Neighborhoods, Ethnicity and School Choice: Developing a Statistical Framework for Geodemographic Analysis

Abstract

Geodemographics as the “analysis of people by where they live” has origins in urban sociology and social mapping, and is experiencing a renaissance in applied spatial demography. However, some commentators have expressed reservations about the statistical limitations of common geodemographic practices, especially focusing on the potential internal heterogeneity of the geodemographic groupings, as well as the problem of clearly identifying predictor variables that might account for or explain the socioeconomic patterns revealed by geodemographic analyses. In this paper we argue that geodemographic typologies are structured methods for making sense of the spatial and socioeconomic patterns encoded within complex datasets such as national census data. By treating geodemographics as more a framework than a tool for analysis in its own right we are able to integrate it with the flexibility and statistical conventions offered by multilevel modeling. We demonstrate this with a case study of whether pupils from different types of neighborhood in Birmingham, England are more or less likely to attend their nearest state-funded secondary school and how that likelihood varies with the ethnic composition of the neighborhood. In so doing we build on previous research suggesting that ethnic segregation between schools is at least equal to that between neighborhoods in England and speculate in this regard on the consequences of current government plans to extend choice to parents within a schools market.

Introduction

In this paper we develop a multilevel statistical framework for geodemographic analysis with a case study of the travel-to-school distances of state-educated secondary school pupils in Birmingham, England. We build on research that has previously shown ethnic segregation in English and Welsh schools to be equal or greater than in the neighborhoods from which the pupils are drawn (Burgess and Wilson 2005; Johnston et al. 2004) by here considering the role ethnic concentration within neighborhoods has in determining whether pupils attend their local (nearest) school or not.

While our empirical findings are relevant to debates about the provision and nature of school choice, and about the function of schooling in promoting a multicultural and racially tolerant society in the U.K., the primary aim of this paper is to consider geodemographics as a method of spatial demographic analysis that is experiencing a renaissance in applied social research (Longley 2005). The paper begins with a brief introduction to geodemographics, focusing on some of its analytical weaknesses. We then provide a case study of how geodemographics can be integrated with multilevel analysis, modeling the geodemographic distribution of secondary school choices in Birmingham—specifically whether a pupil attends his or her nearest school or not—and linking that to geographies of the ethnic composition of neighborhoods.

Birmingham is sometimes described as England’s “second city” and had a population of 977,087 residents (390,792 households) recorded in the 2001 Census. It has been chosen as the study region because, as the local government website states, “the Census confirms Birmingham as a diverse City, with residents from a wide range of ethnic and religious backgrounds” (http://www.birmingham.gov.uk).

About Geodemographics

Geodemographics has been described as “the analysis of people by where they live” (Sleight 2004)—the assumption that where you are says something about who you are and what you do. The geodemographic industry produces classifications of (particularly residential) spaces, places or networks that the entities of interest—usually consumers or their households—inhabit or interact with, sorting the consumers into different groups or “types.” The classifications are sold to clients, including large retail chains and service industries, which then use them to classify their own customer records and from this, ideally, identify a core geodemographic type to which future promotional mailings, radio advertising, new store openings and the like can be targeted.

Geodemographics has a pedigree in socio-spatial research. Historical antecedents include Charles Booth’s Index Map of London (Booth 1902–1903) and the Chicago School of Urban Sociology of the 1920s and 1930s. Whereas Booth developed a multivariate classification of the 1891 U.K. Census data to create a generalized social index of London’s (then) registration districts, the Chicago School (see, in particular, Park et al. 1925) were developing the idea of “natural areas” within cities, conceived as “geographical units distinguished both by physical individuality and by the social-economic and cultural characteristics of the population” (Gittus 1964, p. 6). These ideas coalesced with the increasing availability of national census data and the computational ability to create multivariate summaries of these data by grouping together correlated variables using factor or principal components analysis, and by grouping alike places together using clustering techniques (for further details of this history and the foreshadowing of modern geodemographics in Social Area Analysis during the 1960s, see Batey and Brown 1995). Natural areas and conceptions of neighborhood became specified more formally as census zones (and, more recently, by postal geographies: ZIP or postcode units) or statistical aggregations thereof (Martin 1998). Commercial geodemographics emerged from the late 1970s with the launch of PRIZM by Claritas in the U.S. and ACORN by CACI in the U.K. By the turn of the millennium, Weiss (2000, p. 4) could argue that “cluster-based marketing has gone mainstream and is now used by corporate, nonprofit, and political groups alike to target their audiences,” citing as evidence the estimated $300 million spent annually by U.S. marketers alone. Currently there are geodemographic classifications of most of Western Europe, Northern America, Brazil, Peru, Australasia, South Africa, parts of Asia, and some of China, including Hong Kong (Harris et al. 2005).

