8
The Science of Scenes
With Christopher M. Graziul
Many scientific endeavors rely on highly specialized equipment to probe indirect evidence supporting or disconfirming a theory, and so it goes with Scenescapes. The styles of life we seek to uncover are amorphous, ephemeral, and defy traditional measurement. They are embedded in a built environment whose components are saturated with meaning.
That the physical structures of amenities are saturated with meaning poses classic methodological challenges—quantitative, qualitative, philosophical, or otherwise—for their interpreters. Thus Jürgen Habermas points out, during his account of how coffeehouses helped create a public civil sphere in eighteenth-century England, that cafes were not culturally significant because they served coffee. Yet they still served coffee. Conversely, not every gentleman of the time who frequented a coffeehouse did so to participate in a new form of critical discourse.
Similarly, scenes require the action and energy of human beings to create and support the meanings they offer. This implies that scenes can never be measured completely through any method, let alone by counting the amenities in a community. The “signal” that would indicate the existence of a particular scene at least partially resides in community members (their traits, tastes, and activities) and involves a wider audience attaching certain meanings to places within the community. We therefore primarily deal with indirect evidence of an abstract and emergent cultural phenomenon. In other words, our main data, local amenities, are not constitutive of scenes but, through the science of scenes, become useful indicators of the qualities that constitute scenes.
One helpful way to understand the transformative process that allows us to go from amenities to scenes is to consider the example of experimental particle physics.1 The goal of a science of scenes is to detect the constituent cultural components of a location by observing amenities that have manifold meanings. Particle accelerators are a response to similar challenges faced by physicists. Sometimes the existence of certain unobservable constituent particles can only be inferred by observing how larger, more complex particles interact with one another. Thus many efforts to discover new and exotic particles involve laboriously confirming that a certain soup of particles could only have come about because of the predicted properties of a seemingly invisible ingredient. In the same way, scenes analysis makes inferences about things we cannot see directly (glamour, self-expression, formality, charisma) by observing things that we can see—in this case, amenities.
Beyond facing similar epistemological issues, the science of scenes also shares a surprising number of practical problems with the science of particle physics. Both seek to confirm theoretical predictions by constructing highly specialized tools for collecting indirect evidence. Both must discriminate between “noise” and “signal,” between irrelevant or meaningless data and meaningful data within this indirect evidence. And to do so, again like particle physicists, we start with a very concrete theory of “what is out there” and proceed as if the world truly operates that way, searching for the best indirect evidence we can find that enables us to confirm or disconfirm our theory.
Accordingly, a major methodological challenge for scenes analysis is to develop ways to turn our main indirect indicators—amenities—into measures of scenes. The physics analogy continues: Just as researchers employ multiple detectors to jointly observe the properties of a particle (e.g., trajectory and electrical charge), we create a set of measures for jointly observing the properties of a scene (e.g., transgressiveness and traditionalism). Just as researchers repeatedly perform the same collisions to statistically test the likelihood that the soup of particles they observe implies the existence of unobservable particles, we employ data on tens of thousands of locations to statistically test the likelihood that the amenities we observe imply the existence of particular scenes.
And just as there is a Standard Model of particle physics, we have a “standard model” of scenes. This model describes the constituent units of scenes, which we call dimensions, and allows us to theorize how scenes ought to relate to more traditional social science variables, like income or education. A science of scenes thus needs to demonstrate how the properties of a location’s scene, indicated by its amenities, relate in theoretically consistent ways to other more tangible features of the community.
However, social scientists face methodological challenges that physical scientists do not. Our hypothesized relationships cannot be reduced to a series of syllogisms, and the novelty of our approach suggests that a lack of a relationship may reflect a poorly constructed measure instead of meaningful empirical research. This chapter is accordingly about providing insight into the thinking and decisions that went on—and are still going on—“behind the scenes” of Scenescapes, which is by no means the final word on the topic. This is a science of scenes as an ongoing dialogue between theory and observation. We are not so arrogant as to believe that the 15 dimensions we pose are the only ones, or even the best; they are provisional and pragmatic interpretations from past related work. Our goal here is to point out to readers, especially those with doubts about analytic decisions in previous chapters, the utility of the underlying techniques such that others can adjust our concepts and pursue their own ways to build meaning using the types of data we utilize.
Goals of the chapter. How do we get from (a) information about amenities like public parks, bookstores, and tattoo parlors to (b) quantitatively analyzing how, for instance, self-expressive scenes relate to economic growth? The line from (a) to (b) is not direct or immediately obvious, and accordingly, Scenescapes often opts to provide relatively intuitive and accessible arguments about this process rather than extensive details that may appear esoteric to most readers. This chapter provides technical details that more quantitatively oriented readers (like professional social scientists) would naturally want to know.
One major aim of the chapter is thus to describe and justify our methods for quantifying the presence of scenes and how they relate to urban development. We start with an overview of the national database of local amenities we compiled, one of the most extensive of its kind to date. Next we describe how we transform these data from measures of what is there (amenities data) to measures of what that means (the performance score measures of scene dimensions like self-expression or neighborliness).
We then discuss our standard methodological approach for jointly analyzing how scenes and other more traditional data relate to urban development. To do this, we review multiple ways to model spatial context and apply these techniques to an illustrative result from chapter 5 (that between 1990 and 2000, 25- to 34-year-olds increased in transgressive but not self-expressive scenes).
This discussion of modeling context is important in a number of ways. It helps to validate the robustness of the standard methods in Scenescapes by showing that more sophisticated techniques produce results that are generally consistent with those generated using a simpler approach. Perhaps more importantly, it joins scenes analysis with more recent research involving spatial and contextual analysis. These research efforts have produced new tools for explicitly analyzing the role of context in myriad social processes.2 Indeed, if in the past many social scientists turned to qualitative methods (in part) because of dissatisfaction with the intellectual simplifications inherent in basic quantitative techniques—descriptive maps, frequency distributions, cross tabulations, or linear regression analysis—newer methods permit crucial ideas from qualitative work (like context) to be more directly modeled quantitatively. These novel approaches allow us to explicitly analyze how the situations in which life unfolds affect its meaning and course, and we are able to do so across far larger geographic areas than ever before.
Even though methods have advanced rapidly in recent years, less progress has been made to join them with concepts designed to bring some of the subtlety of ethnography and qualitative research into large-scale statistical analyses. Realizing this potential is one of the major methodological goals of the scenes approach, as we have stressed throughout Scenescapes. In chapter 3, we laid out some of the intuitive logic for how proximity alters performance, that is, how when some scenes are near one another, this spatial configuration may shift their meanings and impacts. And later chapters explicitly analyzed contextual effects, as in the chapter 4 discussion of the economic impacts of Bohemian zip codes in communitarian or urbane counties, and the chapter 6 discussion of how affluent counties are less likely to be Republican when they have a transgressive scene. Here we extend these ideas about context and incorporate more advanced techniques, showing how scenes concepts and data can create more substantive meaning when combined with the modeling of spatial effects.
Another aim of the chapter is to review a series of approaches to the analysis of scenes that we explored but ultimately did not report in the main analytical chapters. We begin with a discussion of causality and the implications of employing scenes as an outcome rather than a predictor. We end with a review of ongoing international scenes research. The intent here is to show how we grappled with various analytic questions and to suggest some of the many possible directions for future research employing the scenes paradigm.
From Theory to Data
Many of our scenes concepts, data, and methods are new, but they have important precedents. One important inspiration is recent work seeking to develop a “cultural sociology” in general and an “urban culturalist perspective” in particular.3 This research has tacitly reenergized an idea shared by classic figures such as Weber, Simmel, Durkheim, and James: the idea to bring seemingly transcendental values down to earth and to see their tangible role in social life.
Thus rather than treating cultural meaning as the necessary condition of social integration, or strategic tool kits,4 many look to the social currency of meanings already embedded in concrete places and practices5—in trees, rivers, parks, street corners, neighborhoods, record shops, sidewalks, baseball parks, blues clubs, bumper stickers, and so on.6 Others have developed proposals for how to measure “meaning structures” and investigate their correlations with, for instance, leadership and occupational patterns.7
A related inspiration includes writings more explicitly devoted to studying phenomena under the heading of “scene.” While the concept has been loosely used by art and music critics for decades, academic researchers have focused on “scenes” as a means to trace the role of national theaters and lifestyle communities in modernization processes, as niches for urban belonging in the metropolis that do not require nostalgia for the premodern village (Straw 2001), and as links to “youth” with respect to a specific phase of the life course.8 Bennett and Peterson (2004) compile descriptions of various music scenes, dividing them into local, global, and virtual. Lena and Peterson (2008) treat “scenes” as one phase in the life course of music genres, as they grow from avant-garde to scene-based to industrial to nostalgic communities. And Haunss and Leach (2007) and Lichterman and Eliasoph (2014) highlight how scenes sometimes animate social movements.
It would be a mistake to call this work a “literature” in the sense of the product of an interconnected research community cumulatively pursuing a scientific program. Most ignore the others and they have largely emerged independently. Still, our review of these works (among others) helped formulate our conception of scenes as the meanings expressed by the people and practices in a place (Haunss and Leach’s 2007 review produced a similar result). At the same time, our multidimensional theory of scenes incorporates some of the main themes stressed by others, such as exhibitionism and transgression (Blum 2003), local authenticity (Zukin 2010; Grazian 2003), and more (as described in chapter 2). But we sought to develop a more comprehensive and integrated analytical framework.
Beyond this theoretical contribution, a primary goal of Scenescapes is to meld cultural theories with data to make qualitative meanings embedded in places not just observable but, in an albeit limited way, quantifiable. Much past effort has been devoted to develop methods for examining and measuring other core aspects of cities and communities—how big they are, how much they produce, how many they employ. Very little work has sought to create measures of their expressive organization, their scenescape.
Finally, both conceptually and methodologically, from the outset we sought to develop an analytical perspective and operational measures that could travel globally. Taking seriously how global we are is a relatively new challenge in general, and asking how to build this globalism into our core concepts and thinking is even newer. This wide-ranging scope was facilitated by working with the ideas, data, and people who developed the Fiscal Austerity and Urban Innovation Project over the previous quarter century into one of the most comprehensively international projects in the social sciences.9
While recognizing these foundations, we still had to chart much new conceptual and methodological territory. The conceptual syntheses in the above chapters emerge from Daniel Silver’s training at the University of Chicago’s Committee on Social Thought, where dialogue with Aristotle, Hegel, and Baudelaire is a daily staple. But he found new specific ways to join these with street conversations and regression coefficients. Terry Clark’s past work is more urban and classically sociological, although he moved toward scenes ideas via art school, the Frankfurt School, five years in Paris, and biking Chicago neighborhoods. We joined in an effort to build a new way to incorporate cultural ideas into sociological discussions, which often downplay their significance.
