models available – and the lack of prior theoretical knowledge about the correct model to choose – this paper compares six types of pooled models

(containing observations for all five regions), selecting the best-performing model for use in the final specification.

Business Method Exposure Variable Non-Zero Distribution (after Type Observations Log(10) Transformation) Raw count None 14687 Right-skewed Raw rate Acres 14687 Normal Spatial rate Acres 95889 Normal Spatial Empirical Slightly left- Acres 95889 Bayes skewed Raw rate All existing businesses in 2011 14687 Normal All existing businesses in Slightly right- Spatial rate 95889 2011 skewed Knowledge Spatial Empirical Slightly left- (NAICS 51- All existing businesses in 2011 9171 52 and 53- Bayes skewed 54) Raw rate 2010 population 12712 Normal Spatial rate 2010 population 93792 Right-skewed Spatial Empirical 2010 population 14127 Normal Bayes Heavily left- Raw rate All new businesses in 2011 14687 skewed Spatial rate All new businesses in 2011 95889 Left-skewed Spatial Empirical All new businesses in 2011 196 Left-skewed Bayes Raw count None 2962 Right-skewed Slightly left- Raw rate Acres 2962 skewed Spatial rate Acres 26721 Normal Spatial Empirical Acres 26725 Normal Bayes Raw rate All existing businesses in 2011 2962 Normal All existing businesses in Spatial rate 26721 Normal 2011 Spatial Empirical High All existing businesses in 2011 2290 Normal technology Bayes Raw rate 2010 population 2381 Normal Slightly right- Spatial rate 2010 population 25735 skewed Spatial Empirical 2010 population 3364 Normal Bayes Heavily left- Raw rate All new businesses in 2011 2962 skewed Spatial rate All new businesses in 2011 26725 Left-skewed Spatial Empirical All new businesses in 2011 79 Left-skewed Bayes

Table 4 (cont’d).

Heavily right- Raw count None 8044 skewed Slightly left- Raw rate Acres 8044 skewed Spatial rate Acres 62240 Normal Spatial Empirical Acres 62240 Left-skewed Bayes Slightly left- Raw rate All existing businesses in 2011 8044 skewed All existing businesses in Spatial rate 62240 Normal Retail, 2011 services, and Spatial Empirical Slightly left- All existing businesses in 2011 5939 food Bayes skewed Slightly right- Raw rate 2010 population 6727 skewed Spatial rate 2010 population 60755 Right-skewed Spatial Empirical 2010 population 8656 Normal Bayes Heavily left- Raw rate All new businesses in 2011 8044 skewed Spatial rate All new businesses in 2011 62240 Left-skewed Spatial Empirical Slightly left- All new businesses in 2011 135 Bayes skewed Heavily right- Raw count None 8936 skewed Slightly left- Raw rate Acres 8936 skewed Spatial rate Acres 67139 Normal Spatial Empirical Acres 67139 Normal Bayes Raw rate All existing businesses in 2011 8936 Normal All existing businesses in Slightly right- Spatial rate 67139 2011 skewed Producer Spatial Empirical Slightly left- services All existing businesses in 2011 6695 Bayes skewed Raw rate 2010 population 7847 Right-skewed Spatial rate 2010 population 65850 Right-skewed Spatial Empirical 2010 population 9817 Normal Bayes Heavily left- Raw rate All new businesses in 2011 8936 skewed Spatial rate All new businesses in 2011 67139 Left-skewed Spatial Empirical All new businesses in 2011 166 Normal Bayes Note: there are 227140 total observations in the data set.

Table 5. Comparison of model results for the four “best” method/variable combinations for each of the four dependent variables of interest.

Other Factors Business Smoothing N Adj. Residual Multicollinearity Breusch- Koenker- Variables Unexpected Overall Pros Cons Model Type Rates R² Std. Condition Pagan Bassett with sign for Score** Selection Error changing significant significance* variables Normality, intuitive Spatial smoothed PCTBACH, Rate - 95,889 0.2 0.555 12.296 7230.916 6503.683 None 11 denominator, None *** POP Acres large number of observations Spatial LR25, HR25, Large number Rate - CR5, CR25, HR5, Slightly right- Knowledge 95,889 0.045 0.283 11.983 1492.607 1152.488 9 of Existing VEHPERHH, POP skewed (NAICS observations 51-52 and businesses NS_ALL 54-55) LR25, Normality, Low number Raw Rate - PCTBACH, LR5, 12,712 0.154 0.509 15.165 8762.823 3336.833 7 intuitive rate of Population VEHPERHH, EXYR11 transformation observations ACRES Normality, Spatial HR25, CR25, LR25, LR5, large number Rate - New 95,889 0.04 0.253 12.208 1369.914 1488.488 RACE_DIV, HR5, 9 Left-skewed of businesses VEHPERHH EXYR11 observations Spatial Normality, Empirical PCTBACH, intuitive Complex rate 26,725 0.211 0.589 12.755 870.078 521.456 POP 12 *** Bayes - EXYR11 smoothed transformation Acres denominator Spatial CR25, LR5, Rate - CR5, LR25, HR5, 26,721 0.108 0.334 12.496 2511.156 1703.268 9 Normality None Existing VEHPERHH, NS_ALL businesses ACRES, POP High LR5, technology Spatial LR25, HR5, RACE_DIV, Slightly right- Rate - 25,735 0.121 0.498 12.437 3947.822 1050.356 7 None CR5, ACRES VEHPERHH, skewed Population EXYR11

HR5, CR5, Spatial PCTBACH, LR25, LR5, Rate - New 26,725 0.13 0.338 12.664 5511.761 4567.508 ACRES, RACE_DIV, 6 None Left-skewed businesses POP, VEHPERHH EXYR11

Table 5 (cont’d).

Normality, intuitive Spatial Rate - PCTBACH, smoothed 62,240 0.22 0.541 12.103 2463.253 2277.16 POP 11 None *** Acres EXYR11 denominator, large number of observations LR25, HR25, CR25, LR5, Spatial Rate - Normality, HR5, Existing 62,240 0.06 0.306 11.785 5096.93 3803.175 NS_ALL 8 large number of None Retail, PCTBACH, businesses observations services, VEHPERHH, and food POP Spatial Complex rate Empirical HR25, LR25, HR5, transformation, 8,656 0.1 0.37 14.675 2396.023 670.819 6 Normality Bayes - ACRES EXYR11 low number of Population observations LR25, HR25, Spatial Rate - LR5, HR5, Large number New 62,240 0.091 0.296 11.998 6787.085 6784.622 RACE_DIV 11 Left-skewed VEHPERHH, of observations businesses ACRES, POP Normality, intuitive Spatial Rate - PCTBACH, smoothed 67,139 0.21 0.539 12.308 5076.311 4608.52 EXYR11 11 None *** Acres POP denominator, large number of observations LR25, LR5, CR5, Spatial Rate - RACE_DIV, HR25, CR25, Large number Slightly right- Existing 67,139 0.059 0.298 11.992 2274.872 1682.242 PCTBACH, 9 HR5 of observations skewed Producer businesses VEHPERHH, services POP, NS_ALL Spatial HR25, HR5, Low number of Empirical CR5, LR25, LR5, observations, 9,817 0.122 0.366 15.254 1848.597 670.816 8 Normality Bayes - VEHPERHH, EXYR11 complex rate Population ACRES transformation LR25, HR25, Spatial Rate - CR25, HR5, LR5, Large number New 67,139 0.08 0.295 12.198 6057.737 6151.06 CR5, 8 Left-skewed PCTBACH, of observations businesses VEHPERHH POP *Based on comparison with non-spatial Poisson regression run in R using all observations and the corresponding background population measure (i.e. acres, existing businesses, population, new businesses) as the offset variable. Variables chosen for regression based on correlation analysis using the full dataset. **Scored based on the following scale: 3 points for best value for a given diagnostic, 2 points for second, 1 point for third, and 0 for worst. Cells are shaded accordingly.

To do this, non-spatial OLS models are first run with a 7 nearest-neighbor spatial weights matrix15 and robust Lagrange Multipliers are calculated (shown in Appendix B.1) to provide an initial indication of model choice (Anselin & Rey, 2014). Then, SAR, SEM, SDM, SDEM, and

SAC models are run in succession16 with AIC and adjusted R2 values calculated for each

(Sparks, 2015). The graphed AIC values for each model (for each business type) are shown in

Figure 8.

In each case, both robust Lagrange Multipliers are highly-significant (though the lag values are generally lower), and the SDM displays the lowest AIC. Given this, the SDM is used for the final specification for the regional regressions used in the paper. Use of a spatial lag-type model also makes sense from a theoretical perspective, as new knowledge and high technology businesses are expected to exhibit spatial dependence based on spillover activity, due to knowledge spillovers, information exchange, and other agglomeration factors (Aharonson et al.

2013; Aharonson, Baum, & Feldman, 2007; Gilbert, McDougall, & Audretsch, 2008).

The specification for the regional spatial Durbin models used in this paper is:

푌 = 훿푊푌 + 푋훽 + 푊푋휃 + 휀, (6)

where 푌 is a vector of observations of the expected density of new businesses per acre in a given region, 푊푌 is the spatially-lagged dependent variable (based on a 7 nearest-neighbor spatial weights matrix, 푊), 훿 is the spatial autoregressive parameter, 푋 is a matrix of exogenous

15 Spatial weights matrices for estimating the spatial Lagrange Multipliers (LM) in these equations used 7 nearest neighbors (Euclidean) of a given observation. This value was chosen based on the average number of contiguous neighbors for the full, contiguous dataset. Since blocks with zero counts cannot be log-transformed – and are not generally of interest for understanding the phenomenon at work (see Acs, Anselin, & Varga, 2002 for a similar decision regarding models of patent counts) – they were removed from the dataset. This however means that there are spatial gaps in the fabric of blocks, which necessitates using a distance-based rather than a contiguity weights matrix. 7 nearest-neighbor weights matrices were also used for the remaining spatial models in this paper. 16 All models were estimated using the “lagsarlm” or “errorsarlm” functions in the “spdep” R package using the Monte Carlo approximate log-determinant method of weights matrix decomposition. 73 covariates (including dummy variables for location within ¼ mile and from ¼ to ½ mile of a transit stations of particular modes), 푊푋 is vector of spatial lags of the independent variables, 휃 is the vector of regression coefficient for these lagged independent variables, and 휀 is the error term (Elhorst, 2014; Sparks, 2015).The basic idea of the spatial Durbin model is that, by including lags of the independent (as well as the dependent) variables, the model can more fully account for residual spatial autocorrelation while also providing estimates of the influence that neighboring values of the independent variables have on the expected density of new business starts. The independent variables included in the final specification are station proximity, racial diversity, percent bachelor’s degree attainment or higher, vehicles per household, population, total number of existing businesses, and total number of all new businesses

It is also important to note that in spatial lag-type models it is necessary to apply scalar summary measures to the regression output, i.e., to calculate the average direct and indirect effects to obtain a value of total partial effects comparable to a non-spatial regression coefficient

(LeSage & Pace, 2009; Elhorst, 2014). This is necessary in spatial lag-type models because a change in the value of an independent variable in a given unit changes the value both of 1) that unit’s own dependent variable (the direct effect), as well as 2) the value of neighboring units’ dependent variables (the indirect or spillover effect). A standard regression coefficient in a non- spatial model necessarily contains both of these effects because there is no accounting for spatial interaction. These effects were calculated for each regional model using the “impacts” function in the “spdep” package in R.

Figure 8. Comparison of AIC for each tested pooled model type (for each business type dependent variable). Lowest AIC shown in red circle.

Results

Descriptive Statistics and Spatial Patterns

Appendix B.2 shows the combined descriptive statistics for the variables used in the spatial models for all regions. In general, it illustrates the low number of average new business starts for all types, as well as the number of blocks located within walking distance to transit stations. On average, only around 5% to 6% of blocks in the model are within ½ mile of any kind of rail transit station; commuter rail stations have a bit more reach, with half-mile buffers covering over 9% of selected blocks. The average block in the dataset has a fairly high educational attainment – 36% bachelor’s degree attainment or higher – and a generally average level of vehicle ownership (1.74 vehicles per household).

Figure 9 shows the broad pattern for expected new knowledge business density and transit station location in each of the five regions of interest at a similar spatial scale. While it is difficult to make conclusive statements about spatial patterns from looking at this figure, a few things stand out: 1) Philadelphia and Boston’s transit systems appear to show much more extensive coverage of blocks with high levels of new knowledge start density (in part due to the larger number of stations in those cities); 2) the pattern of top knowledge blocks and transit stations are seemingly mismatched in Austin and Cleveland; and 3) San Jose appears to have a very large number of the top knowledge blocks (as one might expect), but these also appear to be generally mismatched with the light rail system in particular, which runs along the periphery of the new business-heavy portions of the valley.

Regression Results

Table 66 shows the summarized significant, transit-specific results of the pooled and regional spatial Durbin models for each business type of interest; detailed regression results for 76 each of the 24 individual models can be found in Appendix B.3-6. Given the large number of models to compare and the relative heterogeneity of results across regions, the approach used in this paper is to use the regional model results to find the percentage of positive/significant total effects coefficients17for each category of interest (out of all possible positive/significant results).

These percentages, as indicators of consistent association, are then used to compare across the categories of interest for the study: business type, mode, and region. In addition to this analysis of the regional models, the pooled model results are also used to provide indications of overall trends by business type and mode.

In terms of business type, both the pooled and regional models indicate that new retail, food, and services businesses are the most consistently associated with all types of rail transit stations, with 45% positive/significant coefficients across the regional models (and positive/significant results for every transit variable in the pooled model). All knowledge businesses are the second-most consistently associated with rail transit (at 41%), followed by high technology (36%) and producer services (27%) businesses. While regional heterogeneity is evident in these relationships, this finding generally supports the idea that retail and services businesses rely heavily on visibility and pass-by traffic, while also providing evidence that knowledge businesses benefit from transit proximity in a variety of ways (Credit, 2017;

Chatman, Noland, & Klein, 2016).

17 Significance measured at the p ≤ .05 level. 77

Table 6. Table summarizing pooled and region-specific Spatial Durbin model significant total effects for mode and business type.

Total Regional Overall Avg. by Business Type Rail Mode Distance Pooled San Jose Austin Cleve. Phila. Bost. Positive Percent Mode 0.25 + - + + 2 Light Rail 25% Light Rail 28% 0.50 0 0.25 + 0 Heavy Knowledge Heavy Rail 17% 41% 25% 0.50 + + 1 Rail Commuter 0.25 + + + + 3 Commuter 75% 56% Rail 0.50 + + + + 3 Rail

Light Rail 0.25 0 13% 0.50 + 1 High 0.25 + 1 Heavy Rail 50% 36% technology 0.50 + + + 2 Commuter 0.25 + + + 2 50% Rail 0.50 + + + 2 0.25 + - + + 2 Light Rail 38% 0.50 + + 1 Retail, food, 0.25 + + 1 Heavy Rail 33% 45% and services 0.50 + + 1 Commuter 0.25 + + + + 3 63% Rail 0.50 + + + 2 0.25 + - + + 2 Light Rail 38% 0.50 + 1 Producer 0.25 0 Heavy Rail 0% 27% services 0.50 0 Commuter 0.25 + + + 2 38% Rail 0.50 + + 1 Total Positive 2 1 2 16 12

Regional Percent 13% 13% 13% 67% 50%

As for modal variations, commuter rail is most consistently associated with transit proximity in both the pooled and regional models, corroborating previous research on property values that found larger gains accrue to properties near commuter rail stations than other modes

(Mohammad et al. 2013). Perhaps the generally larger regional mobility offered by commuter rail rather than heavy or light rail accounts for this result; it is also possible that, by their nature as “commuter” transit, these rail stations represent generally larger investments (in terms of size and design) and/or tend to be located in more desirable locations for business development, e.g., suburban locations.

The results also show interesting regional variations in the transit—new business relationship: new businesses of all types in Philadelphia and Boston are most consistently associated with transit (of all modes), at 67% and 50% positive/significant coefficients, respectively. While the results of this analysis do not lend significant insight into its underlying causes, it is possible that the more extensive, connected, mature – and thus useful and visible – nature of the transit networks in these regions increases their attractiveness for new business development. Historic, walkable land use patterns that make everyday access to transit easier also may play a role. These findings also begin to shed some light on the question of the ‘novelty factor’ for transit development – it appears that the association between transit proximity and new business development can remain significant long after construction in regions with long- established, mature transit systems.

Given these results, the final question of relevance is how the pooled and regional spatial

Durbin models presented here compare to the type of non-spatial Poisson regressions commonly used in the literature to this point (Credit, 2017; Chatman, Noland & Klein, 2016).

Figure 9. Census blocks in top 25% for Log10(expected new knowledge business density) and rail transit buffers for all modes and regions.

To do this, pooled and regional Poisson regressions that approximate the spatial models specified in equation (6) for each of the four dependent variables of interest are run; full results can be found in Appendix B.7. These models (by design) use the raw count of new businesses per block with acres as the offset and include a lagged dependent variable to control for lag-type spatial autocorrelation.

The signs and significances (at the p ≤ .05 level) for each of the transit variables in both sets of models are then compared in Table 77; if, for instance, the ¼ to ½ mile light rail transit variable for knowledge businesses was positive and significant in both the pooled spatial Durbin model and the Poisson model, it is marked as “+ Both”; if both models indicated negative significant coefficients, that variable was marked as “- Both”, and “Not sig.” for variables insignificant in both sets of models. More interesting that these corroborations, however, are the instances where the Poisson models show a positive/significant result while the variable is not significant in the spatial models (“+ Poisson”); given the increased information available in the spatial models, we can confidently think of these instances as Type I errors; Type II errors, then, are instances where the spatial models show a positive/significant result that is not captured in the Poisson models (“+ Durbin”, which occurs much less frequently). There are also a small number of instances where both types of models show significant results with opposite signs, which are categorized as “Mixed” results.

Table 7. Table summarizing differences between Spatial Durbin and Poisson model results, i.e., Type I and Type II Errors.

Corroboration Type I Error Type II Error Unknown + Both - Both Not sig. + Poisson - Poisson + Durbin - Durbin Mixed Light Rail 13 0 1 23 0 0 0 3 Rail Mode Heavy Rail 10 0 1 20 0 1 0 0 Commuter Rail 26 0 0 14 0 0 0 0 Knowledge 14 0 0 13 0 0 0 1 High tech 10 0 1 16 0 1 0 0 Business Type Retail 16 0 1 10 0 0 0 1 Producer Services 9 0 0 18 0 0 0 1 Pooled 17 0 0 7 0 0 0 0 San Jose 2 0 1 12 0 0 0 1 Austin 1 0 0 7 0 0 0 0 Region Cleveland 1 0 1 11 0 1 0 2 Philadelphia 16 0 0 8 0 0 0 0 Boston 12 0 0 12 0 0 0 0 Total 49 0 2 57 0 1 0 3 Percent 46% 51% 4% 3%

While there is some corroboration between the spatial and non-spatial approaches (46%), the results show that 51% of the positive/significant coefficients that appear in non-spatial

Poisson models are Type I errors, while 4% are Type II errors. Table 7 shows this percentage broken down by mode, business type, and region – each of those (large) rows conveys the same total counts but shows how the Type I and Type II errors are distributed across those different categories of interest. Perhaps most interestingly, in San Jose, Austin, and Cleveland, the number of Type I errors is larger than the number of corroborations, indicating that non-spatial models may significantly overestimate the relationship between transit and new business creation in regions with less mature transit systems.

Conclusions and Discussion

Based on these results, three primary conclusions can be drawn about the relationship between new businesses and transit in a variety of regions. First, the results show that proximity to rail transit stations does, in fact, have a positive overall relationship with new business starts, even while controlling for several forms of spatial dependence, total existing and new business activity in the block, and other socio-demographic factors. New retail, services, and food business and knowledge businesses are most consistently associated with rail transit variables. At the same time, commuter rail shows the most consistent association with adjacent new business creation of all types, and regions with extensive transit networks and a dense, historic urban fabric – Philadelphia and Boston – show the most consistent association between new business creation and transit proximity.

Second, empirical testing of the performance of different possible “exposure” variables for new business creation – including area, total population, existing business activity, and all new business activity – shows that area provides the most consistent, stable foundation for

83 calculating expected rates of new business activity. It also provides the most easy-to-interpret measure; talking about “expected new business density” is much clearer and easy to visualize than “expected new business creation out of all existing businesses”, for example. As for smoothing functions, spatial smoothers (including spatial rate and Spatial Empirical Bayes) create the most effective estimates of expected new business activity, balancing coverage

(expanding the number of observations) with model performance. Testing a range of spatial econometric models also indicates that the spatial Durbin model provides the lowest-AIC specification.

Third, this paper shows that spatial econometric models are necessary for estimating the relationship between aggregated new business starts and transit proximity – comparisons made between spatial model results and similar non-spatial Poisson models indicate that non-spatial models overestimate the significance of rail transit proximity 51% of the time. These Type I errors are particularly common in regions with less established transit networks. This indicates that previous work on the new business – transit connection may significantly overestimate the effect of transit proximity, perhaps with negative policy implications.

Of course, this paper has several limitations as well. It does not include direct test of other approaches to modeling spatial count data, including spatial GLMM and spatial filtering – a useful extension of this work could directly compare the performance of those approaches to the current one (in terms of computation time, control for spatial dependence, bias in parameter estimates, etc.) in order to see which is most useful for regional science practitioners and researchers. In addition, this paper does not capture causal relationships or evaluate trends over time; future work could expand the approach used here with panel data in order to better evaluate whether the observed relationships change over time.

Despite these limitations, this analysis provides a useful addition to literature on the economic impacts of transit. While previous work has shown a strong connection between knowledge business creation and transit proximity for relatively new light rail systems (Credit,

2017; Chatman, Noland & Klein, 2016), this paper shows that transit proximity also has a significant positive relationship with a range of new business types. These findings could have important implications for urban planners and policy-makers evaluating the economic costs and benefits of creating or extending new rail transit systems. New transit systems can certainly play a role in catalyzing new business development but appears more likely to do so if the systems are extensive, connect employment centers, and are supported by walkable urban environments.

Equally interesting are the implications of this paper for future research. By empirically testing different exposure variables for new business creation, this analysis shows that area provides the most stable background from which to calculate new business creation rates. In a theoretical sense, this also suggests that new business creation is a process influenced more by density than demand (population) or co-location with (all) other types of businesses. This paper also provides a novel method for modeling spatial count data within traditional spatial econometric frameworks, and demonstrates the importance of using spatial models in transit-new business research; without such models, the relationship between transit and new business creation is likely to be vastly overestimated.

CHAPTER 3. PLACE-MAKING AND PERFORMANCE: THE IMPACT OF WALKABLE BUILT ENVIRONMENTS ON BUSINESS PERFORMANCE IN PHOENIX AND BOSTON. Citation: Credit, K., & Mack, E. (2017). Place-Making and Performance: The Impact of Walkable Urban Design on Business Performance in Phoenix and Boston. Environment and Planning – B. DOI: 10.1177/2399808317710466.

Introduction

Urban design and place-making are linked via the opportunities good urban design creates for people to interact with one another and the urban environment (Knox, 2005). Built environment characteristics of urban spaces such as block length, street network layout, building scale, and age provide opportunities for people to interact with one another and explore urban environments on foot. While prior work has acknowledged the importance of urban design to place making and the slow city movement (Knox, 2005; Mayer & Knox, 2006), and urban design as a critical facet of street traffic and the patronization of third places (Knox, 2005), few studies have evaluated how built environment features might also enhance business performance. This is important to consider, because it suggests there are both aesthetic and economic benefits to good urban design.

Previous work on the economic value of good urban design has evaluated the linkages between walkability and property values. (Leinberger & Alfonzo, 2012; Pivo & Fisher, 2011; Li et al. 2014). While valuable, these studies do not consider other potential economic impacts of urban design such as employment, establishment growth, sales tax receipts, or sales volume

(NYCDOT, 2013; Hass-Klau, 1993). As regards the benefits of urban design to businesses, prior studies have hypothesized that compact, walkable urban environments with a diversity of people and businesses facilitate pedestrian activity to create “effective economic pools of use” (Jacobs,

1961, p. 171). This refers to increased foot traffic and window shopping that is beneficial to

86 businesses in the increased patronization of stores, restaurants and cafes. It has also been hypothesized that urban design practices that emphasize walkable urban forms are likely to attract members of the creative class who prefer walkable, mixed use urban spaces to minimize commute times between work and leisure activities (Florida, 2002). Mixed use, walkable environments are also likely beneficial to businesses that employ working-class employees with more limited transportation and employment choices. Unfortunately, there is virtually no information about the link between businesses (which represent one aspect of urban activity) and good urban design.

To address this research need, the goal of this study is to analyze the linkages between built environment aspects of urban design with business performance, as measured by sales volume per employee. Specifically, hierarchical linear models (HLM) are estimated to analyze neighborhood scale features of the built environment (BE) – as characterized by Jacobs’ “four generators of diversity” – and their relationship to business performance. The overarching hypothesis of this study is that the same BE characteristics that promote pedestrian activity will also positively impact business performance. From a theoretical perspective, this is an important yet unassessed dimension of the economic value of good urban design to communities. From a practical perspective, an evaluation of this benefit to urban design will provide important information to planners and economic development practitioners that can enhance their efforts to design economically vibrant places with aesthetic appeal and a sense of place.

The analysis shows that certain features of walkable built environments are positively associated with business performance. However, the relationship between walkable built environments and business performance varies considerably depending on the type of business and city-level context being studied, indicating that significant nuance must be used when

87 considering place-based interventions. Although no causal statements can be made about the built environment and business performance, the results of this paper indicate that (in some contexts) design-based place-making initiatives could be used to generate sustainable local economic development. This provides a welcome alternative to investing in the risky zero-sum game of inter-urban competition for branch plant relocation using traditional economic incentives.

Perspectives on Good Urban Design

Classic theories of urban design emphasize the importance of features including imageability, mixed land uses, short block length, spatial continuity, and human-scaled design

(Jacobs, 1961; Levy, 1999; Moughtin et al. 2003). In recent years, urban design initiatives based on smart growth and new urbanist principles are focused on revitalizing central city and inner ring suburbs to counteract the outward march of people and businesses to the suburbs (Burchell,

Listokin, & Galley, 2000; Addison, Zhang, & Coomes, 2013). Urban design principles to achieve smart growth include: mixed used, walkable neighborhoods, a variety of housing types

(multi and single family), and a diverse choice set of transportation options (Ye, Mandpe, &

Meyer, 2005; Addison, Zhang, & Coomes, 2013). A related but distinct perspective on urban design is the new urbanism. In the charter of the Congress for the New Urbanism (2015), several elements of this design strategy are listed, including: distinctly defined walkable neighborhoods, a connected street network that is lined with buildings, a mix of activities and housing choices, placement of civic places in important areas, amongst others. Given the popularity of these design movements, researchers have attempted to operationalize these ideas for empirical study

(Vale, Pereira, & Saraiva, 2016). In a widely-cited paper, Cervero and Kockelman (1997) introduced the “Three D’s” – density, land use diversity, and street network design – to which

88 the “D’s” distance to transit and destination accessibility were later added (Ewing & Cervero,

2010). Similarly, Krizek’s (2003b) individual neighborhood accessibility indicators provide a useful encapsulation of the generally-accepted principles of walkable urban design that includes high density, small lots, mixed land use, and access to parks, to name just a few.

Benefits of Walkable Urban Design

Evaluations of good urban design have uncovered a range of social, environmental, and health benefits. Since the 1972 Appleyard and Lintell study of livable streets, which called attention to the role of design in improving neighborhood interactions (Lund, 2003), planners have noted the link between social benefits and good urban design (Montgomery, 1998; Knox,

2005). In fact, a fundamental tenet of the new urbanism is to restore lost neighborhood interactions created by suburbanization and the advent of gated communities by creating a sense of community through the strategic placement of public spaces (Talen, 1999). Aside from facilitating neighborhood scale interactions, good design practice also creates third places outside of home and work such as coffee shops, restaurants, and parks which facilitate casual encounters

(Knox, 2005).

Given the health hazards of car-oriented, sedentary, suburban lifestyles (Saelens et al.

2003), the last decade has also witnessed a surge in interest in evaluating the impact of urban design on physical activity (Frank & Engelke, 2001, Saelens et al. 2003). The overarching idea is that good urban design that enhances walkability, will reduce automobile dependence and car- related travel. While early work found little evidence to link design to travel behavior (Crane &

Crepeau, 1998; Crane, 2000), a number of studies have uncovered an association between various facets of urban design and travel behavior. Cervero and Kockelman (1997) found modest impacts of built environment characteristics (density, land-use diversity, and pedestrian-oriented

89 design) on travel demand, which lends support for new urbanist design principles. Krizek

(2003a) found that locating in neighborhoods with higher levels of accessibility decreased car travel, as measured by vehicle miles traveled. Frank et al. (2006) found that people living in walkable neighborhoods with better urban design were more active, less car dependent, and less polluting than residents of less-walkable neighborhoods.

Economic Value of Walkable Urban Design

Recent work has also begun to examine the economic benefits of good urban design with a focus on property values. Song and Knaap (2003) found for example that people are willing to pay more for a range of new urbanist neighborhood characteristics such as mixed use, smaller blocks, more connected streets, and proximity to light rail stations. A follow-up study found that a mix of land uses had a positive impact on property values, but that this relationship depended on the mix of land uses considered because multi-family land uses did not positively impact property values (Song & Knaap, 2004). These findings are tempered however by work which finds that walkable amenities do not increase property values in auto-centric neighborhoods (Li et al. 2015). Findings from property value work are also tempered by studies highlighting equity issues with design-specific features of environments that promote walkability (Davis, 1984).

This calls into question the widespread affordability of walkable neighborhoods (USDOT, 2008;

Pollack, Bluestone, & Billingham, 2010) and the economic accessibility of neighborhoods that personify features of good urban design.

Conceptual Framework

While equity issues and gentrification are negative externalities of good urban design, the majority of studies highlight a wide range of social, environmental, and health benefits. Work on the link between design and vibrant cities highlights the importance of a mix of business sizes 90 and types to city vitality (Montgomery, 1998). Figure 10 presents a conceptual framework for thinking about place-making and urban vitality that combines different perspectives on place

(psychological, activity-based, and design-based) to highlight how these components interact to create a unique sense of place for urban environments (Montgomery, 1998 p. 98). This figure underscores the fact that there is a reciprocal relationship between local economic activity and place-making. A mix of successful businesses drives urban activity and street life, which constitutes a critical component for creating unique urban places (Montgomery, 1998). And, at the same time, the form and image of urban areas influences business success.

Figure 10. Conceptual framework for components of urban place-making (Montgomery, 1998).

Table 88 provides a more detailed description of each of the three elements of place as outlined by Montgomery (1998). Successful business activity underscores many of these activities: restaurants and coffee shops form the foundation of a café culture and make up the transactions component of fine-grained neighborhood economies. Elements of urban form – such as scale, block length, and diverse building stock – create the underlying conditions for economic activity, and thus design dictates a lot about how that activity plays out. In vital urban places, with a mix of businesses competing, innovating, and vying for customers, we would expect business performance to increase - certainly for those businesses that rely on foot traffic and the public or semi-public realm, such as pubs, cafes, restaurants, and retail shops. The impact of image, legibility, sensory experience, and symbolism on business performance also cannot be ignored. Places with strongly shared memories for a large number of people and easy psychological accessibility should also perform better than nondescript, hard-to-remember areas.

This relationship helps to explain the use of nostalgia and place-experience in businesses marketing and advertising, even in "controlled" environments like shopping malls or theme parks

(Harvey, 1989; Relph, 1976; Venturi, Brown, & Izenour, 1972).

The interaction of people and businesses, which is enabled by a well-constituted urban design, is a core component of place-making. In addition, a unique place-based identity – nurtured physically by specific urban design components – contributes to individual business success and urban vitality. While "place" is mutually-constituted through the links between form, activity, and image, this paper chooses to focus on a particular relationship - the connection between features of urban form that enhance walkability and business performance. At the most practical level, this is an important link to study, because planners and cities have a relatively high amount of control over the built environment, and profits are the key economic need for a

92 business to survive and be successful (and thus being able to continue to contribute to the realm of urban activity).

Table 8. Activity-, form-, and image-based components of place-making.

Components of Place-Making

Street life, diversity, vitality, people-watching, café culture, events and local traditions, transaction base, fine-grain economy

Scale, intensity, permeability, landmarks, diverse building stock, public spaces, space to building ratios, block length

Symbolism and memory, imageability, legibility, sensory experience and associations, receptivity, psychological access, lack of fear

Source: Montgomery (1998).

Study Area

To analyze the linkages between specific elements of the built environment and business performance, as measured by sales volume per employee, this paper investigates this relationship in two cities with different historical backgrounds, business characteristics, and urban morphologies. Phoenix is a relatively younger, polycentric Sun Belt city that exemplifies the post-WWII suburban-style development patterns; between 1950 and 1990 Phoenix grew from 17 to 420 square miles (Fink, 1993). Issues prompted by sprawl make this metropolitan area a well- studied case of various maladies associated with unmitigated urban expansion (Heim, 2001;

Bernstein et al. 2014). After decades of struggle to revitalize a downtown area resembling more a

Western ghost town, the downtown core of Phoenix may be on the verge of revitalization (Pela,

2015).

Boston, in contrast, is one of the most historic places in the United States. Its dense historic core dates back to the 1600s and is recognized as a key player in the Revolutionary War. 93

The bustling downtown core of the city is known for its maze of twisted streets in the North end, as well as several renowned institutions of higher education including Boston College, Harvard, and the Massachusetts Institute of Technology (MIT). While Boston, like many major cities, bulldozed blighted areas of the city in urban renewal efforts that displaced thousands of low income families, the city has ongoing urban renewal efforts with increased emphasis on citizen participation and education (Mao, 2015). Recent efforts to strategically guide Boston’s growth via Imagine Boston 2030, are a response to the rapid population growth of the metropolitan area in recent years (6% between 2010 and 2014) (Imagine Boston, 2016).

Data

Given the differences between Boston and Phoenix, it is hypothesized that the built environment is positively related to business performance, but that the strength and direction of this association varies across cities due to differences in regional form, behavioral patterns, and economic structure. In this study, the built environment is operationalized with six variables: a density activity score, block length, transit accessibility, pedestrian and bike accessibility, mixed land use, and a diversity of building ages (Jacobs, 1961; Ewing & Cervero, 2010). To test this hypothesis, secondary data were compiled from a variety of sources. These data are summarized in Table 99 and explained in further detail below.

Business Data

Point-level data about business location and business performance and productivity, as measured by sales volume per employee, were obtained from two sources, the National

Establishment Time Series (NETS) database and the ESRI/Reference USA database.

Table 9. List of variables considered for use in models.

Variable Level Category Name Description Source Dependent variable: natural log of NETS 2010 & LOGSALES sales volume per employee ESRI/Reference USA Dummy for retail business NETS 2010 & RETAIL Business (NAICS 44-45) ESRI/Reference USA Level-1 characteristics Dummy for manufacturing NETS 2010 & MAN business (NAICS 31-33) ESRI/Reference USA Dummy for knowledge business NETS 2010 & KNOW (NAICS 51-52 and 54-55) ESRI/Reference USA MEDAGE Population median age ACS 2008-2012 WHTNH % white non-Hispanic population ACS 2008-2012 BLKNH % black non-Hispanic population ACS 2008-2012 Demographic variables ASNNH % Asian non-Hispanic population ACS 2008-2012 HISP % Hispanic population ACS 2008-2012 % population with Bachelor's BACH ACS 2008-2012 degree or higher education TRANS % commuting to work via transit ACS 2008-2012 DENSITY Population + employees per acre ACS 2008-2012 Average Census block perimeter AVG_SHAP ACS 2008-2012 length (in meters) of the tract Level-2 Features of Herfindahl Index (HI) for walkable built B_AGE_HI building age by decade from ACS 2008-2012 environments¹ 1939-2012 Parcel-level data from MAG MIX Dummy for mixed use² and City of Boston % commuting to work via PEDB ACS 2008-2012 walking or biking TRANSLAG Spatial lag for TRANS variable ACS 2008-2012 POPEDLAG Spatial lag for DENSITY variable ACS 2008-2012 Spatial lag³ MEDALAG Spatial lag for MEDAGE variable ACS 2008-2012 WHITELAG Spatial lag for WHTNH variable ACS 2008-2012 BACHLAG Spatial lag for BACH variable ACS 2008-2012

Note: NETS = National Establishment Time Series; ACS = American Community Survey; MAG = Maricopa Association of Governments. ¹Variables chosen to match Jacobs’ four generators of diversity (1961) and Ewing and Cervero’s “5-D’s” (2010). ²This variable was constructed by calculating the percentage of parcels classified as “commercial”, “residential,” and “public” in each tract; tracts with ≥ 5% of both commercial and residential land use, in addition to any % of public land use, were classified as “mixed use”. ³Spatial lag variables use a first-order queen contiguity spatial weights matrix.

NETS is built in collaboration with Dun and Bradstreet to collect a longitudinal database of business activity that may be tracked over time (Neumark, Zhang, & Wall, 2005). Sales information in this database are taken from reported sales at the firm level (Walls & Associates,

2013). Since individual establishments (which make up firms) also report sales, this is how information is obtained for the majority of establishments (Walls & Associates, 2013). In instances where sales are unavailable, estimates of sales per employees and employment information at the establishment level, are used to estimate establishment sales (Walls &

Associates, 2013).

The 2010 ESRI Business Analyst data used in this paper also comes from the Dun and

Bradstreet database (ESRI Business Analyst, 2014). The primary difference between ESRI

Business Analyst and NETS is the extraction and geocoding process; for the former, it is conducted by ESRI, and for the latter, by Don Walls & Associates. The core source – Dun and

Bradstreet – for the business information is the same. Only businesses with positive sales volume and at least two employees in 2010 were included in the final dataset. In order to test for industry-specific effects in the relationship between BE features and business performance, dummy variables were constructed for three types of businesses: retail (NAICS 44-45), manufacturing (NAICS 31-33), and knowledge (NAICS 51-52 and 54-55).

Tract-Level Data

In addition to business-level data on the location, sales volume per employee, and industry type of individual businesses, this paper also employs a unique Census tract database for

Boston and Phoenix. Built environment variables were chosen to closely match Jacobs’ four generators of diversity (1961), as well as Ewing and Cervero’s “5-D’s” (2010). Parcel-level data from the Maricopa Association of Governments (MAG) in Phoenix and the City of Boston were

96 used to create a dummy variable for mixed use tracts. This variable was constructed by calculating the percentage of parcels classified as “commercial”, “residential,” and “public” in each tract; those tracts with at least 5% of both commercial and residential land use, in addition to any percentage of public land use, were classified as “mixed use”18. This measure operationalizes the concepts of destination accessibility (Ewing & Cervero, 2010) and the need to have more than one primary use in a district (Jacobs, 1961); 5% represents a minimum percentage of land use in a tract that might realistically contribute to its usage pattern. In order to measure the density of activity in a tract, data about the residential population from the 2010

Decennial Census and employment from the ESRI Business Analyst and NETS business point data were added together and divided by the size of the tract (in acres). This variable provides a total measure of aggregate density in an area that combines the economic benefits of two different kinds of activity modes: daytime (employment) and nighttime/weekend (residents).This is important to capture since activity from residents and workers at different times of day – even if they happen to be the same individual human being – is essential to fostering vibrant places

(Jacobs, 1961). The intention of using this kind of density activity score, rather than employment or population density alone, is to create a measure that captures the economic benefits of tracts in which a large amount of people both work and live. These are the dense kinds of areas that provide 24-hour street life and value to businesses of all types.

Block length is another important characteristic of street network design – this measure was obtained by calculating the perimeter of each tract’s nested 2010 Census blocks (which

18 Based on the land use categories obtained from the parcel data, “commercial” land uses were those coded C1 (small-scale retail), C2 (restaurants, coffee shops, bakeries, etc.), C3 (office buildings and banks), MU2 (vertical mixed use without residential), and S1 (commercial services, e.g., dry cleaning). “Residential” land uses included in the calculation of the mixed use dummy were MU1 (vertical mixed use with residential), R2 (single-family attached housing), and R3 (condominiums and multi-family housing). “Public” uses were PO1 (plaza, parks, playgrounds, etc.) and S3 (public recreational buildings, libraries, etc.). In Phoenix, 12 of 357 tracts were classified as mixed use, while in Boston, 18 of 176 met the definition. 97 generally correspond to a city block) and averaging those values across tracts. The Census also provides data on building age by tract – the share of total buildings by decade (from pre-1939 through post-2010). These data were used to create a Herfindahl Index of building age that captures the diversity of business ages and types within neighborhoods (Jacobs, 1961). Finally, transit and pedestrian/bike accessibility, which are important dimensions of walkable design

(Ewing & Cervero, 2010), were estimated by calculating the share of transit commuters and pedestrian/bike commuters (respectively) by tract from the 2008-2012 American Community

Survey (ACS).

In addition to these BE variables, demographic and spatial control variables were also included in the study. Demographic data were collected from the 2008-2012 ACS about median age, the share of population with a Bachelor’s degree or higher, and the ethnic/racial profile of the population. The race/ethnicity variables include: the percentage of white non-Hispanic, black non-Hispanic, Asian non-Hispanic, and Hispanic population per tract. A series of spatially lagged variables were also created in order to control for spatial autocorrelation.

Given the need to include controls for spatial effects in this modeling framework, this study models spatial effects through the independent variables. This enables the use of a first- order queen weights matrix and also exploits little-used information about the way spatial models are estimated. When spatial lag models are estimated, lags of each of the independent variables are produced (Anselin, 1988). This is because regression models produce estimates of the dependent variable as a function of the independent variables. Thus, by lagging key independent variables that are responsible for spatial effects in the dependent variable, it is possible to indirectly account for the bulk of spatial dependence in the dependent variable.