To do so, we (1) compiled a nationwide database with hundreds of types of amenities, supplemented by citizen surveys and other information about neighborhoods. It is largely the product of some 15 University of Chicago students working hard over ten years with, alas, continuous turnover as graduations and other diversions intervened. We then (2) explored many conceptual approaches in the course of our research, continuously building our own analytical framework over nearly a decade. Next we (3) “translated” much of our data into measures of the 15 dimensions of theatricality, authenticity, and legitimacy we present here. After much exploration and experimentation, we found performance scores to be the most theoretically and methodologically compelling measures of scenes. Clemente Navarro was our constant companion in extending these tools, and Lawrence Rothfield was a key contributor in developing our initial conceptualization of scenes.
Throughout we have (4) been in dialogue and debate with leading theorists, especially in Los Angeles, Paris, and New York, about the nature of the “New Chicago School” and how our theories can address the most challenging urban questions nationally and globally.10 We have (5) continually worked with many students and colleagues from around the world, consulting with persons from Beijing to Seoul to Paris to Seville and beyond who were simultaneously developing similar projects elsewhere. For instance, a Seoul Development Institute report and Can Tocqueville Karaoke? centrally address how Western urban theories need modification to understand Asia, especially Tocqueville’s ideas about democratic consensus and recent claims about the role of Bohemia in spurring urban growth.11 Yet we also (6) worked with others (especially some 75 students writing papers on these and related themes) on specific neighborhoods studies, often in Chicago, to explore specific themes like glamour, Bohemia, blue-collar Bobo, and the like, which have frequented the pages of Scenescapes. This cross-fertilization has yielded a large and growing international body of research applying and extending the scenes perspective. Some project websites include scenescapes.weebly.com and scenes.en-linea.eu/en/.
These multiple contacts have (7) made us both more cosmopolitan and more conscious of the contexts, and the limits, of our analytical enterprise. Most theories propose 2 or 3 main types; we use 15! These emerged from challenging ourselves to build a theory general enough to cover the world while incorporating enough subtlety to dialogue with the street-smart participants elaborating on the great traditions of Chicago urban research. Some 15 of us (mainly University of Chicago students and visiting scholars) would generally meet for two hours weekly, joining these many overlapping activities and talking through each key research decision. Scenescapes is emphatically a team effort, and while most data in this volume are for the United States and Canada, they are interpreted in a manner informed by the gaze of many other scenes. When we failed, others corrected us, repeatedly. We started over many times and will likely do so again.
Why Have a National Database of Local Indicators?
This section reviews the main considerations we took into account when searching for suitable scenes data, and then provides a description of the data we ultimately used. The sharpest difference from past work is our inclusion of hundreds of indicators to capture the wide range of cultural dimensions in our theory of scenes. Our decision to explore this seldom-used wealth of data is based on our goal of capturing cultural context as a holistic experience incorporating the built environment.
Beyond atomism. Economists (many at Chicago) have pioneered in theorizing amenities, and we build on their legacy.12 But they largely did so by focusing on just one or a few amenities, which amounts to assuming a vacuum around any single amenity. That is, economists generally reason atomistically and ignore context via the ceteris paribus assumption. Thus, a perfectly reasonable way to study rent prices in urban settings, based on this assumption, is to add January temperature to a standard set of variables known to predict rent in order to estimate the dollar value of January temperature. This is termed hedonic price analysis. One could do the same for restaurants or other amenities, and many have (e.g., Albouy, Leibovici, and Warman 2013).
However, this atomistic approach neglects to consider the fact that patrons of a restaurant next to a tattoo parlor will likely differ substantially from patrons of another restaurant next to an opera, with quite distinct implications for residential and economic growth.13 Thus, we try to capture what makes locations different from one another by including many lifestyle indicators that jointly create the context, giving meaning to the many specific amenities in that scene. The scene in this respect is an emergent property based on its combinatorial logic of adjacency to multiple individual amenities. This holism contrasts with the atomism of most past amenities work.
What Would an Ideal Scenes Data Set Be Like?
Our methodological goal in building the scenes database is to join the subtlety of meaning found in strong ethnographic studies with the wide-scale comparative analysis possible in quantitative research. To achieve this end, we have had to juggle several important yet competing priorities when searching for data sets. Clearly no source is perfect in all these respects, but we list them as our guides, our selection criteria, albeit imperfectly implemented.
An ideal scenes data set would possess several qualities. First, the data should be geographically comprehensive. The data should be nationwide and the product of consistent methods of collection across the country. When we move to international work, we seek to retain such comparability, and we rely on experts (especially national and international statistical agencies) whose job is to assemble and adjust raw local items to make them more comparable. Such data provide maximal comparability of regions with widely varying cultural and historical trajectories, preventing, ideally, these unique features from biasing results. Further, since we make no a priori distinction between urban and rural scenes, we eschew categorical distinctions between these two settings and prefer data that represent a continuous sampling of geographies from the largest and most densely populated to the smallest and least densely populated. Barring this comprehensiveness, a data set should, at the very least, provide information for a large percentage of the population by area—not just the largest metropolitan areas, but as much of the country as possible.
Second, because scenes are largely submetropolitan phenomena, the data should possess a high degree of spatial resolution. Ideally, these data would attach street addresses (i.e., points with latitude and longitude coordinates) to each data item and include information on who frequents these places. If these details are not available, our preferred geographic unit of analysis is the zip code or its equivalent, permitting analysis of some neighborhood-level effects by inferring that local patrons exist for the amenities in these small areas, though even more spatially refined data is preferable and will hopefully become available in the future.14 Less spatially refined data (e.g., cities, counties, metro areas, and states) provide useful but more limited insight into the scenescape by masking the internal variation an average resident would experience while moving through these larger areas (cities, counties, etc.).
Third, the data should differentiate amenities as much as possible. Knowing that there is 1 business or 50 businesses in a given zip code says little about what that business does and what scenes are evinced by its presence. Knowing that a business is a retail establishment adds a little; knowing that it is a restaurant says more; knowing that it is a West Indian restaurant says much more still. Similarly, it is important to know what else is nearby. Is our restaurant in a West Indian enclave? Or is it in a cosmopolitan scene, next door to sushi, Thai food, and a French bistro? Do residents there go out frequently or are they homebodies? Do they enjoy trying new things or fear change? (Self-reports on these types of activities and attitudes are in the DDB survey we also use.) Finally, this differentiation should also include the cultural content. Data on five different kinds of gas stations are far less informative than data on five kinds of ethnic restaurants.
Fourth, an amenity should ideally be spatially invariant in its potential availability and typical practices. Local users should be able to reveal their preferences by patronizing a Thai restaurant, but if citizens prefer catfish restaurants, the local market should not prohibit a catfish restaurant from emerging. Thus in some cross-national work where market entry is more constrained (e.g., beef in India), the regulations and other factors that affect market entry by new firms should be incorporated into the analysis. Further, independent of how its meaning is attenuated by (or contributes to) its surrounding environment to constitute its scene, an amenity’s label, such as “Thai restaurants” in a business directory, should consistently refer to a similar set of practices, such as making and eating Thai food rather than pizza. That is, category labels should be functionally equivalent in their referents. Standardized items such as Starbucks and McDonalds meet this criterion relatively straightforwardly, though even here the issue is complex; less standardized items such as cultural centers (which offer diverse activities) and restaurants (which differ by cuisine and price) are more difficult.15
Finally, the data should permit analysis of trends over time. This is important for multiple, sometimes subtle reasons. First, causal direction is a prime concern when discussing the relationship between scenes and major indicators of economic growth. Second, the stability or impermanence of scenes over time is an important phenomenon worth investigating in its own right. Similarly, a data set of any substantial size tends to have “noise” within it, and being able to average over multiple years can reduce this issue when, for instance, a small town or neighborhood has many coffee shops that open and close within a few years. Finally, time series analysis of scenes data can provide important insight into how economic recessions, neighborhood population composition, gentrification, and so forth, may influence the cultural life of regions, as well as whether certain types of scenes are more or less resilient to these changes.
Scenes Data Sources
A data source that perfectly conforms to these ideals is utopian. So we did what most social researchers do: we used the best of what was available to us, looking for errors and irregularities and ways to assess how much these discrepancies bias our results.
Our scenes database has two main sources: (1) the US Census Bureau’s Zip Code Business Patterns (BIZZIP) from 2001 and (2) an original data set using online business directories often known as yellow pages (YP). We supplemented these two sources with many others, such as the DDB Needham Lifestyle Survey, the Urban Institute’s Unified Database of Arts Organizations, arts organizations’ professional directories, and more. After exploring many possibilities, we decided that BIZZIP and YP best fit our criteria for capturing the holistic scenescape in its richness and subtlety. BIZZIP and YP both have distinct benefits and drawbacks, though each provides critical indirect evidence of scenes. Their joint use, along with supplements, produces a more illuminating picture than any one source alone.
Consider briefly the strengths and weaknesses of each, starting with BIZZIP. BIZZIP utilizes a system of classification known as the North American Industry Classification System (NAICS). This data product provides extensive information on commercial and noncommercial entities—for example, theater companies and musical groups, promoters of performing arts, museums, independent artists, and even human rights organizations, environmental groups, and religious organizations. BIZZIP is also not just a sample; it enumerates the entire population. Its universe is all US zip codes, and its data are drawn from multiple federally collected data sources—primarily the continuously updated Business Register and the periodically undertaken Economic Census.16 As an annual survey of businesses, BIZZIP permits analysis of amenity mix changes over time.
Only a handful of analysts have made use of the BIZZIP data; most economists use a geographically aggregated version, County Business Patterns.17 Use of BIZZIP in social scientific research has been rare since it has not been publicized, and because of the difficulties posed by utilizing hundreds of indicators in a theoretically consistent and meaningful way, not to mention the technical challenges in compiling the raw data into an analyzable format. Scenescapes is thus also an attempt to bring this rich source of data to the attention of more potential users. Many countries report similar data, some in near-identical format.
Yet BIZZIP clearly has its limits. Most information concerns more formal organizations (e.g., businesses and nonprofits, as opposed to public parks) and is classified more from a production and bureaucratic perspective than from a consumption perspective. And while it is good to know how many restaurants are in an area, for instance, BIZZIP does not report what kinds of restaurants are present; BIZZIP simply reports the number of restaurants in each zip code. Do they serve Italian, Korean, or Lebanese food? Is the restaurant local or part of a national or regional chain? To gain access to these kinds of cultural differentiations, we employed a commercial program to systematically download hundreds of categories from online yellow pages sources to create our second scenes data source, YP. We used similar procedures with our scenes collaborators in other countries.