In order to determine which independent variables had strong relationships with the spatial distribution of the dependent variable (and thus were good candidates to lag), the local

Moran’s I was calculated for the average sales volume per employee (aggregated dependent variable at Level-2), as well as each independent variable, in both study areas. The independent variables lagged were those with the highest correlation of local Moran’s I ‘hot spots’ to those of the average sales volume per employee (by tract). In Phoenix, lags were computed for median age, white non-Hispanic, and Bachelor’s degree attainment variables, while in Boston, transit accessibility and density activity score were lagged.

Methodology

This paper uses two-level hierarchical linear modeling to examine the relationship between individual- and neighborhood-level traits and the sales volume of individual businesses

(Raudenbush & Bryk, 2002). In this case, the Level-1 units (individual businesses) are nested within Level-2 units (neighborhoods, operationalized as Census tracts). Conceptually, HLM is similar to estimating a pooled ordinary least squares (OLS) model for all of the individuals within Level-1, where the dependent variable is a characteristic of the individuals (in this case, it is individual establishment sales volume per employee in 2010). In the “random coefficients” model, the intercept and each of the slope coefficients for the Level-1 equation become the dependent variables for a new set of regression equations, with Level-2 independent variables

(e.g., average block length, mixed use dummy, density activity score, etc.) and coefficients included in each (Woltman et al. 2012). The “intercepts- and slopes-as outcomes” or “random slopes” model expands on the random coefficient model by including Level-2 variables to predict the slope of each Level-1 predictor (Raudenbush & Bryk, 2002). In the context of understanding the relationship between the BE characteristics of tracts and business

99 performance, this modeling approach provides detailed information about the association between tract characteristics and individual business determinants of sales performance, including (importantly) industry type. In order to estimate an effective random slopes model to answer the research question of interest, it is necessary to estimate preceding models which provide important information about the variation in sales performance and proposed individual and tract determinants of performance.

Null Model

The first step in HLM model-building is to estimate a “null model” to which additional variables can be added (Hox, 2002). The Level-1 and Level-2 equations for the HLM null model used in this paper are given by (Raudenbush & Bryk, 2002):

퐿푂퐺푆퐴퐿퐸푆푖푗 = 훽0푗 + 푟푖푗 (7)

훽0푗 = 훾00 + 푢0푗 (8)

푡ℎ where 퐿푂퐺푆퐴퐿퐸푆푖푗 is the natural logarithm of the sales volume per employee for the 푖

푡ℎ establishment in the 푗 tract,푟푖푗 is the random Level-1 residual, 훽0푗 is the random intercept for tract j, 훾00 is the grand mean’s Level-2 intercept (which is estimated as a weighted average of tract means), and 푢0푗 is the random Level-2 residual or the dispersion around the grand/overall mean. Taken together, the final equation for the null model is:

퐿푂퐺푆퐴퐿퐸푆푖푗 = 훾00 + 푢0푗 + 푟푖푗 (9)

This model is important because it provides information about the nature of the variation in sales volume that occurs between tracts (휏00) as a proportion of total variability – both

100 between and within tracts (휎2) . This information may be summarized with the intraclass correlation coefficient (휌) in equation (10) (Raudenbush & Bryk, 2002; Woltman et al. 2012):

휏00 . 휌 = 2 (10) 휏00+휎

Larger values of this coefficient highlight more variation in sales volume driven by tract characteristics rather than individual business characteristics. This means that the intraclass correlation coefficient demonstrates the relative importance of neighborhood factors vs. characteristics of the individual businesses themselves in predicting sales volume. While we might expect that specific features of businesses are the most important factor in explaining sales performance (including some unobservable characteristics), the degree to which neighborhood grouping matters provides insight into the role of neighborhood context in the distribution of sales volume per employee, and addresses the first research question of interest. Table 100 presents the results of the intraclass correlation coefficients for Phoenix and Boston and highlights that neighborhood characteristics account for a higher proportion of variation in sales volume per employee in Phoenix (5.3%) than in Boston (3%).

Table 10. Intraclass correlation coefficients for Phoenix and Boston.

Phoenix Boston Random Effect Std. Dev. Variance Std. Dev. Variance INTRCPT1, u0 0.179 0.032 0.105 0.011 level-1, r 0.758 0.575 0.601 0.362 Intraclass Correlation 5.3% 3.0%

101

Random Coefficients Model

Building on these findings, the next step in the analysis is to build a random coefficients model with relevant covariates for each metropolitan area. The specification of the mixed random coefficients model for Phoenix is:

퐿푂퐺푆퐴퐿퐸푆푖푗 = 훾00 + 훾01퐵퐿퐾푁퐻푗 + 훾02퐴푆푁푁퐻푗 + 훾03푇푅퐴푁푆푗 + 훾04퐷퐸푁푆퐼푇푌푗 + 훾05퐴푉퐺_푆퐻퐴푃푗 + 훾06퐵_퐴퐺퐸_퐻퐼푗 + 훾07푃퐸퐷퐵푗 + 훾10푅퐸푇퐴퐼퐿푖푗 + + 훾20푀퐴푁푖푗 + 훾30퐾푁푂푊푖푗 + 푢0푗 + 푢1푗푅퐸푇퐴퐼퐿푖푗 + 푢2푗푀퐴푁푖푗 + 푢3푗퐾푁푂푊푖푗 + 푟푖푗 (11)

The random coefficients specification for Boston is:

퐿푂퐺푆퐴퐿퐸푆푖푗 = 훾00 + 훾01푇푅퐴푁푆푗 + 훾02퐷퐸푁푆퐼푇푌푗 + 훾03퐵_퐴퐺퐸_퐻퐼푗 + 훾04푀퐼푋푗 + 훾05푇푅퐴푁푆퐿퐴퐺푗 + 훾10푅퐸푇퐴퐼퐿푖푗 + + 훾20푀퐴푁푖푗 + 훾30퐾푁푂푊푖푗 + 푢0푗 + 푢1푗푅퐸푇퐴퐼퐿푖푗 + 푢2푗푀퐴푁푖푗 + 푢3푗퐾푁푂푊푖푗 + 푟푖푗 (12)

In the construction of these models, it is critical to assess statistical issues such as confounding variables and collinearity (Hox, 2002; Clark, 2013; Yu, Jiang, & Land, 2015). In order to assess the impact of collinearity, variance inflation factors (VIF) were calculated for each of the covariates of interest (separately) in both Phoenix and Boston. The results of these calculations are shown in Appendix C.1. VIF > 5 are generally considered to be problematic

(Clark, 2013; Yu, Jiang, & Land, 2015); so variables in each region with VIF above 5 were removed (shown in bold in Appendix C.1). Additional confounding variables – discovered when running the random coefficients models in cases where coefficients displayed the opposite sign of the underlying variable’s correlation with the dependent variable – are also identified in

Appendix C.1 and were removed from the final model specification. In both cities, many of the demographic variables are correlated, which explains why these characteristics could not be included in the final models, and also provides insight into the relatively-segregated nature of

102 neighborhoods in both cities. In Boston, for example, the black non-Hispanic population is highly negatively correlated with both white non-Hispanic population (-.83) and Bachelor’s degree attainment (-.66), while the Hispanic population is positively correlated with transit commuting (.55) and negatively correlated with Bachelor’s degree attainment (-.57). In Phoenix, the white non-Hispanic population is negatively correlated with Hispanic population (-.90) and positively correlated with median age (.80) and Bachelor’s degree attainment (.76).

In equations (11) and (12), all variables are grand mean centered. Due to the importance of the intercept in HLM models, centering is often recommended, even for Level-1 dummy variables (Raudenbush & Bryk, 2002, p. 34). If a predictor – for example, TRANS – remains un- centered, the intercept found by the equation is the expected sales volume per employee for a business in tract j with 0% transit commuting percentage. It is more useful (for the purposes of this paper) to set the intercept equal to the expected sales volume per employee for a business in tract j whose transit commuting percentage is equal to the average transit commuting percentage in the study area (grand mean). Thus, all of the independent and dependent variables in this paper are grand mean-centered.

Random Slopes Model

While the random coefficients model shows which BE factors positively relate to business performance, while controlling for industry effects, estimating a random slopes model is necessary in order to find which BE factors positively correspond to the performance of specific types of businesses. This model is an extension of the random coefficients model, with Level-2 predictors added to explain the slope coefficients of each of the Level-1 predictors, creating several cross- level interaction terms. For Phoenix, the mixed random slopes model specification is:

103

퐿푂퐺푆퐴퐿퐸푆푖푗 = 훾00 + 훾01퐵퐿퐾푁퐻푗 + 훾02퐴푆푁푁퐻푗 + 훾03푇푅퐴푁푆푗 + 훾04퐷퐸푁푆퐼푇푌푗 + 훾05퐴푉퐺_푆퐻퐴푃푗 + 훾06퐵_퐴퐺퐸_퐻퐼푗 + 훾07푃퐸퐷퐵푗 + 훾10푅퐸푇퐴퐼퐿푖푗 + 훾11푇푅퐴푁푆 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾12퐷퐸푁푆퐼푇푌 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾13퐴푉퐺_푆퐻퐴푃 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾14퐵_퐴퐺퐸_퐻퐼 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾15푃퐸퐷퐵 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾20퐾푁푂푊푖푗 + 훾21푇푅퐴푁푆 ∗ 퐾푁푂푊푖푗 + 훾22퐷퐸푁푆퐼푇푌 ∗ 퐾푁푂푊푖푗 + 훾23퐴푉퐺_푆퐻퐴푃퐸 ∗ 퐾푁푂푊푖푗 + 훾24퐵_퐴퐺퐸_퐻퐼 ∗ 퐾푁푂푊푖푗 + 훾25푃퐸퐷퐵 ∗ 퐾푁푂푊푖푗 + 훾30푀퐴푁푖푗 + 훾31푇푅퐴푁푆 ∗ 푀퐴푁푖푗 + 훾32퐷퐸푁푆퐼푇푌 ∗ 푀퐴푁푖푗 + 훾33퐴푉퐺_푆퐻퐴푃퐸 ∗ 푀퐴푁푖푗 + 훾34퐵_퐴퐺퐸_퐻퐼 ∗ 푀퐴푁푖푗 + 훾35푃퐸퐷퐵 ∗ 푀퐴푁푖푗 + 푢0푗 + 푢1푗푅퐸푇퐴퐼퐿푖푗 + 푢2푗퐾푁푂푊푖푗 + 푢3푗푀퐴푁푖푗 + 푟푖푗 (13)

The random slopes specification for Boston19 is:

퐿푂퐺푆퐴퐿퐸푆푖푗 = 훾00 + 훾01푇푅퐴푁푆푗 + 훾02퐷퐸푁푆퐼푇푌푗 + 훾03퐵_퐴퐺퐸_퐻퐼푗 + 훾04푀퐼푋푗 + 훾05푇푅퐴푁푆퐿퐴퐺푗 + 훾10푅퐸푇퐴퐼퐿푖푗 + 훾11푇푅퐴푁푆 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾12퐷퐸푁푆퐼푇푌 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾13퐵_퐴퐺퐸_퐻퐼 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾14푀퐼푋 ∗ 푅퐸푇퐴퐼퐿푖푗 + 훾20퐾푁푂푊푖푗 + 훾21푇푅퐴푁푆 ∗ 퐾푁푂푊푖푗 + 훾22퐷퐸푁푆퐼푇푌 ∗ 퐾푁푂푊푖푗 + 훾23퐵_퐴퐺퐸_퐻퐼 ∗ 퐾푁푂푊푖푗 + 훾24푀퐼푋 ∗ 퐾푁푂푊푖푗 + 푢0푗 + 푢1푗푅퐸푇퐴퐼퐿푖푗 + 푢2푗퐾푁푂푊푖푗 + 푟푖푗

(14)

Results

Prior to describing these model results, it is necessary to highlight some important differences between the two metropolitan areas which are critical to understanding the results.

Appendix C.1 displays the descriptive statistics for each of the variables considered in the HLM models described above. In terms of industry breakdown, 17% of the businesses in Phoenix are classified as retail; 6% are manufacturing, and 19% are related to knowledge-based work. In

Boston, the breakdown is 9% retail, 2% manufacturing, and 27% knowledge. Nearly all of the characteristics associated with the chosen variables of walkable built environments are found in

19 As shown below, MAN was removed from this specification due to a lack of significant variability remaining from the random coefficients model. 104 higher average quantities in Boston than in Phoenix, including density activity score (55 residents and employees per acre vs. 11), mixed use tract percentage (10% to 3%), shorter average block length (1,986 meters to 3,773 meters), transit commuting percentage (32% to 4%), and pedestrian/bike commuting percentage (16% to 3%). Rates of building age diversity within tracts are actually higher in Phoenix, with a slightly lower Herfindahl Index value of 0.37 vs. Boston’s

0.41. Since low values of the Herfindahl index correspond to industrially diverse economies, these numbers mean that both Phoenix and Boston have relatively diverse industrial mixes. In terms of demographic indicators, Boston’s median age is 35.9, compared to Phoenix’s 33.2. As for educational attainment, 43% of people in Boston have a bachelor’s degree or higher, compared to

25% in Phoenix. The racial/ethnic mix of people is also distinct between the two metropolitan areas. Boston has comparatively more Black and Asian residents while Phoenix has more Hispanic residents (39% compared to 17%).

Boston also has higher average sales volume per employee ($68,186) than Phoenix

($159)20. Figure 11 shows the results of a local Moran’s I analysis of average sales volume between the two study areas. In Phoenix, the highest concentrations of sales volume are found in downtown

Phoenix and along a stretch south of downtown that includes Sky Harbor Airport and industrial areas along the Salt River. This area of central Phoenix also contains most of the ‘hot spots’ of local spatial autocorrelation of average sales volume per employee (those tracts classified as significantly “high high” and “high low” using a local Moran’s I analysis) (Anselin, 1995).

20 While this is a seemingly large gap, there are several possible explanations based on the significant differences in the economies of Boston and Phoenix. The overall patterns of urban development are quite different in Boston than in Phoenix – as a denser, more urban city with significantly higher property values, it is likely that businesses in Boston need to obtain higher sales volumes in order to offset high operating expenses (including land, labor, and capital). The spending power of residents in each city is also different – according to the 2008-2012 American Community Survey, average household income in Boston was $53,136, while in Phoenix it was $47,866. In addition, the fact that this data represents a cross-section of sales for 2010 could play a role in the difference – since this is directly after the Great Recession, it is possible that there are regional differences in the ways in which these industries were negatively impacted and/or able to recover. 105

Figure 11. Maps showing spatial ‘hot spots’ and distribution of average sales volume per employee by tract in Phoenix and Boston.

106

The high-high category includes tracts with higher than average sales volumes that are surrounded by tracts with similarly high sales volumes. The high-low category contains tracts that have higher than average sales volumes but are surrounded by tracts with lower than average sales volumes.

The I-17 corridor – home to several large research and technology parks (Metro Research

Center, Cave Creek Industrial Center, Karsten Industrial Complex, Eaton Industrial, and the

Black Canyon Commerce Park) and the Metrocenter Mall – also shows concentrations of higher average sales volumes per employee. Newly-developed areas in north Phoenix, such as Desert

View and Deer Valley, also contain higher concentrations of average sales volume per employee.

In Boston, Charlestown (north of the Charles River, near Cambridge) is a hot spot of high sales volume per employee. The area around Logan International Airport, the West End, South Boston

(including the neighborhood surrounding the Boston Innovation District), and the Jamaica Plain neighborhood also show higher average sales volumes per employee. Overall, average sales volumes per employee are much higher in Boston than in Phoenix.

Model Results

While the descriptive statistics and maps provide some insight into the spatial relationship between sales volume and neighborhood features, the model results provide detailed information about the strength of the statistical relationships between these variables. Table 11 shows the results of the random coefficients models specified in equations (11) and (12) with the random slopes models specified in equations (13) and (14). While this table does not report coefficient values, it does indicate important findings for the HLM model-building process and industry-specific effects. In Phoenix, all of the Level-1 variables have a significant p-value

(<0.001), showing that significant variance in the relationship between these variables and

107 individual business performance remains unexplained in the random coefficients model. For

Boston, however, the MAN variable displays a highly-insignificant value (>0.500), meaning that its variance has been sufficiently explained by the random coefficients model and thus should be removed from additional model specifications (i.e., there is nothing significant remaining to explain by adding covariates to better specify its slope) (Raudenbush & Bryk, 2002).

Table 11. Additional variance explained by random slopes model for Phoenix and Boston.

Random Coefficients Model Random Slopes Model

Add. Std. Std. Random Effect Variance p-value Variance p-value Variance Dev. Dev. Explained INTRCPT1, u0 0.1587 0.0252 <0.001 0.1554 0.0242 <0.001 4.1% RETAIL slope, u1 0.1844 0.0340 <0.001 0.1828 0.0334 <0.001 1.8% Phoenix MAN slope, u2 0.1982 0.0393 <0.001 0.1968 0.0387 <0.001 1.4% KNOW slope, u3 0.2449 0.0600 <0.001 0.2345 0.0550 <0.001 8.3% level-1, r 0.7109 0.5054 0.7109 0.5054 INTRCPT1, u0 0.0834 0.0070 <0.001 0.0832 0.0069 <0.001 0.4% RETAIL slope, u1 0.1471 0.0216 <0.001 0.1186 0.0141 0.001 35.0% Boston MAN slope, u2 0.0486 0.0024 >0.500 KNOW slope, u3 0.0946 0.0090 0.001 0.0834 0.0070 <0.001 22.2% level-1, r 0.5829 0.3398 0.5833 0.3402 Note: all variables grand-mean centered

Table 11 also indicates the proportion of additional variance explained by the random slopes model (and thus serves as justification for its use). In a similar way to the intraclass correlation coefficient, this proportion is calculated by subtracting the “conditional” variance explained by the random slopes model from the “unconditional” variance specified by the random coefficients model, and dividing that by the unconditional variance (Woltman et al.

2012; Raudenbush & Bryk, 2002):

108

휏̂푞푞(푢푛푐표푛푑푖푡푖표푛푎푙) − 휏̂푞푞(푐표푛푑푖푡푖표푛푎푙) 푃푟표푝표푟푡푖표푛 푣푎푟푖푎푡푖표푛 푒푥푝푙푎푖푛푒푑 푖푛 훽푞 = 휏̂푞푞(푢푛푐표푛푑푖푡푖표푛푎푙)

(15)

The resulting value – shown in the last column of Table 11 – indicates how much additional variance the random slopes models specified in equations (13) and (14) explain. And, since only Level-2 BE variables were added to explain the Level-1 industry characteristics in these models, this value also indicates how much additional variance in sales volume per employee these BE variables explain for each type of business. In Phoenix, BE variables explain only 1.8% and 1.4% of the performance of retail and manufacturing establishments

(respectively), but add 8.3% to the description of knowledge business performance. Thus, in

Phoenix, the results show that BE variables are more important to knowledge business performance than retail or manufacturing performance. In Boston, 35% of the performance of retail businesses and 22% of the performance of knowledge businesses is explained by the addition of BE predictors in the random slopes model, indicating that these variables play an important role in explaining the performance of these types of businesses.

To understand the relationship between BE characteristics and business performance in specific industries, Table 12 displays model results for the random slopes models for Phoenix and Boston. In Phoenix, model results indicate that businesses outside of the retail, manufacturing, and knowledge sectors that are located in tracts with higher percentages of black non-Hispanic population and transit commuting have better performance. Tracts with lower density activity scores, longer average block length, and a larger diversity of building ages are also positively associated with better business performance. For retail businesses in particular, the built environment has no relationship with performance.

109

Table 12. Final model results for Phoenix and Boston.

Phoenix Boston Fixed Effect Coefficient SE Sig. Coefficient SE Sig. For INTRCPT1, β0 INTRCPT2, γ00 5.021 0.010 *** 11.086 0.008 *** BLKNH 0.323 0.093 *** TRANS 0.661 0.299 ** -0.290 0.106 *** DENSITY -0.006 0.002 *** AVG_SHAP 0.00002 0.00001 * B_AGE_HI -0.175 0.070 ** MIX 0.053 0.027 * For RETAIL slope, β1 INTRCPT2, γ10 0.565 0.016 *** 0.431 0.018 *** TRANS 0.354 0.142 ** DENSITY -0.0007 0.0002 *** B_AGE_HI -0.200 0.091 ** MIX -0.084 0.044 * For KNOW slope, β2 INTRCPT2, γ20 0.503 0.018 *** 0.266 0.013 *** TRANS, -1.391 0.481 *** DENSITY 0.007 0.003 ** B_AGE_HI 0.323 0.132 ** For MAN slope, β3 INTRCPT2, γ30 0.566 0.019 *** TRANS -0.730 0.431 * DENSITY 0.008 0.004 * B_AGE_HI 0.232 0.120 *

p-values: *** ≤ .01, ** ≤ .05, * ≤ .1 Note: all variables grand-mean centered; only significant results shown

For the manufacturing and knowledge sectors, various features of the built environment are related with business performance. These features include: higher density activity score, lower transit commuting percentage, and less building age diversity.

110

In Boston, Table 12 highlights that higher performance for manufacturing and knowledge businesses is significantly related to lower transit commuting percentage and location in a mixed use neighborhood. On the other hand, higher sales volume per employee for retail businesses corresponds to higher transit commuting percentage, lower density activity score, a diversity of building ages, and location in neighborhoods largely dominated by a single use. None of the individual BE variables are significant predictors of knowledge business performance.

Discussion and Conclusion

The goal of this study is to analyze the linkages between good urban design and individual business performance, as measured by sales volume per employee. Results of the hierarchical linear models estimated reveal that the relationship between the performance of individual businesses and the built environment features of the neighborhoods in which they are located is complex, nuanced, and highly-dependent on the type of business and city in question.

In Phoenix, for example, BE characteristics were important to understanding the performance of knowledge but not retail or manufacturing businesses; in Boston, BE characteristics were important to understanding the performance of both knowledge and retail businesses. While there are certainly interesting details to be gleaned from the results, the overarching finding is that there is no ‘one-size-fits-all’ approach to place-based economic development.

Neighborhood-level features play an important role in explaining the variation in business performance in both Phoenix and Boston; while it is clear that business performance and productivity is largely a product of features endemic to individual businesses (such as management, financial status, technology, market demand for the product, location in a specialized business cluster, etc.), a consequential portion is related to characteristics of the local neighborhood, such as demographics and the built environment.

111

Table 13 summarizes the results for the two metropolitan areas and highlights that several walkable BE components are significantly related to higher business performance. In Phoenix, businesses in tracts with higher levels of transit commuting and building age diversity – as measured by a Herfindahl Index of the shares of buildings constructed in different decades (from

1939 – present) – have higher sales volumes per employee, while in Boston, the same is true for businesses located in mixed use tracts, even when controlling for socio-demographic features of the neighborhood and the characteristics of specific business types. This provides some evidence supporting Jacobs’ assertion that visual intricacy and a variety of flexible building space helps foster economic activity (1961).

At the same time, this analysis shows that some elements of walkable built environments are negatively related to business performance. In Phoenix, lower densities and longer average block length are connected with higher performance, which suggests that in some cases auto- centric built environments lead to better business outcomes; this is particularly true in a city like

Phoenix, where auto-centric urban form – and economic behavior – is prevalent. Businesses that require a lot of parking to support their business model, such as big-box retail stores, do not substantially benefit from walkable urban design, which could be driving an insignificant result for the retail variable in Phoenix. This underscores the fact that, while measured in the same way, the walkable built environment variables tested here mean different things in different urban contexts, e.g., transit use is a different economic indicator in Boston than it is in Phoenix.

112

Table 13. Summary of model results.

Observation Method Model Result

Phoenix Boston 1. Neighborhood- Intraclass Confirmed; ρ = 5.3% Confirmed; ρ = 3.0% level correlation characteristics coefficient predict business (ρ) performance 2. Walkable BE Random Confirmed for TRANS (+) Denied for DENSITY (-) & Confirmed for MIX (+) Denied for TRANS (-) features relate to slopes & B_AGE_HI (-) AVG_SHAP (+) higher business model performance, controlling for industry effects 3. Walkable BE Random Industry Confirmed for Denied for Industry Confirmed for Denied for features relate to slopes RET – – RET TRANS (+) & B_AGE_HI (-) DENSITY (-) & MIX (-) higher business model KNOW DENSITY (+) TRANS (-) & B_AGE_HI (-) KNOW – – performance for MAN DENSITY (+) TRANS (-) & B_AGE_HI (-) MAN – – specific industries

113

Certainly, this paper represents the beginning of an analysis of the neighborhood-level micro-foundations of business performance, and future work is needed to illuminate specific relationships in a wider range of contexts. Larger nested models with a variety of regional types could shed light on the ways in which metropolitan-level features influence place-making and business performance. For that to be possible, however, a large sample of individual businesses – perhaps drawn from several years – would be necessary in order to ensure a sufficient amount of within-tract variation. Another interesting extension of this paper would be to provide a more precise breakdown of the interaction effects of various BE features – one of the limitations of this work is that, especially in Boston, several of the design characteristics are too collinear to use together in a regression. Principle Components Analysis (PCA) could perhaps be used to better understand the relationship that these variables have to one another. Similar approaches could also be used to test Jacobs’ assertion that the features of urban design function properly only when they are all concurrently present – that the whole is greater than the sum of its parts, so to speak (1961).

Despite these limitations, the findings indicate that physical design interventions such as historic preservation (to maintain a diverse building stock) and the development of fine-grained mixed use places have the potential to increase the performance of individual businesses. These results suggest economic benefits to urban design above and beyond the social, health, and environmental benefits of walkable urban environments noted in previous studies. However, policy interventions must be context-dependent and sensitive to a locality’s economic structure, aggregate urban form, and behavioral patterns. The results of this paper suggest that walkable built environments have different – sometimes even negative – relationships with business performance in different urban contexts. Planning efforts to design economically vibrant places

114 with aesthetic appeal and a sense of place can be used by economic developers to market profitable place characteristics to prospective businesses, but practitioners must be careful to understand that effects for different types of businesses, customers, and markets will vary.

Economic development strategies based around place-making initiatives should be targeted to the specific businesses that will benefit most from specific built environment features; this study represents the first step in understanding these detailed relationships.

115

CONCLUSION

Limitations

The research presented in this dissertation seeks to quantitatively illuminate the connections between specific features of the urban environment (including public transit systems) and positive economic development outcomes. In terms of the built environment, this approach is novel because previous work on the economic impact of physical urban form is often theoretical or qualitative in nature (Jacobs, 1961). However, by distilling these complex urban relationships into statistical frameworks that inherently simplify the real world, this approach is naturally imperfect in some important ways.

First, all three papers’ variables of interest are features of or distances between physical objects, e.g., block length or distance to light rail stations, which – without the correct framing – could be construed as a kind of crude physical determinism, negating the importance of people

(and their culture, socioeconomic status, age, etc.) as the key actors of significance. Second, the specifications of the statistical models used here lack a complete evaluation of many factors that are relevant to the relationships of interest, some of which – like path dependence and political power relations – that are difficult (or impossible) to quantify. In choosing a systematic quantitative approach to studying the general effect of these relationships across several study areas, the influence of local context is unfortunately marginalized. While this speaks to the importance of complimentary qualitative, case study, and theoretical work to more fully understand these processes, it is also important to understand this fact when interpreting the results of this project: the coefficients presented here are certainly products of model generalizations, data availability, and a multitude of decisions in the framing and execution of the analysis that are inherently subjective (Unwin, 1992). Similarly, the data used in this study –

116

NETS – has several limitations that need to be understood in order to correctly interpret the findings of this dissertation.

Physical Determinism

The tendency towards physical determinism as an implicit explanatory framework in studies of the built environment is also perhaps because the physical environment can be changed directly through policy or construction much more easily than the social environment.

This, combined with the greater political palatability of pursuing planning rather than social policy, especially in the United States, creates a tendency in applied studies to overstate the importance of physical design or built environment elements for influencing behavior (Riggs,

2014). Thus it is important to explicitly acknowledge both the need for additional supplementary research to put the influence of the built environment in its proper context, as well as the inherent

“design bias” in studies such as this (Riggs, 2014). While the approach used in this dissertation attempts to isolate the direct effect of various aspects of the built environment through careful econometric study design, it is still important to recognize that physical features are only one element that influences individual economic behavior, an influence that, as the results of Chapter

#3 indicate, is sometimes relatively small.

Beyond the inherent design bias, this project takes the perspective that the physical environment shapes pathways for human behavior but is not deterministic. This idea, known as the “natural movement hypothesis”, has been developed in previous studies of the built environment (Araldi & Fusco, 2016; Hillier, 1996). This hypothesis posits two pathways for understanding how physical form influences human behavior: on the one hand, the arrangement of visual elements in a pedestrian’s field of view attract further exploration and movement; on the other hand, the structure of the built environment limits or enhances physical access to these

117 visible elements (Araldi & Fusco, 2016; Hillier, 1996). While each individual may react differently to this combination of elements, previous work has identified a number of factors that, on average, provide important indications of the visual arrangement and physical access of urban areas, and thus average human reactions to it (Jacobs, 1961).The purpose of this work – particularly in Chapter #3 – is to assess average relationships between these aggregate built environment features and the economic consequences of increases or decreases in pedestrian accessibility.

Of course, many other factors (free will, culture, race, socioeconomic status, etc.) influence the actual way that any given individual will behave. Some of these factors, e.g., socioeconomic status, are far more important for understanding economic behavior than the pure structure of the physical environment. For example, Vojnovic et al. (2013) show that while the effect of compact built environments on pedestrian behavior is important, proximity to other amenities (such as grocery stores) and income opportunities are more significant; thus wealthy compact areas have higher rates of walking than poor compact areas. This is a very important consideration, and a limitation of the analysis in Chapter #3 in particular. By attempting to control for difference in socioeconomic status through the econometric study design – thus isolating the direct relationship between the built environment and business performance – this analysis necessarily ignores the other factors that might explain economic performance and travel behavior more fully.

Similarly, while the use of distance from transit stations as a proxy for station accessibility in Chapters 1 and 2 does not capture the important intervening characteristics of individuals (culture, race, socioeconomic status, and physical ability), it provides a convenient method for quantifying the concept of average accessibility. Again, the econometric study design

118 in these chapters attempts to control for features of residents (on average) so that the specific relationship between new business creation and station accessibility can be isolated. However, these simple measures of transit accessibility ignore other important features of the built environment that might influence new business creation, including the surrounding pedestrian environment. For instance, a poorly designed pedestrian environment around a transit station might logically have a much different impact on the economic benefits that accrue to new businesses that locate near it than a station with a very high-quality pedestrian context (Newman

& Kenworthy, 2013). While this is an interesting avenue for future research, quantifying the pedestrian environment around each station for implementation in econometric analysis is unfortunately beyond the scope of this project. Since these factors are not measured or proxied for in these chapters’ analyses, one can assume that the variation in new business creation due to transit-area pedestrian environments (and other unmeasured factors) accumulates in the error term of these models, perhaps inducing some (unavoidable) omitted variable bias in the model results.

Situation Within Cultural, Political, and Economic Contexts

By studying the relationship between features of the urban environment and economic behavior, this research is necessarily situated within a multitude of overlapping cultural, political, and economic contexts. The planning of light rail systems, walkable neighborhoods, and the location of new business investment in each of the cities studied in this project are highly path-dependent and tied to both local and global systems, including culture, behavioral norms, development trajectories, economic trends, politics, regulation, and investment. While these features are generally beyond the scope of the econometric analysis presented here, they fundamentally shape the conditions for this analysis and deserve to be addressed in some detail.

119

At the local level, unique behavioral cultures, development history, and land use regulations have a significant impact on the location of transit investments and walkable neighborhoods, the use and value of these features, and the type of possible business uses (and thus their sales performance) that can be located nearby. In Phoenix, as discussed in Chapter #1

(above), the planning of the light rail line was a contentious local political issue. Generally higher socioeconomic status municipalities – most notably Scottsdale – actively opposed connection to the system, while at the same time lower socioeconomic status areas along the final path of the rail line did not receive stations.

Figure 12. City of Phoenix Zoning Map from December 29, 2017 (Phoenix 2017a).

In general, the post-WWII development history of Phoenix is significantly auto-oriented; with this has come a culture that heavily favors driving – Phoenix is one of the largest markets for auto dealerships (Wiles, 2014) – and a land use pattern and regulatory scheme that favors separated, single use zoning. Figure 12 shows the Phoenix zoning designations for the central

120 portion of the city as of December 2017. It clearly displays the classic auto-oriented land use pattern: single-use commercial districts (red) strung along major high-speed arterial roadways with some interspersed (orange) multifamily development and large areas of single family (light yellow) residential that occupy the interstitial areas.

As the analysis in Chapter #3 also indicates, the walkable neighborhoods in the region are confined mostly to the comparatively small historic downtown business districts – the turquoise color in Figure 12 shows the areas in central Phoenix zoned for “downtown” uses. Given this regulatory scheme – and the economic behavior it encourages (or forces) – it is perhaps not surprising that Chapter #3 finds lower densities and longer average block length to be related with higher business performance in Phoenix.

On the other hand, local regulatory schemes that encourage new business development in specific areas – e.g., around transit stations – could also provide a positive ‘bias’ towards new business activity that influence the results of this project. Of the six regions in which the transit – new business relationship was evaluated for this project, only Phoenix (in the cities of Tempe and Phoenix) has adopted and begun to use TOD zoning ordinances in earnest, although TOD is mentioned in various planning documents for all six regions. Phoenix and Tempe’s TOD overlays seek to encourage mixed use development, pedestrian amenities, and compact development patterns by creating specific design development guidelines and favoring specific land uses (Phoenix, 2017b; Tempe, 2017). In particular, Tempe’s Station Area district requires ground floor retail, service, entertainment, hotel, clinical, or daycare uses in non-residential underlying districts on 60% of the street frontage (Tempe, 2017). The location of these districts and new business creation (by block) in three industries of interest over the time period that the light rail has been open is shown in Figure 13.

121

Figure 13. New Knowledge, Retail, and Services Business Creation in TOD Zoning Overlay Districts in the Phoenix region.

122

Close examination of the patterns suggests that the business creation in some areas may be related to the presence of a TOD overlay district – for instance, around the far northern station for retail businesses, and along the Central Avenue corridor for knowledge and service businesses. The Tempe districts are generally very small, but there does appear to be some overlap between areas of new business concentration and the TOD zoning overlay, especially in the retail and service sectors, which makes sense given the focus of the ordinance on encouraging retail and service-based mixed use. While further analysis of the interaction between land use regulations, economic behavior, and features of the built environment is beyond the scope of this project, this brief exploration certainly suggests that local zoning designations might independently contribute to the economic impacts generated by transit systems and/or walkable neighborhoods.

At the global scale, long-term economic trends (including economic restructuring and globalization), political norms such as neoliberalism, and conduits of federal government investment also drive the pattern of local economic outcomes in the cities analyzed here. For instance, the pattern of investment in high-technology industries that eventually resulted in significant clusters of new business activity in Silicon Valley and Boston are a direct result of investment in the military-industrial complex, weapons development during and after WWII, and the geo-political struggles of the Cold War (Saxenian, 1994). Thus the concentration of business activity in these industries shown in this project is necessarily related to these large-scale political decisions. Similarly, global economic forces underlie and reproduce similar patterns of economic disparity in cities, which are inherently reflected in the results of the analyses presented here (Mackinnon & Cumbers, 2011). While an in-depth treatment of the specific ways in which these global forces influence the relationship between physical urban planning

123 interventions and economic outcomes is significant and far-reaching enough to deserve its own separate analysis, it is important to acknowledge that the results of this study are inherently influenced by these strong currents.

NETS Data Limitations

While the National Establishment Time Series (NETS) data used in this dissertation offers almost unprecedented coverage of the location and characteristics of individual business establishments in the US, it also (like any large dataset) has significant limitations and quality concerns. NETS data is compiled and geocoded by a firm named Walls & Associates based on information from Dun and Bradstreet’s database on individual establishment characteristics, beginning in 1990 (Walls & Associates, 2013). Due to the massive coverage of the Dun and

Bradstreet data, NETS provides the “closest thing to a continual census of American business, government and non-profits available to researchers” (Walls & Associates, 2013). Also, in comparison to many federally-collected business datasets, NETS includes almost twice as many establishments with fewer than five employees, as well as contract and temporary employees, sole-proprietors, employees of private households, and those involved in railroad, agricultural production and government industries (Walls & Associates, 2012).

However, NETS has a number of unique characteristics and quality issues that need to be understood when interpreting the results of this research. First, it is vital to understand that the definition of an establishment in NETS is a “’unique line of business (SIC8) at a unique location’” (Walls & Associates, 2012), which contrasts with the traditional definition used by the

US Census of “a single physical location at which business is conducted or services or industrial operations are performed” (2016). While in practice these definitions often overlap, it does mean that in NETS it is possible that a firm with two unique lines of business (e.g., separate spun-off

124 departments) within the same physical location (e.g., one office building), which could inflate establishment counts when conducting aggregate analysis (such as in Chapters #1 and #2).

Second, the nature of the Dun and Bradstreet collection method – which involves database and internet searches, phone calls, marketing information analysis, and other techniques – suggests that the number of establishments captured in NETS may be increasing as time goes on, simply due to better collection methods and data availability. While Walls & Associates do not address this (likely) possibility directly in their informational materials, the aggregate data on the total number of establishments that they provide shows a 59% growth in the total number of US establishments from 2001-2012 (2012). It is currently unknown how much of this increase is due to actual increases in the number of new establishments versus increases in database coverage.

In addition, several issues have been anecdotally identified through the course of conducting this dissertation research. As discussed in Chapter #3, there may be some systematic issues with the individual establishment sales volume data; regional average sales volume per tract was significantly higher in Boston than in Phoenix. While many attempts were made to find a systematic data entry error (e.g., missing 0s or other direct order of magnitude differences) or other cause for the difference, none could be identified. Additionally, the author noticed instances of geocoding errors (e.g., a large number of establishments geocoded to the centroid of a zip code or other areal unit) and changing NAICS codes throughout the time series for individual establishments that may influence the spatial and temporal analyses conducted in this project. While (anecdotally) these errors occur for a relatively small number of establishments, they are certainly worth noting.

Even with these issues, the unprecedented size, coverage, and spatial granularity of the data still means that NETS presents a remarkable resource for studying business location

125 dynamics. Unfortunately, the only way to discover many of the systematic errors in large datasets such as these is to engage in research projects and then to evaluate the results and errors post-hoc. The purpose of this dissertation is to provide an exploratory quantitative look at business dynamics related to active transportation investments, and this exploration has produced a better understanding of the limitations of NETS data that can be incorporated into future research projects. However, the kind of full-scale analysis of data quality that NETS deserves is beyond the scope of this project. With access to the entire dataset in the future, it would be very useful to systematically identify the prevalence of these (and other) errors in the dataset.

Generalizability and Implications for Theory and Practice

Despite these limitations, this study presents an interesting set of results that have important implications for geographic theory and urban planning practice. Substantively, the results show that, generally, proximity to rail transit is related to higher levels of new business creation. Chapter #1 analyzes this relationship using a quasi-experimental study design that controls for overall trends in new business creation by using a long time series of data (1990-

2014) and attempts to isolate the specific effect of rail construction, finding positive relationships with knowledge, service, and retail businesses that decrease with distance from transit stations.

Chapter #2 takes a cross-sectional spatial econometric approach by analyzing the relationship between several rail modes and new business creation across five regions. While this method lacks the time series and experimental design that suggest causality, it makes up for it by assessing variations in the relationships across regions. This approach shows that distance to transit is more likely to be significantly related to new business creation in regions with more well-developed and mature transit networks.

126

The connection between features of walkable built environments and individual business sales volume is even more nuanced and context-dependent. Using a hierarchical linear model,

Chapter #3 shows that some walkable design features are positively related to business sales volumes; however, these connections vary between the regions studied, and some walkable features are negatively related to sales volume, particularly in Phoenix. While this result makes sense given the pattern of urban development and behavior in that city, it also suggests that the results here are driven significantly by regional context, urban form, and even the scale of the analysis. The study also shows that features of walkable built environments make up a relatively small (but significant) proportion of the variation in individual business sales volumes; this finding confirms traditional economic assumptions that characteristics of the firm itself are the most important factor for determining its performance, while also showing that physical features of the built environment do play a role in structuring access to businesses and thus a portion of their overall performance.

In total, these substantive findings confirm the idea that design, location, and access are important economic considerations for businesses, and that their impact can be modeled quantitatively in terms of both monetary sales and new business formation rates. While the economic value of location and access generally has been a central tenet of geographic location theory for some time, this work advances location theory in three specific ways. First, it introduces the notion that the physical layout of both transportation systems and the built environment plays a quantifiable role in business location and performance. Chapter #2 suggests that the size and maturity of the transit network, i.e., its ability to improve accessibility to multiple regional destinations, matters for new business creation, while Chapter #3 shows that specific urban design features are significantly related to individual business sales volumes.

127

While this is a rich field for future research, this dissertation provides a unique bridge between the (mostly qualitative) urban design literature and the (mostly quantitative) work of regional scientists and economic geographers on traditional principles of location theory.

Second, this work expands on the study of the economic value of transportation access by positing and testing unique theoretical connections between different transportation modes and types of businesses. Generally, previous studies of business location have measured transportation accessibility crudely by distance (e.g., Alonso, 1960) or have proxied for it with population density or some other aggregate measure of agglomeration. On the other hand, past work on the economic impacts of specific transportation modes have generally ignored business creation as an indicator variable of interest. The Introduction of this dissertation develops this concept more explicitly by theorizing specific connections between transportation modes and types of businesses based on economic benefits that are likely to accrue to these businesses based on the specific kind of ‘access’ created by a given mode. Chapters #1 and #2 begin to test these theories empirically by evaluating the connection between different rail modes and different types of businesses, and indeed find varying degrees of benefits.