After reviewing several possible providers, we opted for superpages.com, because its geographic coverage was the most extensive and its categorization scheme the most consistent and fine grained, especially for items with rich cultural content.18 YP thus provides a much higher degree of specificity than BIZZIP in terms of categories (Chinese restaurants are differentiated from Mongolian restaurants, Baptist churches from Catholic churches, etc.). YP is also more geographically specific than BIZZIP: data are available by street address, thus providing points in space for Geographic Information Systems (GIS) software. However, for Scenescapes we generally aggregated the total number in each YP category to the zip code level, producing a data set similar to BIZZIP. Clearly, emerging sources like Google’s business directories and street-view images or Yelp’s reviews provide significant methodological opportunities.
The yellow pages also have their limits. Unlike BIZZIP, the YP data are not collected with completeness as a goal. Either by accident or by design (e.g., by omitting businesses that do not advertise or have a business phone, for instance), it is more likely to miss establishments. For instance, our BIZZIP amenities cover almost 4,000 more zip codes than YP amenities do (out of the 42,192 US Postal Service zip codes, there are 35,675 with at least one BIZZIP amenity versus 31,874 with at least one YP amenity). Moreover, there is a documented rationale for BIZZIP’s classification system, while no such information is available for YP. YP expresses a market logic rather than government/bureaucratic logic: individual businesses usually select their categories from a preset list, meaning that the categorizations reflect proprietors’ understandings of the nature of their businesses and intended consumers, rather than a social scientist’s understanding. Moreover, so far we have collected yellow pages data for only one time point, 2006 in the United States and 2009 in Canada. Longitudinal data are hard to acquire as past years’ data are not saved or published electronically.
We supplemented YP and BIZZIP with other sources, most notably the DDB Needham Lifestyle Survey (DDB). DDB is a nationally representative survey of American consumers. Since individuals’ attitudes and behaviors are important elements of a scene, we often look to DDB data (which we typically reported in the notes) to examine lifestyle differences among groups (like postgraduates versus holders of bachelor’s degrees) or to investigate characteristics of individuals who live within certain scenes (such as the political attitudes of people who live in Blue Blood counties). More details on the DDB are in the online appendix (www.press.uchicago.edu/sites/scenescapes).
We also investigated many other sources. Using national surveys, IRS returns, and other sources, the Urban Institute’s Unified Database of Arts Organizations (UDAO) provides useful national information about local concentrations of arts and cultural organizations. These data items are often highly specific by discipline (e.g., modern dance, new music, ethnic dance, puppetry, ceramics) and organizational goals (e.g., ethnic and cultural awareness, community service, media dissemination). In addition, we searched for and selectively employed information from specialized organizations, such as the Federation of Genealogy Societies directory, Nielsen BookScan book sales databases, the Theater Communications Group registry of roughly 700 theaters classified according to 24 special interests (e.g., experimental, Asian American, African American, Native American, puppetry, musicals), music venue information from the Pollstar Talent Buyer Directory (which categorizes venues by size and type), musical activities from myspace.com (which categorizes bands by genres),19 and library usage data from the American Library Association. Even if we do not discuss them systematically, they inform the results we do report, as we have explored them in many unreported analyses (such as checking BIZZIP and YP against them for consistency), and we make them available to other researchers on request.
More detailed summaries of specific amenities are in chapter 3. Table 8.1 provides basic descriptive statistics for our main scenes indicators. Other variables and the main amenity types in our analyses are detailed in the online appendix.
To our knowledge, analysts have never before generated such a massive and comprehensive database of amenities and other indicators of the culture of places. It took us many years and is still growing, especially internationally with collaborators from Canada to China, France to Korea (reviewed below). This database provides an unprecedented opportunity both to understand the range of cultural experiences offered across neighborhoods, cities, and regions and to assess the extent to which scenes either share or help define the fortunes and fates of cities. Many data are available to others.
Measuring Scenes
Our amenities database was developed to provide a rich repository of information about the range of experiences and practices on offer in every US zip code. With over 500 types of amenities and nearly four million data points covering some 36,000 US zip codes from YP and BIZZIP alone, it provides a broad sample of cultural meaning. As digitization and computing power continue to advance, more opportunities for more data continue to emerge, which can enhance and deepen insights from these measures of the scenescape.
More importantly, given our theory of scenes, simple counts of amenities like restaurants do not yet tell us what the scene says about how one is supposed to be theatrical, authentic, or legitimate. For this, we need a method for translating amenity counts into measures of the 15 dimensions of theatricality, authenticity, and legitimacy outlined in chapter 2, allowing us to detect interrelations among scenes and other phenomena (economic, residential, political, etc.). The “performance score” featured throughout Scenescapes is the main measure we settled on. Chapter 3 describes the performance scores and how we calculate them in intuitive terms; here we focus more on assumptions and methods guiding their construction. The aim is to demystify our techniques for measuring scenes, beginning with how meaning is systematically attached to each amenity to produce a “standard model” of scenes, continuing on to describe how this meaning is quantitatively aggregated from a count of amenities to produce a single number, and ending with a discussion of composite measures based on performance scores.
Coding Amenities
A central feature of our efforts to translate what is there to what it means is the question of how to accurately capture the scenic characteristics attached to a wide range of amenities. Particle physicists have it easy by comparison: the Standard Model of particle physics does not change based on what the physical world means to particles. Yet the “standard model” of scenes certainly can (and does) change based on how people interpret their environments. A key step in the science of scenes is therefore accurately capturing typical interpretations of amenities in terms of the scenes dimensions they evince, somewhat in the classical mode of Weber’s ideal types, though less as unitary constructs and more as components that can be synergistically combined in multiple ways. Our approach is to use a standardized coding system for classifying an amenity in terms of the 15 dimensions outlined in chapter 2.
The process of attaching scenic meaning to our amenities represents a simultaneously ontological and epistemological action. By recording what are ostensibly the nominal scenic characteristics of a physical structure, we treat scenes as real and take the position that we make scenes real by surveying the meaning associated with these physical structures. The point is not small or tautological. The symbolic meaning of identically labeled amenities will likely differ elsewhere. For instance, in Canada maple syrup camps may express the kind of nationalism captured in our state authenticity dimension in a way that they would not in the United States. This kind of variability justifies reliance on knowledgeable informants to construct our empirical measures of scenes.
The informant method we utilize is not new or as sophisticated as ethnographic or large N surveys.20 Rather, it is based on six informants (“coders”) who were considered to be knowledgeable about the typical significance of the amenities in our database. The decision to use so few informants was partly due to convenience and the exploratory nature of the scenes project as a whole, but also because the threshold for recognizing the meaning of these amenities is intentionally low—we are talking, after all, not about subtle interpretations of complex texts or social interactions but rather whether adult entertainment venues and opera houses affirm or reject transgressive theatricality, for instance. “Intercoder reliability” is similarly employed in many social scientific studies, often with the agreement of even fewer (trained) coders being taken as a strong sign of a reasonably reliable measure. Robustness of our measures to small changes in coding or the amenities included in their calculation is another way to check their reliability, as described in the online appendix.
Our coders also needed to be able to speak the language of scenes. The scoring system was the key for translating the information of amenities into the meaning of 15 scenes dimensions. Chapter 3 provides intuitive examples of this system at work, which we explain in a more technical way here. In this system, each amenity is assigned a score from 1 to 5 on each of the 15 dimensions. Scores 4 and 5 indicate that the amenity (or more specifically, through the practices it supports) affirms the dimension. Scores 1 and 2 indicate that the amenity (through the practices it supports) rejects the dimension. A score of 3 indicates that the amenity’s practices are neutral on the dimension. The most important decision is between a positive (4 or 5) or negative (1 or 2) score. We advised coders to reserve extreme scores (5 and 1) for cases where an amenity’s label clearly and directly indicates a given dimension as a core part of its meaning. The scores 4 and 2 were for cases where the amenity might often or sometimes indicate a positive or negative stance toward the dimension. Thus improv comedy might receive a 5 on self-expression, but a generic art gallery only a 4.21
Applying the Scoring System to Data
A major goal in applying the scoring system to the data was for our coder-informants to use common, clear, and standardized procedures in making their decisions. Coders were asked to read drafts of important foundational documents for the research, which gave an overview similar to chapter 2 on the scenes dimensions and the project. They were then given our “Coder’s Handbook” which provided guidelines for assigning scores, specific questions to ask in coding each dimension, and common pitfalls, as well as a series of examples with rationales. Finally, they were presented with similar information in an online tutorial. In essence, coders became trained experts in linking the 15 dimensions with their potential indicators.
Coders then followed a standardized set of procedures for scoring each amenity. For instance, they coded all items in the data set on traditional legitimacy rather than coding all Chinese restaurants on all 15 dimensions at once. The intention was to focus attention on one dimension at a time and utilize other amenities’ scores as a comparative tool for difficult cases. Coders were then instructed to first ask themselves, in a specific question tailored for each dimension, whether the value was positively affirmed by the activities associated with the amenity. To determine whether the positives should be scored 4 or 5, and the negatives 1 or 2, they were asked a follow-up question.
For example, to code the amenity category improv comedy clubs for the value self-expressive legitimacy, the coder would read the question from “The Coder’s Handbook”: “Is the amenity worthwhile because it expresses a commitment to the importance of personal, creative, unique, novel expressions?” Improvised comedy clearly cultivates spaces valorizing the legitimacy of individual self-expression, indicating that the amenity would receive a positive score. To determine whether the appropriate score is 4 or 5, the coder would then be asked whether “expressing the importance of creative, unique, novel expressions is essential” to the amenity’s message. Again, the answer to this question is yes; improv comedy—relative to other amenities—is deeply connected to expressing a way of life in which authority is ultimately based on the unique expressions of spontaneous selves. Therefore, the amenity would receive a score of 5.
If, however, the answer to the initial question were no, the coder would then be directed to a negated form of the same question—for example, “Does the amenity express the importance of blocking the creative, unique expression of self?” If the answer were no, the amenity would be given a score of 3 on that value. If the answer to that question were yes, then the coder would be asked a final question to determine whether to score 1 or 2. Figure 8.1 represents this decision tree graphically for traditional legitimacy. And table 8.2 summarizes the questions coders asked as they scored each amenity along the 15 dimensions.
This figure shows the decision-making process coders employed in scoring each amenity on each of the scenes dimensions using traditional legitimacy as an example.