Finally, this project advances research methods in economic geography by developing three generalizable quantitative models for business micro-data. The creation and comparison of these kinds of models is especially important as large new geographic datasets – like the

National Establishment Time Series used in this dissertation – become more commonplace with increasing computational power and the access to Geographic Information Systems (GIS) technology. Modeling the connection between neighborhood-scale planning interventions and business micro-data can be approached in two general ways: either by aggregating the individual points (and their characteristics or counts) to the neighborhood scale or by preserving the

128 individual variation of the businesses and incorporating neighborhood-scale features using a hierarchical linear model. Chapters 1 and 2 take the first approach by aggregating counts but use different methods to try to explain the underlying process that created the distribution of counts:

Chapter #1 uses trends over time within one region to assess the significance of the transit station

– new business count relationship, while Chapter #2 uses spatial trends across five regions in one time period. While the time series approach provides a stronger case for a causal relationship, the spatial econometric model comparisons made in Chapter #2 show that spatial econometric frameworks must be used to reduce Type I errors is the estimation of the significance of transit station proximity for new business creation. Chapter #3 takes the hierarchical approach by developing a two-level hierarchical model to model a characteristics of individual businesses

(performance) based on neighborhood-level characteristics (built environment features). By presenting these three approaches side-by-side, this dissertation demonstrates the applicability of these methods in different contexts, i.e., where time series or cross-sectional data are not available, or when the research question favors individual business characteristics over aggregates, and also (perhaps) points towards future methodological developments that combine the strengths of each individual approach, including spatial panel and quasi-regression models, more advanced models for spatial count data, spatial hierarchical models, and others.

Beyond these theoretical developments, this work also has interesting implications for planning practice. Primarily, the results show that investment in active transportation features can yield real economic benefits to new businesses in a variety of contexts. However, these benefits are very nuanced and depend heavily on both regional characteristics and the integration and design of the investments themselves. Transit systems can attract adjacent new business activity but are more likely to do so if the system provides high quality access to numerous

129

(important) destinations in the region. New rail systems provide benefits to new businesses even in historically auto-oriented regions – in fact, these areas may see a higher initial level of investment because of the lack of any existing transit development – but these benefits may tend to decrease over time unless the system continues to develop and provide truly competitive transportation service. At the same time, business sales volume and the built environment are importantly linked, but this link does not always favor walkable built environments. While density, building age diversity, transit accessibility, and mixed use relate to higher sales volumes in various non-retail industries, auto-oriented features often relate more closely with higher sales for retail businesses, because higher sales for retail businesses occur in big box stores where large amounts of goods can be purchased and carried home in cars. These relationships also vary significantly between cities with different large-scale patterns of urban form, development histories, and industrial concentrations, suggesting that planners must be wary of local context when proposing built environment interventions. Simply put, a dense, walkable neighborhood in the middle of Phoenix is not guaranteed to generate higher sales volumes for its retail businesses that neighborhoods with more auto-oriented characteristics, simply due to the nature of economic behavior in the region (at this point in time).

Avenues for Future Research

To provide a more comprehensive and rich view of the relationship between active transportation features and economic development, and to address the limitations mentioned above, several extensions to this work can be explored. First, future quantitative studies of the relationship between rail transit and new business creation should incorporate local zoning, station-area pedestrian environments, and quality of transit service. While these variables may

(in some cases) be time-consuming to collect, their addition could shed more light on the direct

130 effect of transit station proximity, as well as the interesting ways that zoning and the built environment interact with station area proximity to foster new business creation. Systematic analysis of these interaction effects could be particularly useful for planners attempting to understand the relative importance of local development regulations in enhancing economic development outcomes around transit stations.

Next, from a methodological standpoint, studies of the transit – new business creation relationship could benefit from additional testing and sensitivity analysis for the differences in results that come from using more advanced spatial count models, such as spatial GLMM and/or spatial filtering approaches. An interesting future study could involve direct comparison of the results from the spatial econometric smoothing approach used in Chapter #2 with these (and perhaps other) methods of modeling spatial count data. In addition, future quantitative work on new business creation and transit could benefit from the use of spatial panel models that incorporate both the time series and cross-sectional spatial variation that is available in NETS data, perhaps even employing a quasi-experimental framework (such as the AITS approach used in Chapter #1) to provide strong evidence for a causal relationship.

Building on the approach used in Chapter #3, future work is needed to look at the total impact of walkable built environments on business performance, as well as the interaction effects between different kinds of built environment features. While Chapter #3 seeks to identify the relationship between individual features of the built environment and business performance that ultimately may be less important than understanding the overall effect of walkable built environments on business performance. It would also be interesting to know whether Jacobs’

(1961) conceptualization of four generators of diversity – all present together – is true in all cases, or whether different combinations or thresholds of walkability influence business

131 performance (even non-linearly). In addition, future work on built environment and business performance may want to look in a more detailed way at business types – there is significant heterogeneity even within a fairly discrete category such as “retail”, so a breakdown between, perhaps, “big box” and “neighborhood” scale retail would yield more comprehensible, fine- grained results on the influence of walkable built environments on sales performance. Scale also ultimately plays an important role on the way in which the built environment is measured; an interesting avenue for future research could test built environment measures at different scales to see which makes most sense to use for modeling in hierarchical or other contexts.

In the Introduction, each transportation mode makes a theoretical economic impact on specific types of businesses, each of which needs to be fleshed out empirically in more detail.

While this can be done in a variety of ways, a particularly interesting next step would be to develop work that explores connections between the very recent large-scale pedestrian and bicycle infrastructure improvements built in places like Chicago and Atlanta, i.e., the 606 and the

Beltline, and adjacent new business creation. Since these ‘trails’ have discrete nodes of entry and exit, there is certainly the potential for mode-specific clustering of new businesses.

Another avenue for future exploration of the theoretical discussion in the Introduction is to investigate in more detail the posited mechanisms that create agglomerative potential for businesses in certain industries. It would be interesting to better understand the extent to which transit stations directly influence the social connections and face-to-face contact that provide concrete economic benefits for businesses in the knowledge industries. If transit is shown to directly increase face-to-face interactions, perhaps it could also play an important role in fostering innovation and entrepreneurial ecosystems. In any case, additional work on the direct mechanisms of agglomeration is needed to provide more specific information to urban

132 researchers and planners than the generally-understood idea that larger cities accrue economic benefits due to size. Better understanding the specific ways in which size (which is, after all, a proxy for many related factors) increases economic benefits (and to which types of businesses) would allow policy-makers to directly target the most valuable components of agglomeration, and perhaps provide direction even to smaller cities or rural areas on concrete, valuable ways to increase economic outcomes.

Beyond the quantitative analysis presented here, additional, in-depth qualitative work on the relationship between new business creation and active transportation features is also needed.

While quantitative work can provide (useful) statistical estimates of relationships, another, equally useful way to assess this question is to ask entrepreneurs directly the extent to which various active transportation features influenced their location decisions. Participant observation of these links – particularly in areas where walkability, transit access, and other features of the built environment overlap – would also be extremely helpful to provide a more comprehensive context for interpreting the quantitative results observed here. Both of these approaches could be used to revisit and expand on the work presented in this dissertation, e.g., around the Phoenix light rail line.

Finally, there remains a need to investigate the influence of the large-scale forces mentioned in the Limitations section above, such as the ‘death’ of neighborhood retail, economic restructuring, the ‘back-to-the-city’ movement, historical and path-dependent development patterns, and direct federal investment on patterns of both transportation investment and new business creation. While each of these factors could potentially serve as a path for a completely new line of research, some more conceptual work that ties these trends together and discusses their impact on new business creation and transit development in this dissertations’ regions of

133 interest would provide a much-needed frame for interpreting the results of this project. Most researchers and practitioners understand that urban economic development is a political and path-dependent process, but the specifics of a region’s story on how these large-scale forces directly influence the resulting pattern of investment and disinvestment are often missing. These case studies are needed to provide a better understanding of the complex interactions between global trends, local trajectories, and economic behavior.

Summary

Overall, this dissertation provides an analysis of the relationship between active transportation systems – specifically, rail transit and walkable neighborhoods – and economic development, measured by new business creation and performance. It uses three distinct econometric methods to assess this relationship using fine-grained business micro-data from the

National Establishment Time Series. The results show that transit proximity is significantly related to new business creation, and that built environment features are significantly related to individual business performance. However, there is substantial regional heterogeneity in these relationships, due in part to transit network coverage and regional land use patterns and economic behavior. From a theoretical standpoint, this dissertation provides a unique quantitative investigation – and evidence for – the role of design in economic development outcomes, theorizes and tests the direct relationship between specific transportation modes and types of businesses, and presents a range of econometric methods for studying the interaction between neighborhood-scale features and individual business micro-data. From an applied perspective, this work provides useful information for practicing planners and policy-makers: active transportation investments can produce measurable economic benefits to businesses.

Transit networks can accomplish this goal even in sprawling, auto-centric cities, but are more

134 likely to do so if they provide extensive regional access; built environment features also significantly relate to business performance, but in some cases (such as for retail businesses) auto-oriented development patterns predict higher business sales volumes due to the nature of these businesses and regional transportation norms.

All in all, this project provides an important first look at the quantifiable relationships between active transportation systems and business development. Given the disruptive nature of climate change, economic restructuring, technological development, and increasing inequality, better understanding how to plan sustainable cities that balance all three goals – environmental, social, and economic – will become increasingly important in the years to come. Hopefully, by providing a general framework for studying the economic benefits of active transportation investments, and providing some initial findings of interest, this dissertation can serve as an initial guide to future research on urban and economic development.

135

APPENDICES

136

APPENDIX A: Chapter 1 Supplementary Materials

Chapter 1 Supplementary Materials

137

Appendix A.1. Table showing the pre-intervention (1990-2008) linear equations for the average number of new businesses per block per year.

Possible Treatment Areas Possible Control Areas Type of New Business .25mi .5mi 1mi 2.5mi 5mi 7.5mi Retail (NAICS 44-45) 0.0032x - 6.2865 0.0022x - 4.2392 0.002x - 3.8632 0.0017x - 3.4357 0.0017x - 3.3635 0.0017x - 3.3942

Services (NAICS 72 & 81) 0.0061x - 11.953 0.0043x - 8.4022 0.0031x - 6.159 0.0018x - 3.632 0.0019x - 3.6757 0.0019x - 3.839

Knowledge (NAICS 51-52 & 54-55) 0.0338x - 66.986 0.0135x - 26.671 0.0095x - 18.82 0.0065x - 12.876 0.0064x - 12.684 0.0066x - 13.172

Appendix A.2. Table showing differences in pre-intervention (1990-2008) linear slope coefficients in the average number of new businesses per block per year for nine possible boundary combinations. The boundary combinations with the lowest value – and thus most similar pre-intervention trend – are highlighted.

Possible Boundary Combinations (Treatment/Control Areas) .25mi / .25mi / .25mi / .5mi / .5mi / .5mi / 1mi / 1mi / 1mi / Type of New Business 2.5mi 5 mi 7.5mi 2.5mi 5mi 7.5mi 2.5mi 5mi 7.5mi Retail (NAICS 44-45) 0.0015 0.0015 0.0015 0.0005 0.0005 0.0005 0.0003 0.0003 0.0003 Services (NAICS 72 & 81) 0.0043 0.0042 0.0042 0.0025 0.0024 0.0024 0.0013 0.0012 0.0012 Knowledge (NAICS 51-52 & 54-55) 0.0273 0.0274 0.0272 0.007 0.0071 0.0069 0.003 0.0031 0.0029

138

Appendix A.3. Graphs showing the pre-intervention (1990-2008) linear trends and equations for the average number of new businesses per block per year for the chosen (most-similar) boundary combination: 1mile / 7.5mile.

139

Appendix A.4. Table describing the independent variables used in the final AITS model.

Years Category Name Description Source Available AITS variables (α) Dummy for post-intervention observations in DPOSTIMP 1990-2014 Author's creation treatment area Variables of interest Trend for post-intervention observations in TRPOSTIMP 1990-2014 Author's creation treatment area (e.g., 1 = 2009, 2 = 2010, etc.) Pre-intervention DIMP Dummy for all observations in treatment area 1990-2014 Author's creation absolute difference and Trend for all observations in treatment area TRIMP 1990-2014 Author's creation trend controls (e.g., 1 = 1990, 2 = 1991, etc.) TRALL Trend for all observations in both areas 1990-2014 Author's creation Overall trend controls Trend for all post-intervention observations in TRPOSTALL 1990-2014 Author's creation both areas Exposure (훿) ACRES Size of block in acres 2010 Census TIGER/Line shapefiles Covariates (x) 1990, 2000, Overall economic POP Population Decennial Census 2010 characteristics: market 1990, 2000, National Establishment Time Series size and jobs EMPD Employment density (per acre) 2010 (NETS) National Establishment Time Series RETP % existing retail (NAICS 44-45) businesses 1990-2014 (NETS) % existing manufacturing (NAICS 31-33) National Establishment Time Series MANP 1990-2014 Business dynamics businesses (NETS) % existing information (NAICS 51) National Establishment Time Series INFP 1990-2014 businesses (NETS) BUS_D Location in a downtown business district21 2015 Author's creation

21Defined as the central business districts of Phoenix, Tempe, Mesa, Scottsdale, and Chandler. These boundaries were obtained either from identified district boundaries – as in the case of Phoenix, which uses the “downtown core” neighborhood definition of Downtown Phoenix, Inc. (roughly a square area from Filmore Street to the Union Pacific railroad tracks on the south, and 3rd Avenue to 7th Street on the east) (2015) – or were digitized from Google Maps by drawing polygons around the major streets bounding the central commercial areas of each downtown. For Tempe, the downtown area includes all of the Tempe campus of Arizona State University, as well as downtown Mill Avenue, roughly from Rio Salado Avenue to Apache Road on the south, and Mill Avenue to Rural Road on the east. For Mesa, the area includes everything between Country Club Drive and Central Avenue and W. 1st Street and W 1st Avenue. The 140

Appendix A.4 (cont’d). Table describing the independent variables used in the final AITS model.

INT_DEN Number of street intersections per acre 2015 Census TIGER/Line shapefiles Built environment BLOCK_L Block length in meters 2010 Census TIGER/Line shapefiles 1990, 2000, BLK Black non-Hispanic population % Decennial Census 2010 Demographics 1990, 2000, MID % population 19-64 Decennial Census 2010 Spatial lag of the average value (over the time Average: 1990- AVRETLAG period of the entire study) of new retail Author's creation 2014 businesses Spatial lag of the average value (over the time Average: 1990- Spatial Lags22 AVSERVLAG period of the entire study) of new service Author's creation 2014 businesses Spatial lag of the average value (over the time Average: 1990- AVKNOWLAG period of the entire study) of new knowledge Author's creation 2014 businesses

Scottsdale definition includes the “Entertainment District” and “Downtown Scottsdale” neighborhoods, roughly between N. Drinkwater and N. Goldwater Boulevards and Camelback Road and Osborn Road. For Chandler, the boundary runs on both sides of S. Arizona Avenue between E. Frye Road and E. Chandler Boulevard from S. Delaware Street on the east to the San Marcos Golf Course on the west. Any blocks that intersected the boundaries of these polygons were considered within a downtown business district and given a value of 1 for the variable. 22 All spatial lags were computed using a first-order queen contiguity spatial weight matrix. 141

Appendix A.5. Descriptive statistics for treatment and control areas.

Blocks in Light Rail Treatment (1-mile) Area Blocks in Automobile Control (7.5-mile) Area Variable Obs Mean Std. Dev. Min Max Obs Mean Std. Dev. Min Max Dependent variables RETAIL 50075 0.05 0.31 0 15 990175 0.03 0.31 0 36 KNOW 50075 0.17 1.55 0 258 990175 0.08 0.53 0 104 SERVE 50075 0.07 0.43 0 51 990175 0.04 0.30 0 60 AITS variables (α) DIMP 50075 1 0 1 1 990175 0 0 0 0 DPOSTIMP 50075 0.28 0.45 0 1 990175 0 0 0 0 TRIMP 50075 13 7.21 1 25 990175 0 0 0 0 TRPOSTIMP 50075 1.12 2.08 0 7 990175 0 0 0 0 TRALL 50075 13 7.21 1 25 990175 13 7.21 1 25 TRPOSTALL 50075 1.12 2.08 0 7 990175 1.12 2.08 0 7 Covariates (x)23 AVRETLAG 50075 0.07 0.10 0 0.96 990175 0.06 0.15 0 6.60 AVKNWLAG 50075 0.20 0.37 0 3.48 990175 0.13 0.39 0 22.84 AVSRVLAG 50075 0.09 0.11 0 0.87 990175 0.06 0.14 0 7.08 EMPD 6009 10.92 73.26 0 2799.91 118821 1.83 30.32 0 5591.71 RETP 6009 7.0% 18.9% 0% 100% 118821 4.2% 16.1% 0% 100% MANP 6009 2.8% 12.5% 0% 100% 118821 1.6% 10.0% 0% 100% INFP 6009 1.5% 8.9% 0% 100% 118821 0.7% 6.7% 0% 100% BUS_D 50075 7.9% 27.0% 0% 100% 990175 0.4% 6.4% 0% 100% INT_DEN 50075 1.95 4.89 0 92.67 990175 2.03 11.64 0 1841.70 BLOCK_L 50075 2594.50 1833.79 35.56 19862.70 990175 3094.89 3040.69 45.18 90356.70 BLK 6009 2.9% 8.4% 0% 100% 118821 1.9% 6.6% 0% 100% MID 6009 38.7% 35.5% 0% 100% 118821 36.0% 31.8% 0% 100% Offset (δ) ACRES 50075 8.38 15.58 0 314.36 990175 12.82 64.71 0 4855.92

23Note: The number of observations for several of these variables (e.g., EMPD, RETP, MANP, etc.) are lower, due to the fact that they are measured at only three points in time – 1990, 2000, and 2010. This ensures that the descriptive statistics shown for these variables are correctly divided only by three year-observations, rather than by twenty-five. 142

Appendix A.6. Descriptive statistics for pre- and post-intervention periods for the 1-mile treatment area.

Pre-Intervention (1990-2007) Post-Intervention (2008-2014) Variable Obs Mean Std. Dev. Min Max Obs Mean Std. Dev. Min Max New establishments Retail new starts 36054 0.057 0.34 0 15 14021 0.039 0.22 0 5 Knowledge new starts 36054 0.165 1.67 0 258 14021 0.180 1.20 0 59 Service new starts 36054 0.071 0.46 0 51 14021 0.072 0.33 0 8 Demographic and Business Characteristics Population 4006 56.83 137.51 0 2105.00 2003 58.96 158.27 0 3316.00 Employment density (per acre) 4006 10.34 76.92 0 2799.91 2003 12.09 65.31 0 1181.86 Percent existing retail businesses 4006 6.8% 19.0% 0% 100% 2003 7.4% 18.6% 0% 100% Percent existing manufacturing businesses 4006 3.3% 13.7% 0% 100% 2003 2.0% 9.6% 0% 100% Percent existing information businesses 4006 1.4% 8.6% 0% 100% 2003 1.8% 9.6% 0% 100% Black non-Hispanic population percent 4006 2.6% 8.4% 0% 100% 2003 3.6% 8.3% 0% 100% Percent population aged 19-64 4006 36.7% 34.8% 0% 100% 2003 42.7% 36.4% 0% 100%

143

APPENDIX B: Chapter 2 Supplementary Materials

Chapter 2 Supplementary Materials

144

Appendix B.1. Robust Lagrange multipliers for each business type model.

Robust Lagrange Multipliers Business type Error Sig. Lag Sig. Knowledge 3560.5 2.20E-16 3070.7 2.20E-16 High technology 1145.2 2.20E-16 328.63 2.20E-16 Retail/services 4047.2 2.20E-16 1206.6 2.20E-16 Producer services 3501.8 2.20E-16 1567.4 2.20E-16

145

Appendix B.2. Descriptive statistics for variables used in models.

Variables Type Name Description Min 1st Quartile Median Mean 3rd Quartile Max NS_KNOW Number of new knowledge starts (2011) 0 0 0 0.096 0 101 Dependent NS_HTECH Number of new high tech starts (2011) 0 0 0 0.015 0 17 Variables NS_RETSV Number of new retail, services, and food starts (2011) 0 0 0 0.046 0 26 NS_PRODS Number of new producer services starts (2011) 0 0 0 0.050 0 26 LR25 Location within 1/4 mile of light rail station 0 0 0 0.028 0 1 LR5 Location 1/4 to 1/2 mile of light rail station 0 0 0 0.026 0 1 Rail Transit HR25 Location within 1/4 mile of heavy rail station 0 0 0 0.024 0 1 Variables HR5 Location 1/4 to 1/2 mile of heavy rail station 0 0 0 0.037 0 1 CR25 Location within 1/4 mile of commuter rail station 0 0 0 0.029 0 1 CR5 Location 1/4 to 1/2 mile of heavy rail station 0 0 0 0.064 0 1 Tract-Level PCTBACH Percent Bachelor's degree attainment or higher 0 0.174 0.325 0.358 0.515 1 Covariates VEHPERHH Number of vehicles per household 0 1.530 1.797 1.740 2.051 3.240 RACE_DIV Racial diversity (Herfindahl Index) 0 0 0.6 0.529 0.904 1 POP Population 0 0 30 60.330 77 4535 Block-Level ExYr11 Number of existing businesses per block (2011) 0 0 2 4.810 5 1007 Covariates NS_ALL Total number of new starts per block (2011) 0 0 0 0.341 0 358 Acres Size of block in acres 0 1.95 4.44 26.190 11.1 68599.4 Lagged Avg. number of new knowledge starts (2011) in dependent LAG_KNOW neighbors 0 0 0 0.086 0.143 31.857 variables Avg. number of new high tech starts (2011) in (using 7 LAG_HTCH neighbors 0 0 0 0.014 0 2.429 nearest- Avg. number of new retail, services, and food starts neighbor LAG_RET (2011) in neighbors 0 0 0 0.041 0 5.286 weights Avg. number of new producer services starts (2011) matrix) LAG_PROD in neighbors 0 0 0 0.045 0 6.429

146

Appendix B.3. Results of the pooled and region-specific Spatial Durbin models for knowledge businesses.

Pooled San Jose Business Variable Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.0615 0.0787 0.3498 -0.0670 LR5 0.0446 HR25 0.0049 0.0295 0.5025 0.0605 HR5 0.0075 0.0352 0.5658 0.0664 CR25 0.1021 0.1109 0.1791 -0.0664 0.0413 0.0519 0.2282 CR5 0.0523 0.0677 0.3133 RACE_DIV -0.0249 -0.0315 -0.1344 0.0045 Knowledge PCTBACH -0.0440 -0.0426 0.0273 0.0421 (NAICS 51- VEHPERHH 0.0102 0.0044 -0.1189 -0.0243 -0.0640 -0.0855 -0.4637 52 and 54-55) POP -0.0002 -0.0002 0.0001 0.0002 -0.0001 -0.0001 0.0001 0.0001 ExYr11 -0.0007 -0.0004 0.0059 0.0014 NS_ALL 0.0065 0.0158 0.1889 0.0186 δ 0.8797 0.8697 N 95889 14681 Nagelkerke pseudo-R² 0.7750 0.7106 Seed* 1111 0 Austin Cleveland Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.1635 0.1507 -0.2527 -0.1796 LR5 HR25 HR5 CR25 CR5 0.0732 0.0794 0.1275 -0.0473 RACE_DIV 0.0181 0.0382 0.4197 0.0381 0.0220 0.0332 0.2205 0.0179 Knowledge PCTBACH -0.1498 -0.1533 -0.0701 0.1222 (NAICS 51- VEHPERHH -0.0171 -0.0257 -0.1799 -0.0433 52 and 54-55) POP -0.0001 -0.0001 -0.0001 0.0001 -0.0002 -0.0003 -0.0017 -0.0001 ExYr11 -0.0006 -0.0008 -0.0039 0.0013 NS_ALL 0.0086 0.0291 0.4260 0.0471 0.0025 0.0066 0.0802 0.0111 δ 0.8796 0.8436 N 12821 7632 Nagelkerke pseudo-R² 0.7397 0.6809 Seed* 11 11 Philadelphia Boston Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.0844 0.1037 0.3772 -0.0217 0.0422 0.0711 0.5577 0.0469 LR5 0.0518 0.0590 HR25 0.0547 0.1023 HR5 0.0387 0.0697 0.6059 0.0494 0.1189 CR25 0.0776 0.0914 0.2707 0.1382 0.1449 0.1288 -0.0994 CR5 0.0290 0.0483 0.3786 0.0267 0.0817 0.0954 0.2622 -0.0311 RACE_DIV -0.0106 -0.0110 -0.0087 0.0080 -0.0572 -0.0623 -0.0979 0.0345 Knowledge PCTBACH -0.0748 -0.0860 -0.2192 0.0350 -0.0201 -0.0305 -0.2012 (NAICS 51- VEHPERHH 0.0184 0.0154 -0.0599 -0.0242 -0.0655 52 and 54-55) POP -0.0003 -0.0003 -0.0005 0.0002 -0.0004 -0.0004 -0.0003 0.0003 ExYr11 -0.0012 -0.0013 -0.0025 0.0007 -0.0017 -0.0029 -0.0236 -0.0021 NS_ALL 0.0090 0.0196 0.2081 0.0207 0.0210 0.0555 0.6638 0.0809 δ 0.8717 0.8599 N 32652 28103 Nagelkerke pseudo-R² 0.7857 0.8026 Seed* 111 0 *Used for Monte Carlo approximate log-determinant method of weights matrix decomposition in the "lagsarlm" function of the "spdep" R package used to estimate these models. 7 nearest-neighbor weights matrix used for all models. All reported coefficients significant at p < .05.

147

Appendix B.4. Results of the pooled and region-specific Spatial Durbin models for high tech businesses.

Pooled San Jose Business Variable Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.0854 LR5 0.0878 HR25 HR5 0.1089 0.1362 0.4554 CR25 0.1854 0.1925 0.1187 -0.1026 CR5 0.1064 0.1263 0.3310 RACE_DIV -0.0482 -0.0575 -0.1554 PCTBACH 0.0080 0.0024 -0.0948 -0.0327 High technology VEHPERHH 0.0100 0.0007 -0.1539 -0.0507 POP -0.0006 -0.0006 -0.0001 0.0004 -0.0004 -0.0004 -0.0001 0.0002 ExYr11 -0.0027 -0.0031 -0.0052 0.0006 -0.0049 -0.0052 NS_ALL 0.0263 0.0302 0.0651 -0.0009 0.0431 0.0457 δ 0.7338 0.6738 N 26725 5953 Nagelkerke pseudo-R² 0.5461 0.3979 Seed* 1 9 Austin Cleveland Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 LR5 0.1457 HR25 HR5 0.1139 0.1283 0.2413 CR25 CR5 RACE_DIV 0.1106 PCTBACH -0.2074 -0.2419 -0.6154 High technology VEHPERHH POP -0.0005 -0.0006 -0.0008 -0.0007 -0.0007 -0.0010 ExYr11 -0.0027 -0.0025 0.0025 0.0027 -0.0026 -0.0033 -0.0120 -0.0027 NS_ALL 0.0292 0.0316 0.0418 0.0247 0.0289 0.0705 δ 0.6786 0.6543 N 3806 1990 Nagelkerke pseudo-R² 0.4526 0.4826 Seed* 11 9 Philadelphia Boston Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 LR5 0.2096 0.1220 0.1397 0.2861 HR25 0.1915 0.2163 0.4027 0.1531 HR5 0.1974 0.2219 0.3965 0.1556 CR25 0.2276 0.2330 0.0882 -0.1346 0.2260 0.2279 0.0321 -0.1525 CR5 0.1602 0.1783 0.2942 0.1346 0.1472 0.2034 RACE_DIV 0.0516 -0.0958 -0.0989 -0.0510 0.0534 PCTBACH -0.1471 High technology VEHPERHH 0.0549 0.0505 -0.0708 -0.0607 -0.1554 POP -0.0008 -0.0009 -0.0005 0.0005 -0.0009 -0.0009 0.0000 0.0006 ExYr11 -0.0021 -0.0025 -0.0067 -0.0051 -0.0054 -0.0055 NS_ALL 0.0185 0.0229 0.0709 0.0086 0.0622 0.0673 0.0833 δ 0.7107 0.7173 N 7622 7354 Nagelkerke pseudo-R² 0.5685 0.6361 Seed* 11 7 *Used for Monte Carlo approximate log-determinant method of weights matrix decomposition in the "lagsarlm" function of the "spdep" R package used to estimate these models. 7 nearest-neighbor weights matrix used for all models. All reported coefficients significant at p < .05.

148

Appendix B.5. Results of the pooled and region-specific Spatial Durbin models for retail/services businesses.

Pooled San Jose Business Variable Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.0822 0.1013 0.3587 -0.0232 0.0876 0.0780 -0.1900 -0.1045 LR5 0.0413 0.0549 0.2560 HR25 0.0570 0.0845 0.5176 HR5 0.0498 0.0827 0.6193 0.0403 CR25 0.1096 0.1198 0.1902 -0.0698 0.1161 0.1264 0.2042 CR5 0.0598 0.0764 0.3118 RACE_DIV -0.0328 -0.0389 -0.1149 0.0131 0.0291 Retail, PCTBACH -0.0403 -0.0514 -0.2089 0.0069 services, and VEHPERHH 0.0022 -0.0045 -0.1271 -0.0191 -0.0499 -0.0736 -0.4668 food POP -0.0003 -0.0003 -0.0003 0.0002 -0.0002 -0.0002 -0.0004 0.0001 ExYr11 -0.0013 -0.0014 -0.0017 0.0009 -0.0014 -0.0013 NS_ALL 0.0067 0.0135 0.1282 0.0115 0.0077 0.0142 δ 0.8739 0.8495 N 62240 9347 Nagelkerke pseudo-R² 0.7775 0.6734 Seed* 19 10 Austin Cleveland Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 LR5 HR25 HR5 CR25 CR5 RACE_DIV Retail, PCTBACH -0.0966 -0.1223 -0.4939 -0.1170 -0.1325 -0.2920 services, and VEHPERHH -0.0283 -0.0386 -0.1994 -0.0291 food POP -0.0002 -0.0002 -0.0005 0.0001 -0.0002 -0.0003 -0.0027 -0.0003 ExYr11 -0.0016 -0.0018 -0.0037 0.0009 -0.0006 -0.0006 -0.0003 NS_ALL 0.0117 0.0256 0.2673 0.0255 0.0035 0.0075 0.0763 0.0102 δ 0.8751 0.8383 N 8201 5577 Nagelkerke pseudo-R² 0.7489 0.6802 Seed* 7 10 Philadelphia Boston Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.0840 0.1043 0.3756 0.1209 0.1485 0.4779 LR5 0.0370 0.0551 0.3357 HR25 0.0683 0.1005 0.5972 0.0600 HR5 0.0649 0.1028 0.7014 0.0382 0.0581 CR25 0.1065 0.1182 0.2176 -0.0634 0.1150 0.1274 0.2148 -0.0645 CR5 0.0556 0.0722 0.3069 0.0646 0.0872 0.3933 RACE_DIV 0.0028 -0.0880 -0.0936 -0.0970 0.0599 Retail, PCTBACH -0.0626 services, and VEHPERHH 0.0004 -0.0025 -0.0535 -0.0076 0.0394 0.0216 -0.3094 -0.0819 food POP -0.0003 -0.0004 -0.0005 0.0002 -0.0004 -0.0004 0.0001 0.0004 ExYr11 -0.0016 -0.0021 -0.0091 -0.0020 -0.0034 -0.0242 -0.0020 NS_ALL 0.0090 0.0170 0.1491 0.0123 0.0175 0.0440 0.4602 0.0570 δ 0.8740 0.8537 N 24207 14908 Nagelkerke pseudo-R² 0.7991 0.8024 Seed* 1190 11 *Used for Monte Carlo approximate log-determinant method of weights matrix decomposition in the "lagsarlm" function of the "spdep" R package used to estimate these models. 7 nearest-neighbor weights matrix used for all models. All reported coefficients significant at p < .05.

149

Appendix B.6. Results of the pooled and region-specific Spatial Durbin models for producer services.

Pooled San Jose Business Variable Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.0836 0.1021 0.3466 LR5 HR25 0.0664 HR5 0.0704 CR25 0.0922 0.1040 0.2215 -0.0496 CR5 0.0504 0.0724 0.4154 0.0135 RACE_DIV -0.0177 -0.0197 -0.0385 0.0337 0.0461 0.2479 PCTBACH -0.0196 -0.0269 -0.1377 Producer services VEHPERHH -0.0009 -0.0062 -0.1011 -0.0770 -0.0979 -0.4188 POP 0.0002 -0.0002 -0.0002 -0.0004 ExYr11 -0.0016 -0.0015 NS_ALL 0.0082 0.0151 δ 0.8713 0.8591 N 67139 10332 Nagelkerke pseudo-R² 0.7690 0.6762 Seed* 110 110 Austin Cleveland Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.2457 0.2321 -0.2504 -0.2485 LR5 0.1097 0.1439 0.6267 HR25 HR5 CR25 CR5 RACE_DIV PCTBACH Producer services VEHPERHH POP -0.0002 -0.0002 -0.0003 0.0001 -0.0002 -0.0003 -0.0016 ExYr11 -0.0009 -0.0010 -0.0033 -0.0007 -0.0007 -0.0001 NS_ALL 0.0082 0.0237 0.3020 0.0342 0.0041 0.0083 0.0777 0.0090 δ 0.8718 0.8484 N 9819 5659 Nagelkerke pseudo-R² 0.7270 0.7059 Seed* 9 110 Philadelphia Boston Estimate Direct Indirect Lag Estimate Estimate Direct Indirect Lag Estimate LR25 0.1419 0.1625 0.3749 0.1059 0.1310 0.4516 LR5 HR25 0.0726 HR5 0.0597 CR25 0.0922 0.1085 0.2961 0.1128 0.1220 0.1643 -0.0717 CR5 0.0716 0.0922 0.3714 RACE_DIV 0.0015 0.0044 0.0535 0.0065 -0.0591 -0.0601 -0.0177 PCTBACH -0.0323 -0.0530 -0.3779 -0.0273 Producer services VEHPERHH 0.0082 0.0058 -0.0432 -0.0134 -0.0491 POP -0.0004 -0.0004 -0.0004 0.0003 -0.0004 -0.0005 -0.0005 0.0003 ExYr11 -0.0017 -0.0019 -0.0035 0.0009 -0.0018 -0.0029 -0.0199 NS_ALL 0.0097 0.0163 0.1218 0.0095 0.0171 0.0423 0.4535 δ 0.8633 0.8577 N 22705 18624 Nagelkerke pseudo-R² 0.7865 0.8025 Seed* 1110 7 *Used for Monte Carlo approximate log-determinant method of weights matrix decomposition in the "lagsarlm" function of the "spdep" R package used to estimate these models. 7 nearest-neighbor weights matrix used for all models. All reported coefficients significant at p < .05.

150

Appendix B.7. Results of the pooled and region-specific Poisson models for all business types.

Pooled San Jose Austin Cleveland Philadelphia Boston Business Type Variable Estimate Estimate Estimate Estimate Estimate Estimate LR25 1.3300 0.6886 1.4710 0.8077 1.2310 LR5 1.3740 0.2287 1.3570 0.7907 0.8414 HR25 1.2980 1.0920 1.2840 1.6930 HR5 0.9320 0.4775 1.2490 1.6270 CR25 1.1470 1.2300 2.0590 1.2410 1.4470 CR5 1.2660 0.6239 0.3744 1.3700 0.9986 Knowledge RACE_DIV -0.7794 -0.8913 -0.2576 -0.7913 -0.4084 (NAICS 51-52 PCTBACH 0.8295 2.1770 1.0080 -0.1604 -0.1315 and 54-55) VEHPERHH -0.1784 -0.8305 0.1710 -0.1866 -0.0859 -0.2555 POP 0.0001 0.0002 -0.0003 -0.0003 ExYr11 0.0061 -0.0032 0.0024 0.0059 0.0034 -0.0170 NS_ALL 0.0089 0.1370 0.0874 0.0053 0.0221 0.3022 7NN - lagged DV 0.2494 0.6802 0.7617 0.4039 0.2368 0.8032 N 227140 22345 25003 15334 85559 78899 LR25 1.4830 1.0020 1.4390 0.8590 1.2590 LR5 1.5600 0.5110 1.1270 0.6000 1.3020 HR25 1.2020 1.8120 1.9750 HR5 0.8230 1.2000 1.8940 CR25 1.2070 1.4770 2.6470 1.4380 1.1240 CR5 0.9980 0.6590 0.9520 1.2240 0.6960 High RACE_DIV -0.9910 -0.5870 -0.9010 -0.7330 -0.5640 technology PCTBACH 1.3200 2.7300 0.9940 0.3360 VEHPERHH -0.1100 -0.7600 0.1770 -0.1950 POP 0.0001 0.0002 -0.0010 0.0000 ExYr11 0.0070 0.0040 0.0080 0.0050 -0.0180 NS_ALL 0.0070 0.1390 0.0700 0.0110 0.3090 7NN - lagged DV 1.2140 0.8010 4.6180 1.7200 0.5480 3.1140 N 227140 22345 25003 15334 85559 78899 LR25 1.2500 0.7440 1.3430 0.7720 1.5320 LR5 1.1390 1.0990 0.7590 1.0280 HR25 1.3820 0.9480 1.8420 1.4110 HR5 1.1720 0.4420 1.5720 1.3680 CR25 1.0810 1.4660 2.4900 0.8890 1.5200 CR5 1.0750 0.5830 0.6470 0.9940 1.0010 Retail, services, RACE_DIV -0.8040 -0.2510 -1.2750 -0.2570 -0.7870 -0.5580 and food PCTBACH 0.3910 1.6620 0.3990 -0.5250 -0.3330 -0.4540 VEHPERHH -0.2240 -0.9320 0.1110 -0.1330 -0.0740 -0.2320 POP 0.0003 -0.0010 -0.0002 ExYr11 0.0060 -0.0030 0.0040 0.0070 0.0040 -0.0160 NS_ALL 0.0070 0.1460 0.0750 0.0030 0.0120 0.3000 7NN - lagged DV 1.0710 0.5700 1.0670 0.4910 0.9400 1.5320 N 227140 22345 25003 15334 85559 78899 LR25 1.2950 0.7430 1.4120 1.0840 1.2210 LR5 1.3730 0.3430 1.4030 0.9380 0.8390 HR25 1.3070 1.2340 1.5130 1.7630 HR5 1.0760 0.4960 1.3120 1.6040 CR25 0.9360 1.2720 1.9950 1.2860 1.3800 CR5 1.2930 0.5690 0.7220 1.3520 1.1100 Producer RACE_DIV -0.5780 0.0860 -0.7190 -0.2640 -0.4510 -0.2860 services PCTBACH 0.6700 1.9180 0.8950 -0.1470 VEHPERHH -0.2130 -0.8880 0.1740 -0.1960 -0.0560 -0.2760 POP 0.0002 0.0002 -0.0004 -0.0003 ExYr11 0.0060 -0.0020 0.0020 0.0070 0.0040 -0.0170 NS_ALL 0.0080 0.1260 0.0860 0.0030 0.0120 0.3030 7NN - lagged DV 0.6230 1.3970 1.1890 0.6520 0.4130 1.3120 N 227140 22345 25003 15334 85559 78899 All reported coefficients significant at p < .05. Models estimated with log(acres) as offset.

151

APPENDIX C: Chapter 3 Supplementary Materials

Chapter 3 Supplementary Materials

152

Appendix C.1. Descriptive statistics and Variance Inflation Factors (VIF).

Level Variable Name Phoenix Boston Final Final VIF Confd. N Mean Min. Max. VIF Confd. N Mean Min. Max. Model Model LOGSALES - 27185 5.07 -0.29 8.86 26513 11.13 2.91 17.45 RETAIL 1.08 X 27185 0.17 0 1 1.05 X 26513 0.09 0 1 Level-1 MAN 1.07 X 27185 0.06 0 1 1.01 26513 0.02 0 1 PROF_OFF 1.14 X 27185 0.19 0 1 1.19 X 26513 0.27 0 1 MEDAGE 5.06 357 33.22 0 55.30 1.39 DENSITY 175 35.89 0 76.30 WHTNH 30.64 357 0.48 0 0.95 43.12 175 0.48 0 1 BLKNH 2.45 X 357 0.06 0 0.38 22.74 175 0.22 0 0.90 ASNNH 1.99 X 357 0.03 0 0.36 7.61 175 0.08 0 0.70 HISP 11.90 357 0.39 0 0.94 8.45 175 0.17 0 0.72 BACH 6.97 357 0.25 0 0.73 5.94 175 0.43 0 1 TRANS 1.74 X 357 0.04 0 0.27 2.97 X 175 0.32 0 0.73 DENSITY 2.12 X 357 10.76 0.02 54.07 4.27 X 175 54.88 0.03 262.38 Level-2 AVG_SHAP 1.88 X 357 3773 1985 16588 3.96 DENSITY 175 1986 969 5035 B_AGE_HI 1.66 X 357 0.37 0 0.98 1.67 X 175 0.41 0 1 MIX 1.32 TRANS 357 0.03 0 1 2.43 X 175 0.10 0 1 PEDB 1.80 X 357 0.03 0 0.30 6.49 175 0.16 0 0.66 TRANSLAG - 2.64 X 175 0.32 0.10 0.62 POPEDLAG - 5.71 175 52.14 5.94 183.94 MEDALAG 10.29 357 33.44 22.68 47.85 - WHITELAG 22.04 357 0.49 0.08 0.92 - BACHLAG 9.65 357 0.25 0.04 0.61 -

153

REFERENCES

154

REFERENCES

Acs, Z. J., & Audretsch, D. B. (1990). Innovation and Small Firms. Cambridge, MA: MIT Press.