The key to consistency was elaborating and detailing the coding guidelines, which we did iteratively as we discovered different assumptions that led coders to different decisions. As the work on coding progressed, all the coders met regularly to discuss significant differences in coding as well as problems that emerged. This often led to revisions in the coding manual and coding process. The coding lasted around a year, with dozens of meetings that led to repeated revisions and clarifications, until we found relative consensus. Where there was doubt or disagreement, instead of asking for the votes of different coders, we encouraged open discussion and writing out the criteria for the decision more explicitly. These discussions verified that all coders were on the same page mentally, generating similar scores for similar reasons (see the helpful discussion of similar issues by Abend, Petre, and Sauder [2013]).
The convergence process was continually monitored numerically through intercoder correlations, for which we set a minimal reliability threshold (Pearson r) of 0.7 and considered scores of 0.8 or above to be good. When we found items with r’s below 0.7, we met and extensively discussed reasons for the differences, seeking to agree on as many criteria as possible that were concrete and replicable in such a manner that others who faced these same issues should reach the same basic conclusions. These discussions led to revising and adding specific criteria for coding in our manual that in subsequent rounds led to r’s greater than 0.7 among coders. Ultimately, we utilized coders’ average scores for each dimension for each amenity, capturing remaining dissension that we provisionally assume represents real differences in symbolic interpretation.
The foundation of the approach was iteratively building a high level of specificity into “The Coder’s Handbook.” The creative part of the coding process was in the discussions, where each coder reported reactions to each amenity. This led us to revise the handbook until it had a level of precision that generated consistency among coders.
No doubt it is still possible to disagree with specific coding decisions.22 One might even imagine utilizing subsets of coders from different regions of the United States and applying their coding scheme to those regions, and the like. We do not insist that our categories or coding decisions are the only valid ones. But we do hold that they conform to the standards of measurement theory similar to most social science.23 Others may organize the same data in different ways to see what happens, or posit the existence of different dimensions of scenes. We repeated essentially similar methods in other national contexts and performed several reliability and validity checks (see the online appendix).
An initial assessment of our coding efforts is their “face validity.” The question is how effective are the resulting measures at supplying evidence for the presence of particular scenes based on the total set of relations among the many amenities in our database. Showing that they are rather effective was the methodological burden of chapter 3. A further test of the coding comes through evaluating the “hypothesis validity” of our measures based on that coding, to which we return below.
Computing Performance Scores
The result of our coding process is a matrix with rows for each amenity and columns for the 15 scenes dimension scores that coders, on average, assigned to an amenity. The online appendix summarizes the coding scores for all BIZZIP and YP amenities, respectively. Table 8.3 shows an example of one row.
| Table 8.3. Sample scene profile of one amenity, art dealers | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| NAICS | d11i | d12i | d13i | d14i | d15i | d21i | d22i | d23i | d24i | d25i | d26i | d31i | d32i | d33i | d34i |
|
453920 |
3 |
4 |
2.25 |
4 |
3 |
3 |
3.6 |
3 |
3.75 |
3 |
2 |
3 |
3 |
2.75 |
3 |
In table 8.3, 453920 is the US Census Bureau’s NAICS industry code for art dealers, an amenity in the BIZZIP data set. The variables d11i, d12i, . . . d34i represent the 15 scenes dimensions, with the cell numbers indicating the average coder score associated with each dimension for art dealers. For example, coders judged that the existence of an art dealer indicates more self-expression (d12i = 4) and less utilitarianism (d13i = 2.25).
This sort of matrix is the “standard model” of scenes, capturing how our amenities should relate to each dimension according to the symbolic characteristics they embody. Such matrices position each amenity within a complex of cultural meaning. Similarly scored amenities indicate a similarly supportive or suppressive connection to each dimension. Many researchers will recognize this matrix as similar to hypothesized factor loadings associated with a confirmatory factor analysis. Typically the next step is to use these factor loadings to construct the hypothesized covariance matrix, essentially capturing which amenities should or should not colocate in the same geographic area, based on their scenes characteristics. The final step would be to compare this covariance matrix with a covariance matrix produced by examining the actual clustering of amenities in our data. Too poor a fit between the two matrices would indicate that amenities do not cluster together according to the “standard model” of scenes that coders created.
We have opted to not apply the typical fit criterion for confirmatory factor analysis for two important reasons. First, we explicitly acknowledge that cultural meanings may not be the primary driver for the creation and support of certain amenities. For instance, economic incentives provided by localities (e.g., tax breaks or funding for public recreation) will likely affect which amenities exist in a community. This is critical to recognize for many of our analyses, since they are primarily about near-term consequences or what happens given the presence of a certain scene, not why a certain scene came into being in the first place.
Second, historic processes that far precede our data collection have almost certainly had a path-dependent effect on the configuration of amenities in an area. Communities are always changing, evolving, and refinding their pasts, and not only are our amenities data largely cross-sectional at this time but most current data sources only extend back to the early 1990s. Development processes that span multiple decades are simply not something we can accurately measure. There is clearly “more” in our database than we are in a position to interpret—it contains patterns that are beyond our analytic scope and/or are highly localized. Since all these issues correlate with poorer fit between our hypothesized factor loadings and the actual clustering of amenities, and since we constructed these theoretical categories to represent something different from the mere copresence of amenities, we constructed our own pragmatic method for testing how well our measures of scenic meaning perform.
Rather than rely on traditional measures of statistical fit, we tested the construct validity of our scenes dimensions by postulating an extensive set of falsifiable hypotheses involving (what we believe to be) measures of particular scenes and then determining whether these hypotheses are supported empirically, as reported in chapters 4, 5, and 6. Our logic of validity assessment is that if we have poorly measured scenes or our theory of scenes does not provide a consistent and meaningful way of understanding relevant cultural meanings, then we should not observe the relationships posed by our hypotheses. Part of providing an extensive set of hypotheses in Scenescapes has been to demonstrate a broad pattern of results confirming the presence of expected relationships, which we take to mean that we have created a sufficiently specialized and sensitive way of measuring what we call scenes using the indirect evidence of amenities.24
To carry this out, we used our matrix of coder scores to calculate what we call performance scores, which represent a way to aggregate information about the expressive content embodied in our data sets. In essence, performance scores summarize the degree to which the typical amenity in a place affirms or attacks each of the 15 dimensions of theatricality, authenticity, and legitimacy outlined in our theory of scenes. They provide an overall profile of the scene indicated by the set of amenities in a place, allowing for the fact that the same dimension (e.g., transgression) can be affirmed by many different amenities (e.g., body piercing studios and tattoo parlors).
The performance scores combine the scores assigned by coders to each amenity with data on the number of each type of amenity located in a zip code. The discussion of “transgressies” in chapter 3 outlines how performance scores are computed in clear and intuitive terms. Computing performance scores is actually quite simple conceptually. Take a hypothetical zip code, Zip Code #1. One calculates Zip Code #1’s transgression performance score as follows: Suppose Zip Code #1 has five total amenities: four body piercing studios and one Catholic church. Suppose also that body piercing studios were scored 5 on transgressive theatricality while Catholic churches were scored 1. Multiply the number of each type of amenity (4 body piercing studios, 1 Catholic church) by that type’s transgression score (5 and 1). Sum the product and you get 21. Zip Code #1’s total output of transgressive theatricality, which we sometimes refer to as its intensity score, is 21.25 Now divide that total output by the total number of amenities in the zip code (in this case, 5). The result of that division, 4.2, is Zip Code #1’s transgression performance score.
A different zip code, say, Zip Code #2, with four Catholic Churches and one body piercing studio, would thus have a transgression performance score of 1.8. A greater share of the amenities in Zip Code #1 indicates the presence of transgressive theatricality. Here is another similarity to factor analysis—but instead of showing that the amenities in a zip code tend to point toward a statistical regularity, performance scores show which zip codes’ amenities point toward theoretically significant dimensions of cultural meaning. To include an additional amenity in the calculation of a performance score, one need only add it into our matrix and include its dimensional score. Figure 8.2 illustrates the computation process.
The 15 performance scores taken together thus represent a standardized measure of a location’s scene, a holistic reconstruction of the scene detected in that location. Other measures are no doubt possible and necessary, as is on-the-ground ethnographic work that can capture subtle variations of amenity usage and meanings—we have explored alternative constructions of our scenes measures and encourage others to do so as well.26 That said, this approach represents a way of constructing scenes measures that can be easily repeated. For example, performance scores have been successfully created using data sets similar to BIZZIP and YP from Canada, France, Spain, Japan, Korea, and Poland, and we are refining methods of performing cross-national scene comparisons.
We tested the sensitivity of the performance scores to the number of amenities included in their calculation by recomputing several scores with smaller numbers of amenities. Changes were generally small, which indicated to us that individual amenities are unlikely to be biasing our measures. Such small changes also increased our confidence in the construct validity of performance scores. The online appendix section on sensitivity checks details some of these analyses.
We have pursued many such alternative constructions, methods, and sensitivity tests. Combining amenities, moreover, yields much more cultural specificity than any single amenity can, so that national averages become modified by local context. Thus, as we saw in chapter 3, the correlations among amenities and dimensions often vary across regions and cities in ways that reveal their specific characters: for example, charisma is correlated with neighborliness in Chicago but with self-expression in Los Angeles. Looking for how relations (e.g., between charisma, self-expression, and neighborliness) shift across contexts in this way permits us to join our cross-national focus with local specificity.
Measures That Combine Performance Scores
The performance scores are our main measures of the 15 dimensions of scenes. We analyzed them in many ways. The simplest is to treat them separately, for instance, in analyzing the separate impacts of self-expression, glamour, and tradition on economic development, as in chapter 4. However, the theory of scenes suggests we look to combinations of dimensions as well. So too does descriptive analysis of the performance scores. For instance, if we look only at the national average performance scores for all US zip codes, we find that that the average zip code positively affirms both the legitimacy of tradition and of self-expression—that is, the national average performance scores for self-expression and tradition are both positive. This is somewhat misleading, however, as the two performance scores are negatively correlated. That is, many places are high in tradition; many are high in self-expression; but places high in one are usually low in the other. To account for the fact that scenes are combinations of positive and negative relations to many dimensions of meaning, we placed a high priority on capturing combinations of the performance scores. We did so in a number of ways.
One way to combine performance scores is simply by creating a multiplicative interaction term for two or more dimensions, as in our analysis of self-expression times transgression in chapter 5 and below. Another approach we explored (but ultimately did not include) is to calculate the average sum of several dimensions for a geographic area. And still another approach is to compare a zip code’s actual performance scores across the 15 dimensions to the scores that a theoretically important ideal-typical scene would receive, as in our “bliss point” measure of Bohemia in chapter 4 (examples of the formulas for computing these types of combinations are in the online appendix).27
Exploratory factor analysis offers another helpful way to draw out how scenes combine in situ. We created factor scores using all 15 dimensions of the YP and BIZZIP data. The six factors on which we have focused most attention in our analyses are summarized in table 8.4. We discussed these factors because they had clear resonance with key scenes from classical social, urban, and cultural theory.
| Table 8.4. Factor score weights for performance scores | ||||||
|---|---|---|---|---|---|---|
| Urbanity (YP1) | LA-LA Land (YP2) | Rossini’s Tour (YP3) | City on a Hill (YP4) | Nerdistan (BZ3) | Renoir’s Loge (CBP3) | |
|
Traditionalism |
−0.786 |
0.384 |
0.031 |
0.438 |
−0.176 |
0.100 |
|
Self-expression |
0.307 |
−0.326 |
0.860 |
0.005 |
−0.210 |
0.398 |
|
Utilitarianism |
0.836 |
0.003 |
−0.056 |
0.190 |
0.355 |
−0.363 |
|
Charismatic |
−0.788 |
−0.030 |
−0.132 |
0.492 |
−0.127 |
0.734 |
|
Egalitarian |
−0.087 |
0.185 |
−0.005 |
0.948 |
−0.038 |
0.176 |
|
Neighborliness |
−0.899 |
0.230 |
−0.237 |
0.353 |
−0.021 |
0.107 |
|
Formality |
−0.840 |
−0.040 |
−0.390 |
0.040 |
0.603 |
0.571 |
|
Exhibitionism |
−0.015 |
−0.885 |
0.006 |
−0.049 |
−0.007 |
−0.166 |
|
Glamour |
0.417 |
−0.756 |
0.194 |
−0.450 |
0.063 |
0.870 |
|
Transgressive |
0.853 |
−0.353 |
0.275 |
−0.320 |
−0.181 |
−0.114 |
|
Rational |
0.834 |
−0.064 |
0.394 |
0.060 |
0.651 |
−0.217 |
|
Localism |
0.051 |
0.235 |
0.885 |
0.047 |
−0.051 |
−0.015 |
|
Statism |
0.873 |
−0.154 |
0.498 |
−0.135 |
0.102 |
0.033 |
|
Corporatism |
0.856 |
−0.214 |
0.048 |
−0.170 |
0.287 |
−0.320 |
|
Ethnicity |
0.183 |
−0.173 |
0.295 |
−0.192 |
−0.716 |
−0.010 |
|
Note: This table shows results of factor analyses applied separately to our YP and BIZZIP performance scores. It shows the first four factors from our US YP performance scores. Because we also analyzed them in chapter 5, we show Nerdistan, the third factor from the US BIZZIP performance scores, and Renoir’s Lodge (CBP3), the third factor from the Canadian CBP performance scores. The first two BIZZIP factors extracted joined dimensions similar to YP. Results for LA-LA Land reported in the text reverse its sign for easier interpretability, that is, in the chapter 5 bubble figures, a positive coefficient for this factor indicates a positive relationship to glamour and exhibition. Key dimensions are in bold. Extraction method: principal component analysis. Rotation method: Oblimin with Kaiser normalization. |
||||||
The first US factor accounts for approximately 44 percent of the total variance among the 15 dimensions, and we described it in chapter 3 under the label Communitarianism versus Urbanity. On one end of the spectrum are Urbane scenes of transgression, utilitarianism, reason, state, and corporation, which have the strongest loadings, followed by glamour and then self-expression. On the other end are Communitarian scenes of tradition, neighborliness, charisma, and formality. In essence, the first factor summarizes in a single number the most powerful cleavage dividing the American scenescape. We therefore include it in our core set of control independent variables. This is methodologically significant: though we describe and analyze Communitarianism/Urbanity at times, we are often less interested in the development consequences of a generic Gemeinshaft/Gesellschaft scene and more interested in the specific impacts of, say, a self-expressive scene net of, or over and above, this generic difference. Including both the Communitarian/Urbanity factor and some individual variables (like the self-expression performance score) permits us to identify their separate effects, though not perfectly since some factors include multiple individual variables, as we discuss. The other three YP factors (LA-LA Land, Rossini’s Tour, and City on a Hill) were featured in chapter 5, along with the third BIZZIP factor, Nerdistan. We include the third Canadian factor (from Canadian Business Patterns, similar to BIZZIP), Renoir’s Loge, in table 8.4 because it was analyzed in chapter 4.
We experimented with other ways to measure the sorts of complex scenes outlined in chapter 2. One technique was coding our data directly to measure, for instance, the concepts of Bohemia or Disney Heaven. Another was to create ad hoc indexes based on smaller subsets of amenities, usually by summing them, as in our New Con or Blue Blood indexes.28 Others can adapt these and additional approaches to create indexes of their favorite concepts for urban comparisons. Dozens of students have used these data for all sorts of purposes.
Methods of Analysis: General Analytical Approach, Standard Method of Analysis, and Modeling Context
One part of detecting scenes involves the careful construction of the best possible measures of their presence. But an equally important part of demonstrating the validity of scenes measures is confirming that these measures exhibit relationships with other variables that are both expected and consistent with theory.
The question then becomes, what conditions have to be met before we consider our results robust? Part of the answer involves accounting for known relationships and likely confounders among common (and not so common) variables and our outcomes—that is, accounting for potential spuriousness. The other significantly more complex part of the answer involves choosing the appropriate multivariate methods of analysis. We discuss both in turn.
General Analytical Strategy: The Core and More
Most multivariate analysis in Scenescapes includes a set of variables that measure key factors associated with urban development. Largely drawn from the US Census of Population, they typically represent aspects of the urban environment commonly studied by urban economists, geographers, sociologists, and public policy advisors when analyzing the outcomes we employ. Our general strategy is to include this set of independent variables drawn from past research (which we term “the Core”) to account for the influence of more traditional factors. The Core includes eight independent variables: population size, percent of population who are nonwhite, median gross rent, percent college graduates, percent Democratic vote for president, crime rate, the location quotient of cultural jobs, and the Urbanity factor score. Including the Core in multivariate analysis tests for possible spuriousness of our scenes measures. We also often explored how scenes variables mediate relationships of other variables with our outcomes.29
In order to connect our hypotheses to ongoing academic debates, our main analytical chapters used basic outcome measures similar to those employed by leading researchers in the subfield studying each dependent variable. In many analytical chapters, we employed other variables relevant to the specific issues discussed there; usually these were included in addition to the Core group of independent variables. These other variables include, for instance, technology jobs, neighborhood walkability, evangelical Christians, age, zip code rent, and the like. Often they are an integral part of the hypothesis being tested. Sometimes they supplement the Core, representing additional controls when theory or past research suggests their relevance. All variables are listed in the online appendix, together with basic descriptive statistics.
Performance Scores as Independent Variables
Performance scores are often key independent variables of interest in our models.30 That is, Scenescapes typically asks what follows from the presence of a type of scene rather than what encourages that type of scene. This by no means implies that we believe scenes are uncaused causes—we illustrate some sources of scenes below. Nor does highlighting performance scores and amenities imply that these are the only or primary drivers of development. We often stress how our amenities-based measures shift relationships of other important variables. We seek to add scenes to the analytical mix, not to replace other important variables.
Nevertheless, as our primary interest is in scenes, we have constructed our hypotheses to explicitly test whether our measures demonstrate theoretically consistent relationships with these outcomes. As suggested above, these hypotheses themselves help assess the robustness of patterns in which scenes may operate as independent variables. Yet the possibility that performance scores are actually measuring some other more traditional feature of the urban landscape is always a concern to address systematically.
To assess the degree to which scenes measures might be proxies for other factors, we examined their correlations with many sociodemographic and other variables—the Core and many of the “other variables” listed in the online appendix (26 in all). The lack of strong relationships is very striking. The average absolute value of Pearson correlations between the YP and BIZZIP performance score measures of scenes and the 26 variables is below 0.1 (0.087 for YP, 0.074 for BIZZIP). As we might expect, education and rent show significant average correlations with many performance scores, though the connections are weak to moderate: their average correlation with the YP and BIZZIP performance scores is about 0.16 (and change in college graduates, often given much weight in analyses of urban change, is even lower: 0.03).
To further scrutinize the analytical independence of the performance scores, we analyzed each of the 15 performance scores as dependent variables, including the aforementioned 26 as independent variables. The extent to which variance in performance scores is explained by this expanded model is low: the adjusted R-squared for nearly all performance scores is below 0.2 (only a few are slightly above), with most being considerably lower. The same holds even for the combinatorial measures of scenes: the Core predicts our main factor score relatively weakly, with the adjusted R-squared at only 0.13. All of which is to say that scenes cannot be explained away as relabeling old concepts and measures with new names. What the scene is cannot be read off reports of education levels or average incomes, since many places with similar levels of each have very different scenes.
Similarly, our main analyses presume that scenes are relatively stable. But scenes clearly are often in flux, changing in nonrandom ways that may influence our analyses. A scene may owe its self-expressive qualities in 2000 to the fact that college graduates moved there in 1980 and slowly started building (or inspiring the building of) amenities to suit their tastes, changing the scene accordingly. Still, even though scenes change to reflect the tastes and values of residents (not to mention visitors, business owners, and politicians), after they emerge, they can “take on a life of their own” and acquire “emergent properties” that go beyond whatever processes generated them. For instance, once a scene that offers opportunities to cultivate individual self-expression emerges, however it got there, it can provide continuing attractions that often outlast the specific people or processes that stimulated its creation.31
Given the limits of our data, we have explored this hypothesis in a number of ways. One analysis of stability of our scenes measures over time used four time points (1998, 2000, 2001, and 2004) and county-level data. The goal was to understand the “stickiness” of scenes by looking at how quickly scenes change given a certain degree of churn in the numbers and types of amenities within a county. On average, the total number of coded amenities in a county increased by 7 percent from 1998 to 2004. Conversely, performance scores associated with all 15 dimensions shifted, on average, less than 1 percent during the same time period. Analysis of Canadian performance scores also revealed that scenes are not temporally fleeting: the average correlation of the 2001 and 2011 Canadian FSA performance scores was 0.77; 2007 and 2011 were correlated 0.86 (these are the years for which we have data). Here is some evidence of the “rolling inertia” that Molotch, Freudenburg, and Paulsen (2000) suggest make up the character of places. More broadly, we locate the difficulties we face in making causal claims about scenes with a set of general concerns shared in the social sciences at large, considered below in our discussion of “the chicken and the egg.”