Acs, Z. J., & Audretsch, D. B. (2003). Innovation and Technological Change. In Z. J. Acs & D. B. Audretsch (Eds.), Handbook of Entrepreneurship Research. New York, NY: Springer.

Acs, Z. J., Anselin, L., & Varga, A. (2002). Patents and innovation counts as measures of regional production of new knowledge. Research Policy, 31(7), 1069-1085.

Addison, C., Zhang, S., & Coomes, B. (2013). Smart growth and housing affordability: A review of regulatory mechanisms and planning practices. Journal of Planning Literature, 28(3), 215-257.

Agostini, C. A., & Palmucci, G. (2008). The Anticipated Capitalization Effect of a New Metro Line on Housing Prices. Fiscal Studies, 29, 233-56.

Aharonson, B. S., Baum, J. A. C., & Feldman, M. P. (2007). Desperately Seeking Spillovers? Increasing Returns, Social Cohesion and the Location of New Entrants in Geographic and Technological Space. Industrial and Corporate Change, 16, 89-130.

Aharonson, B. S., Stettner, U., Amburgey, T.L., Ellis, S., & Drori, I. (2013). Understanding the Relationship Between Networks and Technology, Creativity and Innovation. Bingley, UK: Emerald Group Publishing Limited.

Alexander, C. (1977). A Pattern Language: Towns, Buildings, Construction. New York, NY: Oxford University Press.

Alonso, W. (1960). A Theory of the Urban Land Market. Papers in Regional Science, 6(1), 149- 157.

American Planning Association (APA). (2014). Investing in Place: Two Generations’ View on the Future of Communities. American Planning Association. Retrieved from: https://www.planning.org/policy/polls/investing/pdf/pollinvestingreport.pdf.

American Public Transit Association (APTA). (1994). Glossary of Transit Terminology. American Public Transit Association. Retrieved from: http://www.apta.com/resources/reportsandpublications/Documents/Transit_Glossary_199 4.pdf.

Amin, A. (1994). Post-Fordism: Models, fantasies, and phantoms of transition. In A. Amin (Ed.), Post-Fordism: A Reader. Cambridge, MA: Blackwell.

Anselin, L. (1988). Spatial Econometrics: Methods and Models. Dordrecht, Germany: Kluwer Academic Publishers.

Anselin, L. (1995). Local Indicators of Spatial Association–LISA. Geographical Analysis, 27(2), 93-115.

155

Anselin, L, Kim, Y. W., & Syabri, I. (2004). Web-Based Analytical Tools for the Exploration of Spatial Data. Journal of Geographical Systems, 6, 197–218.

Anselin, L., & Rey, S. (2014). Modern Spatial Econometrics in Practice. Chicago, IL: GeoDa Press LLC.

Appleyard, D., & Lintell, M. (1972). The environmental quality of city streets: the residents' viewpoint. Journal of the American Institute of Planners, 38(2), 84-101.

Araldi, A., & Fusco, G. (2016). Urban Form from the Pedestrian Point of View: Spatial Patterns on a Street Network. Conference Proceedings from 9th International Conference on Innovation in Urban and Regional Planning (INPUT). September 2016: Turin, Italy. Retrieved from: http://www.input2016.it/conference_2016.

Asheim, B. T. (1996). Industrial Districts as “Learning Regions”: A Condition for Prosperity. European Planning Studies, 4, 379-400.

Asheim, B. T., Lawton, S. H., & Oughton, C. (2011). Regional innovation systems: theory, empirics and policy. Regional Studies, 45(7), 875-891.

Batty, M., Besussi, E., & Chin, N. (2003). Traffic, Urban Growth, and Suburban Sprawl. University College London Working Paper Series: Paper 70. Retrieved from: https://www.bartlett.ucl.ac.uk/casa/pdf/paper70.pdf.

Beckmann, M. (1968). Location Theory. New York, NY: Random House.

Bernstein, M. J., Wiek, A., Brundiers, K., Pearson, K., Minowitz, A., Kay, B., & Golub, A. (2016). Mitigating urban sprawl effects: a collaborative tree and shade intervention in Phoenix, Arizona, USA. Local Environment, 21(4), 414-431.

Besser, L., & Dannenberg, A. (2005). Walking to Public Transit: Steps to Help Meet Physical Activity Recommendations. American Journal of Preventive Medicine, 29(4), 273–280.

Birch, D. (1987). Job Creation in America: How Our Smallest Companies Put the Most People to Work. New York, NY: The Free Press.

Blackburn, R., & Ram, M. (2006). Fix or Fixation?? The Contributions and Limitations of Entrepreneurship and Small Firms to Combating Social Exclusion. Entrepreneurship and Regional Development, 18, 73-89.

Boarnet, M., & Crane, R. (2001). The influence of land use on travel behavior: Specification and estimation strategies. Transportation Research Record – A, 35(9), 823-845.

Bolin, B., Grineski, S., & Collins, T. (2005). The Geography of Despair: Environmental Racism and the Making of South Phoenix, Arizona, USA. Human Ecology Review, 12(2), 156- 168.

Bollinger, C. R., & Ihlanfeldt, K. R. (1997). The Impact of Rapid Rail Transit on Economic Development: The Case of Atlanta’s MARTA. Journal of Urban Economics, 42, 179- 204. 156

Boschma, R., & Frenken, K. (2006). Why is Economic Geography not an Evolutionary Science? Towards an Evolutionary Economic Geography. Journal of Economic Geography, 6, 273-302.

Boschma, R., & Lambooy, J. (1999). Evolutionary Economics and Economic Geography. Journal of Evolutionary Economics, 9, 411-429.

Bradley University. (2008). Choosing a Retail Location. Turner Center for Entrepreneurship. Retrieved from: http://www.bradley.edu/academic/colleges/fcba/centers/turner/established/resources/locat ion.dot.

Brueckner, J. K. (2000). Urban Sprawl: Diagnosis and remedies. International Regional Science Review, 23(2), 160-171.

Burchell, R. W., Listokin, D., & Galley, C. C. (2000). Smart growth: More than a ghost of urban policy past, less than a bold new horizon. Housing Policy Debate, 11(4), 821-879.

Calthorpe, P. (1993). The Next American Metropolis: Ecology, Community, and the American Dream. New York, NY: Princeton Architectural Press.

Camagni, R., Gibelli, M. C., & Rigamonti, P. (2002). Urban mobility and urban form: the social and environmental costs of different patterns of urban expansion. Ecological Economics, 40(2), 199-216.

Carree, M. A., & Thurik, A. R. (2003). The Impact of Entrepreneurship on Economic Growth. In Z. J. Acs & D. B. Audretsch (Eds.), Handbook of Entrepreneurship Research. New York, NY: Springer.

Cervero, R. (1984). Light Rail Transit and Urban Development. Journal of the American Planning Association, 50(2), 133-147.

Cervero, R. (1994). Rail Transit and Joint Development: Land Market Impacts in Washington, D.C. and Atlanta. Journal of the American Planning Association, 60(1), 83-94.

Cervero, R. (2004). Effects of Light and Commuter Rail Transit on Land Prices: Experiences in San Diego County. Journal of the Transportation Research Forum, 43(1), 121-38.

Cervero, R., & Kockelman, K. (1997). Travel Demand and the 3Ds: Density, Diversity, and Design. Transportation Research Part D, 2(3), 199-219.

Cervero, R., & Landis, J. (1997). Twenty Years of the Bay Area Rapid Transit System: Land Use and Development Impacts. Transportation Research Part A, 41(4), 309-333.

Chapple, K., Markusen, A., Schrock, G., Yamamoto, D., & Yu, P. (2004). Gauging metropolitan “high-tech” and “I-tech” activity. Economic Development Quarterly, 18(1), 10-29.

Chatman, D. G. (2013). Does TOD Need the T? Journal of the American Planning Association, 79(1), 17-31.

157

Chatman, D. G., & Noland, R. B. (2011). Do Public Transport Improvements Increase Agglomeration Economies? A Review of Literature and an Agenda for Research. Transport Reviews, 31(6), 725-742.

Chatman, D. G., Noland, R. B., & Klein, N. J. (2016). Firm births, access to transit, and agglomeration in Portland, Oregon, and Dallas, Texas. Transportation Research Record, 2598, 1-10.

Christaller, W. (1966). Central Places in Southern Germany. Englewood Cliffs, NJ: Prentice Hall.

Christensen, O. F., Roberts, G. O., & Sköld. M. (2006). Robust Markov chain Monte Carlo Methods for Spatial Generalized Linear Mixed Models. Journal of Computational and Graphical Statistics, 15(1), 1-17.

Clark, P. (2013). The effects of multicollinearity in multilevel models. Dissertation: Wright State University.

Clayton, D., & Kaldor, J. (1987). Empirical Bayes Estimates of Age-Standardized Relative Risks for Use in Disease Mapping. Biometrics, 43, 671–81.

Cobb, J. C. (1993). The Selling of the South: The Southern Crusade for Industrial Development, 1939-1990. Chicago, IL: University of Illinois Press.

Congress for the New Urbanism (2015). Charter of the New Urbanism. Congress for the New Urbanism. Retrieved from: https://www.cnu.org/who-we-are/charter-new-urbanism.

Crane, R. (2000). The influence of urban form on travel: An interpretive review. Journal of Planning Literature, 15(1), 3-23.

Crane, R., & Crepeau, R. (1998). Does neighborhood design influence travel? A behavioral analysis of travel diary and GIS data. Transportation Research Part D: Transport and Environment, 3(4), 225-238.

Credit, K. (2017). Transit-Oriented Economic Development: The impact of light rail on new business starts in the Phoenix, AZ region. Urban Studies. DOI: 10.1177/0042098017724119.

Curtis, C., Renne, J. L., & Bertolini, L. (2009). Transit Oriented Development: Making it Happen. Burlington, VT: Ashgate.

Damm, D., Lerman, S. R., Lerner-Lam, E., & Young, J. (1980). Response of Urban Real Estate Values in Anticipation of the Washington Metro. Journal of Transport Economics and Policy, 14(3), 315-36.

Darden, J. T., Rahbar, M., Jezierski, L., Li, M., & Velie, E. (2010). The measurement of neighborhood socioeconomic characteristics and black and white residential segregation in metropolitan Detroit: Implications for the study of social disparities in health. Annals of the Association of American Geographers, 100(1), 137-158.

158

Darden, J. T., & Kamel, S. M. (2000). Black Residential Segregation in the City and Suburbs of Detroit: Does Socioeconomic Status Matter? Journal of Urban Affairs, 22(1), 1-13.

Davis, J. E. (1984). Reallocating Equity: A Land Trust Model of Land Reform. In C. C. Geisler & F. J. Popper (Eds.), Land Reform, American Style. Totowa, NJ: Rowman & Littlefield.

Delgado, M., Porter, M., & Stern, S. (2010). Clusters and entrepreneurship. Journal of Economic Geography, 10(4), 495-518.

DeVol, R., & Wong, P. (1999). America’s High-Tech Economy: Growth, Development and Risks for Metropolitan Areas. Santa Monica, CA: Milken Institute.

De Bruijn, H., & Veeneman, W. (2009). Decision-making for light rail. Transportation Research Part A: Policy and Practice, 43(4), 349-359.

De Oliveira, V. (2013). Hierarchical Poisson Models for Spatial Count Data. Journal of Multivariate Analysis, 122, 393-408.

Diggle, P. J., Tawn, J. A., & Moyeed, R. A. (1998). Model-based geostatistics. Applied Statistics, 47(3), 299-350.

Diggle, P. J., & Ribeiro, P. J. (2007). Model-based Geostatistics. New York, NY: Springer.

Dittmar, H., & Ohland, G. (2004). The New Transit Town: Best Practices in Transit-Oriented Development. Washington, DC: Island Press.

Downtown Phoenix, Inc. (2015). Downtown Core. Downtown Phoenix, Inc. Retrieved from: http://dtphx.org/live/neighborhoods/downtown-core/.

The Economist. (2017). Snow belt to sun belt: Migration southward and westward is picking up again. The Economist. Retrieved from: http://www.economist.com/news/united- states/21721215-why-chicagoans-are-leaving-droves-migration-southward-and- westward-picking-up-again?frsc=dg%7Cc.

Edwards, R. (2008). Public Transit, Obesity, and Medical Costs: Assessing the Magnitudes. Preventive Medicine, 46(1), 14–21.

Eisinger, P. (1988). The Rise of the Entrepreneurial State: State and Local Economic Development Policy in the United States. Madison, WI: University of Wisconsin Press.

Elhorst, J. P. (2014). Spatial Econometrics: From Cross-Sectional Data to Spatial Panels. New York, NY: Springer.

Elhorst, J. P. (2017). Spatial Panel Data Analysis. In S. Shekhar, H. Xiong, & X. Zhou (Eds.), Encyclopedia of GIS (2nd Edition). New York, NY: Springer.

ESRI. (2014). Methodology Statement: 2014 ESRI US Business Locations and Business Summary Data. ESRI. Retrieved from: http://downloads.esri.com/esri_content_doc/dbl/us/J9940_Methodology_Statement_Esri_ US_Business_Locations_Business_Su.pdf.

159

ESRI Business Analyst. (2014). ESRI Business Locations and Business Summary [Data file]. ESRI. Retrieved from: http://www.esri.com/data/esri_data/business-overview/business.

Ewing, R. (1997). Is Los Angeles-style Sprawl desirable? Journal of the American Planning Association, 63(1), 107-126.

Ewing, R., & Cervero, R. (2010). Travel and the Built Environment: A Meta-Analysis. Journal of the American Planning Association, 76(3), 265-294.

Faraway, J. J. (2006). Extending the Linear Model with R: Generalized Linear, Mixed Effects and Nonparamteric Regression Models. New York, NY: Chapman & Hall.

Figueiredo, O., Guimaraes, P., & Woodward, D. (2002). Home-field advantage: location decisions of Portuguese entrepreneurs. Journal of Urban Economics, 52(2), 341-361.

Fink, M. (1993). Toward a sunbelt urban design manifesto. Journal of the American Planning Association, 59(3), 320-333.

Florida, R. (1995). Toward the Learning Region. Futures, 27(5), 527-536.

Florida, R. (2002). The Rise of the Creative Class: And How It’s Transforming Work, Leisure, Community, and Everyday Life. New York, NY: Basic Books.

Florida, R. (2004). Cities and the Creative Class. New York, NY: Routledge.

Frank, L. D., & Engelke, P. O. (2001). The built environment and human activity patterns: exploring the impacts of urban form on public health. Journal of Planning Literature, 16(2), 202-218.

Frank, L. D., Sallis, J. F., Conway, T. L., Chapman, J. E., Saelens, B. E., & Bachman, W. (2006). Many Pathways from land Use to Health: Associations between Neighborhood Walkability and Active Transportation, Body Mass Index, and Air Quality. Journal of the American Planning Association, 72(1), 75-87.

Freeman, C., & Perez, C. (1988). Structural crises of adjustment: Business cycles and investment behavior. In G. Dosi, C. Freeman, R. Nelson, G. Silverberg, & L. Soete (Eds.), Technical Change and Economic Theory. London, UK: Pinter Publishers.

Freemark, Y. (2014). Have U.S. Light Rail Systems Been Worth the Investment? Citylab. Retrieved from: http://www.citylab.com/commute/2014/04/have-us-light-rail-systems- been-worth-investment/8838/.

Frenken, K., Van Oort, F., & Verburg, T. (2007). Related Variety, Unrelated Variety and Regional Economic Growth. Regional Studies, 41(5), 685-697.

Galster, G., Temkin, K., Walker, C., & Sawyer, N. (2004). Measuring the Impacts of Community Development Initiatives: A New Application of the Adjusted Interrupted Time-Series Method. Evaluation Review, 28(6), 502-538.

160

Garrett, M., & Taylor, B. (1999). Reconsidering Social Equity in Public Transit. Berkeley Planning Journal, 13(1), 6-27.

Gehl, J. (2010). Cities for People. Washington, DC: Island Press.

Gilbert, B. A., McDougall, P. P., & Audretsch, D. (2008). Clusters, knowledge spillovers and new venture performance: An empirical examination. Journal of Business Venturing, 23(4), 405-422.

Giuliano, G. (2004). Land Use Impacts of Transportation Investments: Highway and Transit. In S. Hanson & G. Giuliano (Eds.), The Geography of Urban Transportation. New York, NY: Guilford Press.

Golub, A., Guhathakurta, S., & Sollapuram, B. (2012). Spatial and Temporal Capitalization Effects of Light Rail in Phoenix: From Conception, Planning, and Construction to Operation. Journal of Planning Education and Research, 32(4), 415-429.

Gordon, I. R., & McCann, P. (2000). Industrial Clusters: Complexes, Agglomeration and/or Social Networks? Urban Studies, 37(3), 513-532.

Graham, S., & Marvin, S. (2001). Splintering Urbanism: Networked Infrastructures, technological Mobilities and the Urban Condition. London, UK: Routledge.

Guerra, E., Cervero, R., & Tischler, D. (2011). The Half-Mile Circle: Does It Best Represent Transit Station Catchments? Working Paper: Institute of Transportation Studies, University of California, Berkeley. Retrieved from: http://www.its.berkeley.edu/sites/default/files/publications/UCB/2011/VWP/UCB-ITS- VWP-2011-5.pdf.

Hammann, E., Habisch, A., & Pechlaner, H. (2009). Values that Create Value: Socially Responsible Business Practices in SMEs – Empirical Evidence from German Companies. Business Ethics: A European Review, 18(1), 37-51.

Hanson, S., & Giuliano, G. (Eds.). (2004). The Geography of Urban Transportation. New York, NY: Guilford Press.

Harris, C., & Ullman, E. (1945). The Nature of Cities. Annals of the American Academy of Political Science, 242, 7-17.

Harvey, D. (1989). The Condition of Postmodernity: An Enquiry into the Origins of Cultural Change. Cambridge, MA: Blackwell.

Hass-Klau, C. (1993). Impact of pedestrianization and traffic calming on retailing: A review of the evidence from Germany and the UK. Transport Policy, 1(1), 21-31.

Hecker, D. E. (2005). High-technology employment: a NAICS-based update. Monthly Labor Review, 128(7), 57-72.

Heim, C. E. (2001). Leapfrogging, urban sprawl, and growth management: Phoenix, 1950–2000. American Journal of Economics and Sociology, 60(1), 245-283. 161

Hilbe, J. (2014). Modeling Count Data. Cambridge, MA: Cambridge University Press.

Hillier, B. (1996). Space is the machine: A configurational Theory of Architecture. Cambridge, MA: Cambridge University Press.

Holzer, H. J. (1991). The Spatial Mismatch Hypothesis: What Has the Evidence Shown? Urban Studies, 28(1), 105-122.

Hoover, E. M. (1937). Location Theory and the Shoe and Leather Industries. Cambridge, MA: Harvard University Press.

Hox, J. (2002). Multilevel Analysis Techniques and Applications. Mahwah, NJ: Lawrence Erlbaum Associates, Publishers.

Hoyt, H. (1939). The Structure and Growth of Residential Neighbourhoods in American Cities. Washington, DC: Federal Housing Administration.

Hyra, D. (2015). The Back-to-the-City Movement: Neighbourhood Redevelopment and Processes of Political and Cultural Displacement. Urban Studies, 52(10), 1753-1773.

Imagine Boston. (2016). Imagine Boston 2030. Imagine Boston. Retrieved from: http://imagine.boston.gov/.

Innovation District. (2015). Boston’s Innovation District: The Strategy. Boston Innovation District. Retrieved form: http://www.innovationdistrict.org/the-strategy/.

Jacobs, J. (1961). Death and Life of Great American Cities. New York, NY: Random House.

Jacobs, J. (1967). The Economy of Cities. New York, NY: Random House.

Jessop, B. (1992). Fordism and post-Fordism: Critique and reformulation. In M. Storper & A. Scott (Eds.), Pathways to Industrialization and Regional Development. London, UK: Routledge.

Jing, L., & De Oliveira, V. (2015). geoCount: An R Package for the Analysis of Geostatistical Count Data. Journal of Statistical Software, 63(11), 1-33.

Kafadar, K. (1996). Smoothing Geographical Data, Particularly Rates of Disease. Statistics in Medicine 15, 2539–60.

Kawabata, M., & Shen, Q. (2007). Commuting Inequality between Cars and Public Transit: The Case of the San Francisco Bay Area, 1990-2000. Urban Studies, 9, 1759-1780.

Kelejian, H. H., & Prucha, I. R. (1998). A Generalized Spatial Two-Stage Least Squares Procedure for Estimating a Spatial Autoregressive Model with Autoregressive Disturbances. Journal of Real Estate and Finances Economics, 17(1), 99-121.

Klepper, S. (2009). Spinoffs: A review and synthesis. European Management Review, 6(3), 159- 171.