Multivariate Methods
Choosing a multivariate method for testing relationships between various measures of scenes and our outcomes is a crucial decision that means balancing parsimony and accuracy. Simpler methods are preferable for ease of interpretation, but scenes pose a unique challenge by their explicitly spatial nature. Residents and nonresidents frequent a scene, and the expanse of a scene may be a few buildings, a few city blocks, or even a few miles. How do we make sense of these scales analytically, acknowledging variation while still applying a method that consistently picks out the appropriate “signal” for exploring the relationships we hypothesize?
A major issue in choosing a multivariate method for scenes analysis involves units of analysis and their geographic scope. Substantial questions arise as to the appropriate geographic scope of many variables in our models, but especially regarding a scene’s impacts. Scenes might cross zip codes, and their attractions might extend to people who live nearby but not in the same zip code. Similarly, a variable like crime rate is reported at the county level but depending on population density (and other factors) the effects of crime might be significantly more localized, or spread around the county as a whole. We address these concerns in the section below on modeling context. In general, however, we started with the smallest reasonable units and then compared results for successively larger units as one method to assess the size of distinct catchment areas, since catchment areas vary for different scenes and processes.
Another important issue is connected with the nature of large data sets. Since scenes analysis tends to involve thousands of cases, we often observe highly statistically significant yet quite modest numerical relationships, like a Pearson correlation of 0.15. This is both the blessing and curse of having so many cases, requiring us to be circumspect in the claims we make about the strength of observed relationships. One simple solution we have employed is to not only report standardized regression coefficients, or significance levels, but also often graph the relative size of these coefficients so that readers can easily compare the importance of scenes with the importance of other variables.
Along with large numbers of cases, our scenes database involves some items that were either not available in a consistent manner nationally or simply exhibit distributionally troublesome traits for statistical modeling, such as an excess of zeros. Typically, we observe excess zeros in relatively rare amenities, like film studios, and largely solve this issue by creating additive indexes of several amenities when exploring clusters of amenities outside of performance scores. Similarly, variables are transformed when reasonable to do so, though persistent issues of missing data were more difficult to address. For the most part we assume that data are missing completely at random (MCAR), but we do not employ imputation methods since we lack the expertise/resources to implement a proper imputation scheme or decide whether to impute amenities or performance scores (for instance).32
Finally, the indirect nature of our evidence had to be considered when choosing a multivariate method for scenes analysis. Just as our performance scores represent imperfect indicators of a scene, so too will any large-scale quantitative analysis imperfectly capture the dynamics of the hypotheses. The complex narrative that leads some young professionals to live in high crime areas due to the presence of an appealing scene cannot be adequately told by standard methods.
This is where work like that of Richard Lloyd in Neo-Bohemia and other more ethnographic approaches becomes extremely valuable for validating or challenging our results. Where possible we include tidbits of ethnographic observation intended to flesh out the cold numbers produced by multivariate analysis. The boxes in the text are often critical in adding more subtle meaning to our interpretations.
Standard Method of Analysis
Determining the simplest multivariate method possible for analyzing the relationship between scenes and our outcomes turns on addressing the issues identified above. The goal is to detect evidence for a relationship that assesses our original hypothesis and can be clearly translated into a narrative for further investigation.
Linear regression is the simplest multivariate method we could have employed, and largely for this reason we began with ordinary least squares (OLS) regression as our standard multivariate method of analysis. We follow a long tradition here. Our analyses generally began with OLS models for each outcome, included the Core as independent variables, and explored both listwise and pairwise deletion. This method of beginning with a set of core variables and adding other measures to the analysis as needed has been robust in past works (cf. Clark 2004, 2011a). To these models we add other variables, often based on past research, one or a few at a time, depending on the substantive proposition. This allows us to assess both their direct effects on the outcome and their possible indirect effects on the relationships involving Core variables.
Some variables were too strongly intercorrelated to permit including them all simultaneously. We normally omitted independent variables when two or more exhibited Pearson correlations that exceeded 0.5, but we also relied on variance inflation factors (VIFs) to detect problematic multicollinearity. For most models, the VIF of independent variables does not exceed 4, and in no case does it exceed 10.
The assumption of independence of cases in classic linear regression (ordinary least squares) is clearly broken when we combine data whose units of analysis differ in geographic scope (e.g., assigning the same county total population to each of its zip codes). Moreover, linear regression fails to account for spatial configurations, by assuming independence even between two zip codes in the same city neighborhood, as if they were tiny prisons whose residents never left and which outsiders never visited. This is one of several standard assumptions of OLS that are often hard to justify in practice.
To check whether violating linear regression assumptions was substantively biasing our results, we tested many variations in model specification and replicated our main analyses using other more complex methods (like mixed/multilevel models) in comparison to OLS. Out of 527 total coefficients we examined, we found that 12 percent disagreed between OLS and mixed models (i.e., switched from positive or negative to insignificant, from insignificant to positive or negative, or from one sign to the other). Differences were slightly smaller for scene variables (10 percent). While not as large as one might imagine, this was enough to convince us to make mixed models the default method of analysis in Scenescapes, rather than OLS, because they more directly take account of the embededness of zip code scenes in broader (e.g., county) contexts.33 Still, multilevel models are just one approach to explicitly modeling context, and combining this approach with others can enrich our understanding of some of the processes investigated in Scenescapes. To illustrate this general statement, we take a deep dive into one analysis—the chapter 5 examination of the relationship between self-expression, transgression, and change in the 25- to 34-year-old share of the population.
Modeling Context
We explored several additional methodological strategies to check and extend the standard methods typically reported in Scenescapes. One key area for such exploration concerns spatial context—the idea that the meaning and impacts of scenes may vary depending on where they are located, just as much as other variables may operate differently depending on their proximity to a given scene. “Proximity alters performance” was the slogan for this idea in chapter 3. We illustrated this concept with relatively simple methods, such as showing in chapter 3 how correlations between charisma, self-expression, and neighborliness vary in Chicago versus Los Angeles, or showing in chapter 4 how the impact of Bohemian zip codes on economic growth depends on how communitarian or urbane their surrounding counties are.
Here we link a particularly intriguing finding from chapter 5—that 25- to 34-year-olds were increasing between 1990 and 2000 in transgressive but not self-expressive scenes—to more sophisticated methods that progressively add more depth and detail to our understanding of the role of spatial context in scenes analysis. Scenes, we have stressed, are not just one more variable to add to a multivariate model; they are also contexts that potentially transform the very meaning of other variables. Systematically analyzing such contexts, however, is challenging methodologically. Indeed, only recent innovations make it possible to consider a range of indirect, interactive, and multilevel processes in ways that allow us to test (and relax) some of our assumptions regarding space and scenes.
Consider these more sophisticated techniques for the chapter 5 analysis of transgression, self-expression, and 25- to 34-year-olds. We present three ways of further analyzing this relationship: multiplicative interaction terms, parameterization of spatial adjacency among scenes, and calculation of performance scores using different catchment areas—often exploring the extent to which the nested structure of our data biases results from traditional OLS models. Each approach has its distinctive strengths and weaknesses, but together they help illustrate some contextual dynamics of scenes and link scenes analysis with cutting-edge work on modeling local area and spatial effects. Perhaps more importantly, attacking the same problem from multiple methodological angles permits a more creative and subtle interpretation of how the scenic context of a location can transform its meaning.
Interaction terms—within zip code. One classic approach to modeling context in a linear regression framework is to create multiplicative interaction terms. This is a common way of gaining a potentially useful picture of how two variables mediate each other. Given the standard nature of this method of parameterizing context for use in modeling, we often explored the significance of such multiplicative interaction terms. Table 8.5 summarizes results of this and subsequent analyses.
Consider the base model in table 8.5 reporting OLS regression coefficient estimates for our chapter 5 example of change in 25- to 34-year-olds across all US zip codes from 1990 to 2000. This model separates self-expression from transgression: self-expressive scenes show less than average growth in the 25- to 34-year-old population, while transgressive scenes show more than average growth. We may then raise the question of context: Does the level of transgression in a self-expressive scene alter these relationships or exhibit its own unique effect?
In our example, the column of table 8.5 labeled “Interaction—within zip code” presents results for an OLS regression where a multiplicative interaction term for self-expression and transgression has been added to the base model. We call this a within–zip code interaction to highlight how a heightened level of transgressiveness within a given zip code may alter the impacts of that self-expressiveness within that same zip code. We contrast this with a neighbor interaction approach that highlights how a heightened level of transgressiveness among adjacent zip codes alters the effects of a given zip code’s self-expressiveness, elaborated in more detail below.
As table 8.5 shows, in the case of the within–zip code interaction, the main effects of self-expression and transgression on change in 25- to 34-year-olds remain the same: negative for the former, positive for the latter. At the same time, the estimated regression coefficient for their interaction is positive, indicating that the presence of one amplifies the effect of the other.
Interpreting interaction terms can be arduous and requires a sound understanding of how the magnitude of estimated coefficients for both main effects and the interaction term operate. Reasoning in terms of how scenes shape a context can sometimes add substance and specificity to the interpretation. One method is to graphically display the specific interaction result, as we did in several cases in chapters 4, 6, and 7.34 More specifically, one can graph predicted values of the outcome based on the average value of other variables and a demonstrative manipulation of the two variables that constitute the interaction. One way to do so is by plotting predicted responses in the outcome given continuous variation in one variable while holding the other variable at a constant value.
Plotting three lines shows the predicted response of our outcome given a continuous change in one component of the interaction term and three categorically different levels of the other component, holding other variables constant. Thus in chapter 4 we used this technique to show how low, average, and high levels of the Renoir’s Loge scene (a factor score combination of charisma, glamour, and self-expression) shift the relationship between artist concentrations and economic growth.
For table 8.5, we simply note that the regression coefficient for self-expression is on par with that of the interaction term. This implies a relatively complicated state of affairs. Self-expressive scenes are negatively related to change in 24- to 35-year-olds, and they exhibit a net neutral relationship in the presence of a transgressive scene but an especially negative relationship in the presence of antitransgressive scenes.
Interaction terms—neighboring zip codes. While within–zip code interactions involve relatively standard methods, an open methodological question regarding scenes, mentioned already, is their typical geographic extent. A more novel method to tackle this problem is to extend traditional multiplicative interaction terms to incorporate a form of spatial dependence. The idea is to create variables that capture characteristics of neighboring zip codes, and parameterize them for use in subsequent modeling.
Other ways of modeling spatial dependence are possible, and we present one alternative below, but adjacent zip codes (those that touch geographic borders) are a reasonable place to start. After the zip code of residence, residents likely have easiest access to the scenes in the immediate vicinity of where they live—these adjacent areas are the most likely to be within walking distance and require the lowest travel time. (This is one implicit assumption of studying how scenes relate to demographic and economic trends using zip codes.) Thus we constructed measures for each scene dimension for both YP and BIZZIP based on the average performance score of immediately surrounding zip codes.35
Avoiding spatial autocorrelation is often the main rationale for investigating how a given analysis may be biased by failing to take into account space. A classic case is racial segregation. If underinvestment in inner-city neighborhoods occurs because such neighborhoods all tend to be physically clustered (e.g., in ghettos) and thus share a larger area of economic deprivation, then analyzing the characteristics of each neighborhood separately will not capture this larger context. Since it is now well established that neighborhoods where many African Americans live are highly likely to be near other such neighborhoods, ignoring this context can lead policy makers to develop inefficient or ineffective programs for urban revitalization, to the extent that the success of local interventions depends in some part on factors operating at a scale larger than the zip code.
This well-known example provides a convenient way to assess how well our adjacency technique captures spatial autocorrelation, since the mean value of adjacent zip codes is not a common way to address potential spatial dependence of a variable.36 We thus performed the same procedure on percent nonwhite in a zip code as we did for our performance scores: we computed the average percent nonwhite for all adjacent zip codes for every US zip code. We then calculated a simple Pearson correlation between percent nonwhite in a zip code and average percent nonwhite in adjacent zip codes. Results indicated a very strong direct relationship between the two (above 0.8), suggesting that if scenes are strongly spatially autocorrelated, then our technique would pick this up as well.
Are zip codes with similar scenes likely to be near one another? By comparing the performance score of a given zip code (i.e., the target zip code) with the mean performance score of adjacent zip codes (i.e., neighbors), we find quite different results than for nonwhites: there is little spillover of scenes across zip codes, or general spatial ordering of scenes. Tables 8.6 and 8.7 report Pearson correlations between each dimension’s performance score in a zip code and the neighboring zip codes’ mean performance scores along each dimension, for YP and BIZZIP respectively.
| Table 8.6. Pearson correlations between yellow pages (YP) performance scores of target zip code and neighboring zip codes |
|---|
Note: This table indicates that the scene in one zip code is a poor predictor of scenes in adjacent zip codes.
**p < 0.01 (2-tailed)
*p < 0.05 (1-tailed)
| Table 8.7. Pearson correlations between Zip Code Business Patterns (BIZZIP) performance scores of target zip code and neighboring zip codes |
|---|
Note: This table indicates that the scene in one zip code is a poor predictor of scenes in adjacent zip codes.
**p < 0.01 (2-tailed)
*p < 0.05 (1-tailed)
Tables 8.6 and 8.7 go beyond looking for simple spatial autocorrelation, however. Scenes may not be spatially autocorrelated along single dimensions but rather show spatial patterns across different or multiple dimensions. For example, self-expressive scenes may usually be next to transgressive scenes but rarely next to other self-expressive scenes.37 Even casting such a wide net, we find no strong correlations suggesting a clustering of scenes across zip codes. In fact, Pearson correlations never reach 0.4 in magnitude and rarely exceed 0.3. This suggests that the “right” scale for studying scenes—at least using the kinds of data and analyses in Scenescapes—is generally unlikely to be much larger than the zip code. This result offers some justification for using zip codes as the main unit for analyzing scenes, and suggests spatial autocorrelation is generally less a source of bias in the case of scenes than in the case of race (as measured for the United States in this time period).
Analyzing adjacency is valuable not only for avoiding statistical bias. One can also interpret how much spatial proximity shifts the impacts of other variables.38 Thus, armed with the knowledge that neighboring zip codes generally do not replicate the scene of the target zip code, we can gain further insight into how the effect of that scene is mediated by its surrounding scenic context. That is, because zip codes are generally weakly correlated with the scenes of their neighboring zip codes, we can test hypotheses about how the scenes in the neighboring zip codes may alter the character and consequences of scenes in the original target zip code.
Returning to our example, this approach allows us to explicitly model how the spatial proximity of self-expression and transgression shifts their impact on 25- to 34-year-old population growth. We do this by adding measures of average transgressiveness in neighboring zip codes to the model. The last column of table 8.5, labeled “Interaction—Neighbor” adds (a) the average transgression performance score of neighboring zip codes and (b) the multiplicative interaction of this measure with the original zip code’s self-expression performance score. The estimated regression coefficients for these variables tell us (a) whether neighboring zip codes’ transgressiveness has an independent effect on a zip code’s change in 25- to 34-year-olds and (b) whether self-expressive scenes next to transgressive scenes shift outcomes.
Results suggest that this spatial configuration modeling not only matters but accounts for the independent effect of self-expressive scenes. Self-expression alone, recall, suppresses change in 25- to 34-year-olds. However, when we add the transgressiveness of adjacent zip codes, self-expression becomes insignificant. Further, the positive relationship between transgression and change in 25- to 34-year-olds extends beyond the target zip code—whether a zip code’s neighboring zip codes have transgressive scenes is actually more strongly related to our outcome than the transgressiveness of the target zip code itself.
Finally, in considering the joint effect of transgressiveness and adjacent self-expressiveness, we observe a surprisingly powerful relationship that would have gone unnoticed otherwise. Specifically, zip codes high in self-expression also adjacent to zip codes high in transgression were more likely to increase in 25- to 34-year-olds than zip codes that were only more transgressive. This cross–zip code interaction is also more strongly related to growth in 25- to 34-year-olds than the within–zip code interaction between self-expression and transgression. The spatial configuration of scenes, in other words, may have a pronounced effect on certain outcomes—in this case being near a transgressive scene heightens the allure of self-expression over and above simply being in a transgressive scene. For many young people, that is to say, transgression may be more appealing when they do not have to spend all night surrounded by red lights and loud bars but can retire to a less edgy, but still interesting, self-expressive scene.39
Multilevel models. Employing this relatively simple spatial measure thus provides a refined level of analytical control over the social environments we are studying. Yet thus far we have still been violating a principal assumption of linear regression—independence of cases—by including county-level variables in multivariate models without accounting for the clustering of zip codes within counties. We next explore how analyses may be further refined through a method designed to model such clustering.
Social scientific data is often collected at multiple, nested levels. The now classic example, explored by Raudenbush and Bryk (2002), is the nesting of students within classrooms, where student performance is not just a function of individual characteristics but also of classroom characteristics. Their hierarchical linear modeling (HLM) is largely associated with such one-to-one hierarchical relationships between levels of analysis, though the analytical approach is known by many names: mixed models, multilevel models, random-effects models, and so on. Whatever the label, the central premise is that we must account for the multilevel structure of the data because not doing so may produce biased results. For example, in the case of student performance, comparing two demographically identical students fails to account for teacher effectiveness and peer influences unless these differences in the two different classrooms are explicitly modeled.
In the case of scenes, comparing two demographically identical zip codes from two different counties fails to account for county-level characteristics—economic, demographic, or political, like the presence of a Superfund site. An initial approach to this issue was to utilize some county-level variables to capture context in zip code–based OLS regressions. We did this by joining the two levels in a single data file, which simply repeats the county values for each zip code within that county. This has the effect of underestimating the standard errors for county-based variables, thus biasing the normal significance computations.
Why investigate counties? The county is a larger and more sensible catchment area for certain characteristics, like cost of living or wages or rent, where low transportation costs allow individuals wider choices in location of residence. In other cases, such as crime rate, we were limited to county-level measures because the data are not disclosed at a lower level (by the FBI which adjusts noncomparable local crime data). OLS is a good first step, but it does not fully account for the specific nesting of zip codes within counties.
We thus estimated the appropriate mixed models for our initial OLS regressions. While multilevel models can allow estimates of regression coefficients that vary according to the higher-level units of analysis (e.g., counties), we typically estimate random intercept models (i.e., intercepts vary by counties), though we did explore some random-effects models. This means that an average value for our outcome variable is estimated for each county, but a single value is estimated for the regression coefficients associated with our zip code–level variables. Our county-level variables (e.g., total population) are used to model the county-specific average values of our outcome.
See below for the generic model specification for our outcomes using the Core as independent variables:
(8.1a) Yij = β0j + β1j × CollProfLv90 + β2j × LevelNonWhite90
+ β3j × ARTGOSLG98a + β4j × YPFactorScore + rij
where
(8.1b) β0j = γ00 + γ01 × ITEM005 + γ02 × ITEM218
+ γ03 × ITEM108 + γ04 × CrimeRate1999county + u0j
HLM/mixed models in table 8.8 replicate the key OLS analyses in table 8.5.40 The OLS regression results are effectively confirmed. Some standard errors shrink, but others actually increase. This is unsurprising as the use of HLM allows us to more precisely analyze the structure of our data, and we expect some differences in the magnitude and/or significance of coefficient estimates in HLM versus OLS models. In every case, however, estimates are consistent in terms of the direction and approximate magnitude of relationships between our variables of interest and our outcome.
One further opportunity for spatial analysis that HLM affords is the ability to estimate cross-level interaction terms. Rather than exploring the effect of having transgressive neighboring zip codes, this approach explores how the impacts of self-expressive zip codes depend on the transgressiveness of their surrounding counties. Doing so effectively casts the spatial net wider, not only to the opportunities for transgressive experiences adjacent to a given zip code, but to those experiences available (on average) throughout its entire county. The “Interaction—Cross-level” model in table 8.8 reports results for such a model.
The “Interaction—Cross-level” results indicate that zip codes with self-expressive scenes exhibit no independent effects on change in 25- to 34-year-olds. Only when self-expressive zip codes are located in a transgressive county do such scenes exhibit a positive relationship with change in 25- to 34-year-olds. This result suggests that the effect of scenes, in this case transgression, may not be as localized as suggested initially by our adjacent zip codes approach. It suggests that county effects may operate in addition to the adjacent zip codes—as in the case of people who regularly commute across many neighborhoods in the same county for jobs, religious worship, shopping, and scenes.41
These combined results of direct measures, standard within–zip code interactions, neighbor interactions, and multilevel modeling suggest a few intuitive scenarios: (1) 25- to 34-year-olds are drawn to zip codes which are more transgressive, with the presence of a self-expressive scene effectively unimportant in such zip codes; (2) 25- to 34-year-olds did not move to self-expressive scenes unless such scenes were also transgressive or adjacent to a transgressive scene, with the latter being the more powerful dynamic; (3) 25- to 34-year-olds are drawn to self-expressive scenes located next to transgressive scenes. One can also imagine a combination of (1) and (3).
A look at some examples helps clarify and interpret these scenarios. Consider first some zip codes in Chicago’s North Side, near the Loop. These contain neighborhoods like Wicker Park, Bucktown, Logan Square, DePaul University, Lincoln Park, Old Town, the Gold Coast, and River North. All score high on self-expression.