Klepper, S., & Sleeper, S. (2005). Entry by Spinoffs. Management Science, 51(8), 1291-1306. 162

Knaap, G., Ding, J. C., & Hopkins, L. D. (2001). Do Plans Matter? The Effects of Light Rail Plans on Land Values in Station Areas. Journal of Planning Education and Research, 21(1), 32-39.

Knight, R. L., & Trygg, L. L. (1977). Evidence of Land Use Impacts of Rapid Transit Systems. Transportation, 6(3), 231-247.

Knox, P. L. (2005). Creating ordinary places: Slow cities in a fast world. Journal of Urban Design, 10(1), 1-11.

Kunstler, J. H. (1994). The Geography of Nowhere: The Rise and Decline of America’s Man- Made Landscape. New York, NY: Simon & Schuster.

Krizek, K. J. (2003a). Neighborhood services, trip purpose, and tour-based travel. Transportation, 30, 287-410.

Krizek, K. J. (2003b). Operationalizing Neighborhood Accessibility for Land Use-Travel Behavior Research and Regional Modeling. Journal of Planning Education and Research, 22, 270-287.

Landis, J., Guhathakurta, S., & Zhang, M. (1994). Capitalization of Transit Investments into Single-Family Home Prices: A Comparative Analysis of Five California Rail Transit Systems. The University of California Transportation Center, 246, 1-38. Retrieved from: http://www.uctc.net/papers/246.pdf.

Lawson, A. B., Banerjee, S., Haining, R. P., Ugarte, M. D. (2016). Handbook of Spatial Epidemiology. New York, NY: CRC Press.

Leigh, N. G., & Blakely, E. J. (2013). Planning Local Economic Development: Theory and Practice. Los Angeles, CA: Sage.

Leinberger, C. B., & Alfonzo, M. (2012). Walk This Way: The Economic Promise of Walkable Places in Metropolitan Washington, D.C. Brookings Institution. Retrieved from: http://www.brookings.edu/research/papers/2012/05/~/media/Research/Files/Papers/2012/ 5/25%20walkable%20places%20leinberger/25%20walkable%20places%20leinberger.pdf

LeSage, J., & Pace, R. K. (2009). Introduction to Spatial Econometrics. New York, NY: CRC Press.

Levy, A. (1999). Urban morphology and the problem of the modern urban fabric: Some questions for research. Urban Morphology, 3(2), 79-85.

Li, W., Joh, K., Lee, C., Kim, J., Park, H., & Woo, A. (2014). From Car-Dependent Neighborhoods to Walkers’ Paradise: Estimating Walkability Premiums in the Condominium Housing Market. Transportation Research Record, 2453, 162-170.

Li, W., Joh, K., Lee, C., Kim, J., Park, H., & Woo, A. (2015). Assessing Benefits of Neighborhood Walkability to Single-Family Property Values: A Spatial Hedonic Study in Austin, Texas. Journal of Planning Education and Research, 35(4), 471-488.

163

Lindstrom, M. (2008). Means of Transportation to Work and Overweight and Obesity: A Population-Based Study in Southern Sweden. Preventive Medicine, 46(1), 22–28.

Liu, H., Davidson, R. A., & Apanasovich, T. V. (2007). Statistical Forecasting of Electric Power Restoration Times in Hurricanes and Ice Storms. IEEE Transactions on Power Systems, 22(4), 2270-2279.

Lösch, A. (1967). The Economics of Location. New York, NY: John Wiley & Sons.

Lund, H. (2003). Testing the claims of new urbanism: Local access, pedestrian travel, and neighboring behaviors. Journal of the American Planning Association, 69(4), 414-429.

Mack, E., & Credit, K. (2016). The Intra-Metropolitan Geography of Entrepreneurship: a Spatial, Temporal, and Industrial Analysis (1989-2010). In E. Mack & H. Qian (Eds.), Geographies of Entrepreneurship. New York, NY: Taylor and Francis.

Mack, E., & Mayer, H. (2015). The Evolutionary Dynamics of Entrepreneurial Ecosystems. Urban Studies. DOI:10.1177/0042098015586547.

Mackinnon, D., & Cumbers, A. (2011). Introduction to Economic Geography: Globalization, Uneven Development and Place. New York, NY: Routledge.

Malecki, E. J. (1984). High Technology and Local Economic Development. Journal of the American Planning Association, 50(3), 262-269.

Malecki, E. J. (1991). Technology and Economic Development: The Dynamic of Local, Regional, and National Change. London, UK: Longman.

Malmberg, A., & Maskell, P. (2002). The Elusive Concept of Localization Economies: Towards a Knowledge-Based Theory of Spatial Clustering. Environment and Planning A, 34(3), 429-449.

Mao, J. (2015). Boston Designers Want to “Refresh” Urban Renewal. Next City. Retrieved from: https://nextcity.org/daily/entry/boston-urban-design-urban-renewal-city-planning-2015- meetings.

Marshall, A. (1890). Principles of Economics. London, UK: Macmillan.

Mayer, H., & Knox, P. J. (2006). Slow cities: sustainable places in a fast world. Journal of Urban Affairs, 28(4), 321-334.

McCann, P., & Sheppard, S. (2003). The Rise, Fall, and Rise Again of Industrial Location Theory. Regional Studies, 37(6-7), 649-663.

McCormack, G., Giles-Corti, B., & Bulsara, M. (2008) The Relationship between Destination Proximity, Destination Mix and Physical Activity Behaviors. Preventive Medicine, 46(1), 33-40.

164

Mejia-Dorantes, L., Paez, A., & Vassallo, J. M. (2012). Transportation infrastructure impacts on firm location: the effect of a new metro line in the suburbs of Madrid. Journal of Transport Geography, 22, 236-250.

Mohammad, S. I., Graham, D. J., Melo, P. C., & Anderson, R. J. (2013). A Meta-Analysis of the Impact of Rail Projects on Land and Property Values. Transportation Research Part A, 50, 158-170.

Molotch, H. (1976). The City as a Growth machine: Toward a Political Economy of Place. American Journal of Sociology, 82(2), 309-332.

Montgomery, C. A. (1994). Corporate diversification. Journal of Economic Perspectives, 8, 163- 178.

Montgomery, J. (1998). Making a city: Urbanity, vitality and urban design. Journal of Urban Design, 3(1), 93-116.

Moughtin, J. C., Cuesta, R., Sarris, C., & Signoretta, P. (2003). Urban design: Methods and techniques. Burlington, MA: Routledge.

National Historic Geographic Information System (NHGIS). (1990). U.S. Geographic Summary Data and Boundary Files – 1990 Metropolitan Statistical Area Population [Data file]. National Historic Geographic Information System. Retrieved from: https://www.nhgis.org/.

Nelson, M. (2005). Rethinking agglomeration economies and the role of the central city. Journal of Planning Education and Research, 24(3), 331-341.

Neumark, D., Zhang, J., & Wall, B. (2005). Business establishment dynamics and employment growth. Hudson Institute Research Paper.

Newman, & Kenworthy. (2013). Urban Sustainability in an Australian Context. Urban Sustainability, 227-254.

New York City Department of Transportation (NYCDOT). (2013). The Economic Benefits of Sustainable Streets. New York City Department of Transportation. Retrieved from: http://www.nyc.gov/html/dot/downloads/pdf/2013-dot-sustainable-streets-lowres.pdf.

O'hUallachain, B., & Reid, N. (1991). The location and growth of business and professional services in American metropolitan areas, 1976-1986. Annals of the Association of American Geographers, 81(2), 254-270.

Oldenburg, R. (1999). The Great Good Place: Cafes, Coffee Shops, Bookstores, Bars, Hair Salons, and Other Hangouts at the Heart of a Community. New York, NY: Marlowe & Company.

Orlitzky, M., Schmidt, F. L., & Rynes, S. L. (2003). Corporate Social and Financial Performance: A Meta-Analysis. Organization Studies, 24(3), 403-441.

165

Ozdil, T. R. (2006). Assessing the Economic Revitalization Impact of Urban Design Improvements: The Texas Main Street Program. Dissertation – Texas A&M University. Retrieved from: http://oaktrust.library.tamu.edu/bitstream/handle/1969.1/5885/etd-tamu- 2006A-URSC-Ozdil.pdf?sequence=1&isAllowed=y.

Pagliara, F., & Papa, E. (2011). Urban Rail Systems Investments: An Analysis of the Impacts on Property Values and Residents’ Location. Journal of Transport Geography, 19, 200-211.

Pela, R. (2015). Roosevelt Row: The Fight to Keep Downtown Phoenix “Authentic” Is a Fight to Let It Become Something It’s Never Been. Phoenix New Times. Retrieved from: http://www.phoenixnewtimes.com/arts/roosevelt-row-the-fight-to-keep-downtown- phoenix-authentic-is-a-fight-to-let-it-become-something-its-never-been-6569021.

Phoenix. (2017a). City of Phoenix Zoning Map. City of Phoenix. Retrieved from: https://phoenix.maps.arcgis.com/apps/webappviewer/index.html?id=13428321a9f84e95a 634be1beab5fe96.

Phoenix. (2017b). City of Phoenix Zoning Regulations. City of Phoenix. Retrieved from: (http://www.codepublishing.com/AZ/Phoenix/html/PhoenixZ06/PhoenixZ0662.html.

Tempe. (2017). City of Tempe Zoning Regulations. City of Tempe. Retrieved from: https://www.tempe.gov/home/showdocument?id=1913.

Piore, M., & Sabel, C. (1984). The Second Industrial Divide: Possibilities for Prosperity. New York, NY: Basic Books.

Pivo, G., & Fisher, J. D. (2011). The Walkability Premium in Commercial Real Estate Investments. Real Estate Economics, 39(2), 185-219.

Pollack, S., Bluestone, B., & Billingham, C. (2010). Maintaining Diversity in America’s Transit- Rich Neighborhoods: Tools for Equitable Neighborhood Change. Boston, MA: Dukakis Center for Urban and Regional Policy at Northeastern University.

Porter, M. (2000). Location, Competition, and Economic Development: Local Clusters in a Global Economy. Economic Development Quarterly, 14(1), 15-34.

Quinn, B. (2006). Transit-Oriented Development: Lessons from California. Built Environment, 32(3), 311-322.

Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear models: Applications and data analysis methods (2nd ed.). Newbury Park, CA: Sage Publications.

Relph, E. C. (1976). Place and Placelessness. London, UK: Pion Press.

Renski, H. (2008). New Firm Entry, Survival, and Growth in the United States: A Comparison of Urban, Suburban, and Rural Area. Journal of the American Planning Association, 75(1), 60-77.

166

Resource Systems Group (RSG). (2014). Who’s on Board 2014: Mobility Attitudes Survey. Prepared for TransitCenter. Retrieved from: http://transitcenter.org/wp- content/uploads/2014/08/WhosOnBoard2014-ForWeb.pdf.

Richardson, H., & Bae, C. (2016). Urban Sprawl in Western Europe and the United States. New York, NY: Ashgate.

Richman, J. (2015). Here’s Why Tech Startups are Flocking to Austin. Tech.co. Retrieved from: https://tech.co/make-way-silicon-valley-heres-tech-startups-flocking-austin-2015-04.

Riggs, W. (2014). Steps toward validity in active living research: Research design that limits accusations of physical determinism. Health & Place, 26, 7-13.

Robertson, K. (1999). Can Small-City Downtowns Remain Viable? A National Study of Development Issues and Strategies. Journal of the American Planning Association, 65(3), 270-283.

Robinson, K. L., Dassie, W., & Christy, R. D. (2004). Entrepreneurship and small business development as a rural development strategy. Southern Rural Sociology, 20(2), 1–23.

Rodriguez, G. (2007). Chapter 4: Poisson Models for Count Data. Lecture Notes on Generalized Linear Models. Retrieved from: http://data.princeton.edu/wws509/notes/c4.pdf.

Rosenthal, S.S., & Strange, W. C. (2001). The Determinants of Agglomeration. Journal of Urban Economics, 50(2), 191-229.

Rutten, R., & Boekema, F. (2007). The Learning Region: A Conceptual Anatomy. In R. Rutten & F. Boekema (Eds.), The Learning Region: Foundations, State of the Art, Future.

Sabel, C. (1994). Flexible specialization and the re-emergence of regional economies. A. Amin (Ed.), Post-Fordism: A Reader. Cambridge, MA: Blackwell.

Saelens, B. E., Sallis, J. F., Black, J. B., & Chen, D. (2003). Neighborhood-based differences in physical activity: an environment scale evaluation. American Journal of Public Health, 93(9), 1552-1558.

Saviotti, P.P., & Pyka, A. (2004). Economic development by the creation of new sectors. Journal of Evolutionary Economics, 14(1), 1-35.

Saxenian, A. (1994). Regional Advantage: Culture and Competition in Silicon Valley and Route 128. Cambridge, MA: Harvard University Press.

Schuetz, J. (2014). Do Rail Transit Stations Encourage Neighborhood Retail Activity? Urban Studies. DOI: 10.1177/0042098014549128.

Schumpeter, J. (1934). The Theory of Economic Development. Cambridge, MA: Harvard University Press.

167

Smart Growth America. (2014). The Best Complete Streets Policies of 2014. Smart Growth America. Retrieved from: http://www.smartgrowthamerica.org/complete-streets-2014- analysis.

Song, Y., & Knaap, G. J. (2003). New urbanism and housing values: a disaggregate assessment. Journal of Urban Economics, 54(2), 218-238.

Song, Y., & Knaap, G. J. (2004). Measuring the effects of mixed land uses on housing values. Regional Science and Urban Economics, 34(6), 663-680.

Song, Y., Lee, K., Anderson, W. P., & Lakshmanan, T. R. (2012). Industrial agglomeration and transport accessibility in metropolitan Seoul. Journal of Geographical Systems, 14(3), 299–318.

Sparks, C. S. (2015). DEM 7263 Fall 2015 - Spatially Autoregressive Models 2. Rpubs. Retrieved from: https://rpubs.com/corey_sparks/109650.

Spencer, G. M. (2015). Knowledge neighbourhoods: Urban form and evolutionary economic geography. Regional Studies, 49(5), 883-895.

Spigel, B. (2015). The Relational Organization of Entrepreneurial Ecosystems. Entrepreneurship Theory and Practice. DOI: 10.1111/etap.12167.

Squires, G. (2002). Urban Sprawl: Causes, Consequences, and Policy Responses. Washington, DC: The Urban Institute Press.

Stam, E. (2007). Why Butterflies Don’t Leave: Locational Behavior of Entrepreneurial Firms. Economic Geography, 83(1), 27-50.

Stam, E. (2015). Entrepreneurial Ecosystems and Regional Policy: A Sympathetic Critique. European Planning Studies, 23(9), 175-1769.

Storper, M., & Scott, A. J. (2009). Rethinking Human Capital, Creativity and Urban Growth. Journal of Economic Geography, 9(2), 147-167.

Storper, M., & Venables, A. J. (2004). Buzz: Face-to-Face Contact and the Urban Economy. Journal of Economic Geography, 4(4), 351-370.

Sturtevant, L. A., & Jung, Y. J. (2011). Are We Moving Back to the City? Examining Residential Mobility in the Washington, DC Metropolitan Area. Growth and Change, 42(1), 48-71.

Talen, E. (1999). Sense of community and neighbourhood form: An assessment of the social doctrine of new urbanism. Urban Studies, 36(8), 1361-1379.

Teece, D. J. (2007). Explicating dynamic capabilities: the nature and microfoundations of (sustainable) enterprise performance. Strategic Management Journal, 28(13), 1319-1350.

Tiefelsdorf, M., & Griffith, D. A. (2007). Semiparametric filtering of spatial autocorrelation: the eigenvector approach. Environment and Planning A, 39(5), 1193-1221.

168

Tranel, M., & Handlin, L. B. (2006). Metromorphosis: Documenting Change. Journal of Urban Affairs, 28(2), 151-167.

UCLA. (2007). R Data Analysis Examples: Poisson Regression. UCLA: Statistical Consulting Group. Retrieved from: http://www.ats.ucla.edu/stat/r/dae/poissonreg.htm.

United States Census Bureau. (2015). Census TIGER/Line Road Shapefile – Maricopa County [Data file]. United States Census Bureau. Retrieved from: https://www.census.gov/geo/maps-data/data/tiger-line.html.

United States Census Bureau. (2016). Business Dynamics Statistics – Definitions. United States Census Bureau. Retrieved from: https://www.census.gov/ces/dataproducts/bds/definitions.html.

United States Department of Transportation (USDOT). (2008). Realizing the Potential: One Year Later, Housing Opportunities near Transit in a Changing Market. Washington, DC: US Department of Transportation.

Unwin, P. T. H. (1992). The Place of Geography. Minneapolis, MN: Longman Scientific and Technical.

Urban Land Institute (ULI). (2015). America in 2015: A ULI Survey of Views on Housing, Transportation, and Community. Urban Land Institute. Retrieved from: http://uli.org/wp- content/uploads/ULI-Documents/America-in-2015.pdf.

Vale, D. S., Pereira, M., & Saraiva, M. (2016). Active accessibility: A review of operational measures of walking and cycling accessibility. The Journal of Transport and Land Use, 9(1), 1-27.

Venturi, R., Brown, D. S., & Izenour, S. (1972). Learning from Las Vegas: The Forgotten Symbolism of Architectural Form. Cambridge, MA: MIT Press.

Vessali, K. V. (1996). Land Use Impacts of Rapid Transit: A Review of the Empirical Literature. Berkeley Planning Journal, 11(1), 72-105.

Vojnovic, I., Lee, J., Kotval-K, Z., Podagrosi, A., Varnakovida, P., Ledoux, T., & Messina, J. (2013). The burdens of place: A socio-economic and ethnic/racial exploration into urban form, accessibility and travel behavior in the Lansing Capital Region, Michigan. Journal of Urban Design, 18(1), 1-35.

Walljasper, J. (2007). The Great neighborhood Book: A Do-it-Yourself Guide to Placemaking. Gabriola Island, BC: New Society Publishers.

Walls & Associates. (2012). NETS Database vs. BLS Establishment & Employment Estimates. Walls & Associates: Insights into Business Dynamics. Obtained from NETS datafile.

Walls & Associates. (2013). National Establishment Time-Series (NETS) Database: 2012 Database Description. Walls & Associates. Retrieved from: http://exceptionalgrowth.org/downloads/NETSDatabaseDescription2013.pdf.

169

Wang, Y., Kockelman, K., & Wang, X. (2013). Understanding Spatial Filtering for Analysis of Land Use Data Sets. Journal of Transport Geography, 31, 123-131.

Ward, M. D., & Gleditsch, K. S. (2008). Spatial Regression Models. Thousand Oaks, CA: Sage Publications.

Weber, A. (1929). Theory of the Location of Industries. Chicago, IL: University of Chicago Press.

Weinberger, R. R. (2001). Light Rail Proximity: Benefit or Detriment in the Case of Santa Clara County, California? Transportation Research Record, 1747, 104-11.

Weinstein, B., & Clower, T. L. (2003). DART Light Rail’s Effect on Taxable Property Valuations and Transit-Oriented Development. University of North Texas. Retrieved from: http://www.valleymetro.org/images/uploads/general_publications/2003_DART_Study.pd f.

Weisbrod, G., Duncan, C., & Moses, S. (2014). The Evolving Connection of Transit, Agglomeration and Growth of High Tech Business Clusters. Transportation Research Record, 2452.

Weissmann, J. (2012). Why Don’t Young Americans Buy Cars? The Atlantic. Retrieved from: http://www.theatlantic.com/business/archive/2012/03/why-dont-young-americans-buy- cars/255001/.

Wennekers, S., & Thurik, R. (1999). Linking Entrepreneurship and Economic Growth. Small Business Economics, 13(1), 27-55.

White, S. B., & Kotval, Z (Eds). (2012). Financing Economic Development in the 21st Century. London, UK: Routledge.

Wiles, R. (2014). Billionaire Warren Buffet buys group of Arizona auto dealers. AZcentral.com. Retrieved from: https://www.azcentral.com/story/money/business/2014/10/02/billionaire- warren-buffett-buys-group-of-arizona-auto-dealers/16620611/.

Woltman, H., Feldstain, A., MacKay, J. C., & Rocchi, M. (2012). An introduction to hierarchical linear modeling. Tutorials in Quantitative Method for Psychology, 8(1), 52-69.

Ye, L., Mandpe, S., & Meyer, P. B. (2005). What is “smart growth?”—Really? Journal of Planning Literature, 5(3), 301-315.

Yu, H., Jiang, S., & Land, K. C. (2015). Multicollinearity in hierarchical linear models. Social Science Research, 53, 118-136.

Zhang, M. (2004). The role of land use in travel mode choice: Evidence from Boston and Hong Kong. Journal of the American Planning Association, 70(3), 344-361.

170

Zhao, F., Chow, L. F., Li, M. T., Gan, A., & Ubaka, I. (2003). Forecasting transit walk accessibility: Regression model alternative to buffer. Transportation Research Record, 1835, 34-41.Zhao, F., Chow, L. F., Li, M. T., Gan, A., & Ubaka, I. (2003). Forecasting transit walk accessibility: Regression model alternative to buffer. Transportation Research Record, 1835, 34-41.

171