But so too do zip codes in many small towns in Illinois. One example is Maeystown. About an hour outside of St. Louis, Maeystown is a historic German village that is home to Oktoberfest events, arts and crafts fairs, garden shows, and quaint B&Bs.42 Maeystown accordingly has a high self-expression score. Looking at self-expression alone effectively treats all these zip codes—Maeystown, Wicker Park, the Gold Coast—as the same.
But of course they are not the same, and a young person can tell the difference. We can too, once we look beyond self-expression in isolation. Take Maeystown again: it has a high self-expression score, a high local authenticity score, but a very low transgression score. This is a mix potentially attractive to a baby boomer, but less to the typical 25- to 34-year-old.
Now return to Chicago. The zip codes containing Wicker Park, Bucktown, and Lincoln Park have high self-expression scores. But they also have high transgression scores—they are examples of a strong within–zip code interaction between self-expression and transgression. To move here is to live among art galleries and festivals as well as nightclubs, tattoo parlors, and alternative music venues.
The within–zip code interaction highlights the experiential difference between places like Bucktown and places like Maeystown. The adjacency-interaction and multilevel models, however, bring out what is distinct about places like the Gold Coast and River North. Like Maeystown and Bucktown, they have high self-expression scores—but in comparison to Bucktown and the other North Side zip codes, the Gold Coast transgression score is relatively low, right around the national mean. In this respect the Gold Coast is closer to Maeystown.
But, of course, there is a big difference: location. The Gold Coast and River North are near many other zip codes with high transgression scores: south to the Loop, west to Fulton River Park, and north to the Near North Side. And the average transgression score in Cook County (in which the Gold Coast is located) is considerably higher than the average transgression score of Maeystown’s county (Monroe). To move to the Gold Coast/River North, that is, is to move to a place with interesting, but perhaps not very edgy, galleries, boutiques, and restaurants. But just down the street are Chicago’s more risqué haunts, which a Gold Coast resident can partake of—followed by a quick cab ride home to a safer, and quieter, apartment (with a doorman). This would be hard to do in Maeystown.
The example illustrates the general results. If we analyze self-expression scores alone, we would not be able to tell the difference between a place like Maeystown and those many Chicago zip codes that also have high self-expression scores. But once we see that Maeystown (and similar places) (a) do not also have a high transgression score and (b) are not near anywhere else with a high transgression score, then the difference between Maeystown-esque and (a) Wicker Park/Bucktown-style scenes and (b) the Gold Coast/River North–style scenes comes out loud and clear.
These contextual differences are all clear enough intuitively, but modeling them statistically is difficult. At the same time, more sophisticated statistics often mean losing sight of the lived experiences they are supposed to illuminate. This scenes approach to modeling context shows how to do both at once, and how similar techniques could be extended to other propositions about scenes and spatial context.
Catchment areas. Accounting for the scenes in adjacent zip codes is only one way to incorporate space into a scenes analysis. Another approach is to incorporate the concept of adjacency and catchment areas into the very construction of scenes measures. While adjacency only incorporates immediately surrounding zip codes, catchment areas expand this area to include various rings of zip codes extending outward from the target zip code. Importantly, we can construct performance scores for a target zip code based on the amenities within a certain radius of zip codes that surround it. Figure 8.3 illustrates the conceptual and spatial differences between neighboring scenes and amenity catchment areas.
Key implications of catchment areas arise from how we theorize their meaning: Catchment areas are a way to model the effective limit of residents’ vision when it comes to available options for consumption. That is, residents may not consider certain amenities to be available because they are “too far” away. Conversely, the amenities that constitute their local scene may extend well beyond their zip code of residence. To examine these possibilities, we recalculated performance scores using various radii.43 Table 8.9 reports the Pearson correlations between these self-expressive performance scores, suggesting that, despite producing similar values, choice of catchment area may add value to our analysis of scenes. For example, scores calculated using only amenities in the target zip code have a Pearson correlation of 0.343 with scores calculated using amenities from the target and immediately adjacent zip codes. Zip code and immediately surrounding scenes are similar but not identical.
To explore how different catchment areas may reveal further subtleties about the dynamics of self-expression, transgression, and young people, we reestimated our “Interaction—Neighbor” mixed model using performance scores based on four different radii (R = 0 to R = 3). In table 8.10, the R = 0 model is identical to the original “Interaction—Neighbor” model reported in table 8.8.
Catchment areas do matter. In particular, the R = 1 model exhibits stronger relationships between our outcome and self-expression in the target zip code, transgressive scenes in the neighboring zip codes, and their interaction. At the same time, the within–zip code interaction between self-expression and transgression shifts from positive in the original model to negative (and barely significant) in the R = 1 model, then insignificant in the R =2 and R = 3 models. The R = 2 and R = 3 models exhibit their own patterns, where the independent main effects of within–zip code self-expression and transgression become increasingly important, but the effect of neighboring zip codes remains on par with the original R = 0 model. That is, the impacts of local zip code scene qualities become even more pronounced the more we take account of their surrounding scenes. These results all add nuance to our initial finding: young people are drawn to transgressive scenes, not to self-expressive scenes in isolation. But when self-expressive scenes are joined by transgression nearby, they become more attractive to 25- to 34-year-olds. Proximity alters performance.
In sum, both the scenes within zip codes and nearby surrounding scenes make a difference for changes in the 25- to 34-year-old population. Scenes are localized phenomena that operate within and across zip code boundaries. While exploratory, these models collectively suggest that there are quantifiable boundaries to scenes and their effects, with catchment areas of different radii transforming these effects. Pursuing these lines of inquiry is an active area of ongoing research, which we illustrate here with this initial analysis to show how it can deepen our understanding of some of the processes identified in Scenescapes.
This “rings” method of analysis builds on a large geography literature on catchment areas, with empirical estimates for certain very specific items, but clearly there are large variations by amenity, activity, and subgroup (cowboys who drive 200 miles for dinner versus Manhattan residents who don’t drive, the reach of opera houses versus that of a coffee shop, etc.). One point here is simply that the specific tools and concepts of scenes analyses help identify coherent patterns of variations, which can be joined with new data as they emerge to better conceptualize and measure catchment areas. For example, adding scenes indicators can distinguish between the effect of restaurants (alone) or restaurants near a more transgressive scene, for separate age groups, and more.
These multiple approaches illustrate alternatives to modeling context. Results from OLS models, HLM, interactions, catchment areas, and adjacency analyses each add insights. Each makes different assumptions and employs different computational techniques.44 Despite these differences, we have been pleasantly surprised to discover that the simplest possible multivariate modeling we performed also happened to be quite robust. More involved methods produce more intricate pictures of the general relationships, which is exactly what we would expect.
Scenes as Elements of Multicausal Processes
This book has mostly treated scenes as independent variables and other variables as outcomes, such as economic growth, residential patterns, and partisan voting. We chose this path in part because of the extreme complexity involved in evaluating any claim about what causes a specific scene, and in part because we wanted to insert scenes into academic and policy discussions about the drivers of income and job growth, residential patterns, and partisan voting. This latter goal implied analyzing scenes as independent variables, together with classic social science variables thought to drive these outcomes (education, population, and the like) to minimize spuriousness.
In treating scenes as independent variables, Scenescapes may seem to resemble a sort of “cultural sociology,” where “cultural” variables drive “material” outcomes.45 This is misleading in that it is impossible to draw any hard-and-fast distinction between what is cultural and what is material. More important, we are not dogmatic culturists—reversing the equation is just as possible and plausible.
Being open to this reversal—considering scenes as either consequence or cause—implies at least two further types of questions. First are the chicken-and-egg questions: Which comes first, the scene, or other variables, like college graduates or jobs? Does one always precede the other? Or does sometimes one, and sometimes the other, come first? If so, when, where, and why? Questions more in line with much research and theorizing in the sociology of culture come next: If we treat scenes as results of other factors, what explains where and why one type of scene is stronger than another? Class, ethnicity, religion, education, or something else entirely? That is, how can we bring our concepts and data to bear on questions deriving from theories in which scenes are dependent rather than independent variables? We briefly address both questions.
The Problem of Causality
Causality is a sensitive issue in all our lives. Max Weber even suggested stressing elective affinities among variables rather than seeking causality. By the late twentieth century, social scientists had often grown more ambitious about causality. Yet most accept Weber’s fundamental ideas about multicausality, contingency, close interrelations among key variables, and especially changes in all of these over time. Few rally around the nineteenth-century flags of “culture” or “materiality” as the ultimate causes of social processes “in the final analysis.” Most social scientists have accepted the interdependence of many types of processes, the difficulties and arbitrariness of definitions, and the shifts over time and within different contexts, all of which are compounded by feedback effects.46 Thus feedback is one type of specific causal mechanism, which helps to clarify what may look impenetrable at a distance.
One way to think about feedback is in terms of the chicken-and-egg problem. Which comes first—the chicken or the egg? This labeling of the problem underlines the degree to which, over time, each is essential (i.e., changes in one feed back and affect the other), even if one seems primary at any specific point in time, and distinct causal dynamics can be identified at multiple stages.
Scenes analysis is linked with the rising salience in many scientific domains of this broader multicultural and multicausal perspective. It implies its own version of the chicken-and-egg question. An example of one of its cruder forms is, Do locations with more income attract more Starbucks, or do Starbucks attract persons with more income? This is the starkest formulation of the problem, as sometimes discussed in the urban literature.47
Our answer to “the Starbucks question” is that this is indeed a classic chicken-and-egg phenomenon. In any one short time period a statistical analysis may suggest that the driver is income but in a longer slice of time causality can reverse. How to interpret and measure these processes?
The migration and urban development literatures have slowly but increasingly recognized such interdependence over several decades. Clark and Ferguson’s (1983) review of the migration and development literatures found only a handful of papers by economists who had explored this chicken-and-egg question at the time. Richard Muth and Michael Greenwood were key figures. They used cross-sectional analysis of job growth and population growth for the entire United States, applying two-stage least squares to estimate the relative impact of population growth on jobs and vice versa. They found substantial impact both ways, but the strength of each was difficult to disentangle. Given the many other associated variables, they did not try to explore this further.
Yet Clark and Ferguson suggested that, contra the general theories of Muth and Greenwood, their results could be reinterpreted. Rather than trying to disentangle production and consumption as separate or unidirectional causal arrows, they could be seen as co-constituting each other. That is, people would choose a place to live as well as work simultaneously. Moreover, this overlap was increasing as work itself became more similar to a lifestyle decision about how to meaningfully spend one’s time, especially among self-employed persons and highly skilled professionals. This general orientation informs much of the analysis in Scenescapes: people select a location for many reasons, including amenities, and once they are there, they contribute to the character of the place, including its scenes.