The success of geodemographics has drawn critical attention. Some commentators provide social critique, focusing on the representational (Goss 1995), discriminatory (Burrows et al. 2005; Graham 2005) and intrusive (Monmonier 2002; Curry 1998) effects of geodemographic practices. Others outline statistical concerns that this paper heeds. The starting point is the accusation of ecological fallacy which is, in the sense it is made against geodemographics, the contentious assumption that members of a geodemographic group are sufficiently alike to be analyzed as one. The assumption can be questioned at two scales. First, the census or postcode areas that are assigned to and comprise a geodemographic cluster may not be especially similar—inevitably some clusters will be more uniform in regard to their data attributes than other. Second, even if all the areas were identical within a cluster, it does not follow that the population (individuals or households) within any one specific area need also be homogeneous.¹

Voas and Williamson (2001) suggest that apparent differences between geodemographic classes conceal a much greater diversity within the classes. If their finding generally is true then apparent geodemographic differences (where found) could be more an artifact of the classification process than a consequence of real-world socioeconomic cleavages. Their finding may not generally be true but it is hard to disprove. Geodemographic analyses usually calculate an index value summarizing the prevalence of a particular event (e.g., consumer behavior) within a cluster group, relative to its prevalence across all groups and standardized against a score of 100, which is the mean average. What is rarely provided is a measure of variation (variance) within each group and therefore of the statistical significance of differences between the groups.²

Furthermore, geodemographic analyses usually are conducted outside of more traditional statistical frameworks making it difficult to assess either the significance of apparent trends found in data or the importance of predictor variables that might explain them. This may not matter for the sorts of commercial and service planning applications to which geodemographic analysis is a strategic tool of proven value. However, geodemographics—benefiting from increased collaboration between commercial data vendors, governmental organizations and public sector researchers—is reentering areas of social research akin to those from which it originated (Ashby and Longley 2005; Williamson et al. 2005) and which include monitoring whether there is fair access to U.K. universities for all socioeconomic groups. These examples of applied data analysis are characteristically inductive, undertaking “knowledge discovery” by geodemographic classification of extensive microdatasets. While neither trivial to undertake nor unimportant (indeed they are arguably more relevant to public policy than the conceptual obfuscation apparent in much academic writing!) such research lacks focus on theory, model building, and hypothesis testing. In short, the spotlight is more on finding (geodemographic) patterns in data than on explaining them.

At its simplest, geodemographics is only a structured method of making sense of the spatial and socioeconomic patterns encoded within complex datasets. It does so by imposing a strict hierarchy on the data: in “classic” geodemographics, individuals reside in census or postal zones that are grouped into geodemographic clusters. Such hierarchies are efficiently handled by the wealth of analytical techniques developed under the rubric of multilevel modeling, often to measure differences in educational attainment between schools and pupils (Goldstein 2003). In those areas of research it is easy to imagine a regression relationship between a pupil’s performance in higher level examination and performance on previous exams, gender, and so forth. However, it is also likely that the relationship varies at a “higher level”—specifically, between schools when they have different resources, specialist interests, and pupil composition. While a separate regression relationship could be fitted to all the schools, to do so is neither parsimonious nor efficient. A better option is to pool all the pupil level data while at the same time acknowledging that pupils “nest” into schools, consequently estimating how the pupil level relationship also varies between schools and thence adjusting the standard errors associated with the regression coefficients to incorporate the nonindependence of pupils within schools.

The exact methods of multilevel estimation are beyond the scope of this paper (see instead Snijders and Bosker 1999).³ Nevertheless, incorporating geodemographics in these methodological frameworks permits new opportunities for a more statistically robust and model-based approach to social area analysis, and might provide more concrete evidence of the sorts of “neighborhood effects” that geodemographics is often said to reveal but does so ambiguously.

Modeling Geodemographics, Ethnicity, and Least Distance to School

From the landmark Education Act of 1870, the intervention of the state in funding and directing education in the U.K. has been premised on both the social and economic capital that accrue to society as a whole as it has the benefits of knowledge to the individual learner. Beyond the transmission and nurturing of subject-based facts, ideas, and practices, education is seen to serve a wider but politicized social role, exemplified by the resurgent language of citizenship and embodied by the statutory provision of citizenship classes to pupils aged 11–16 years in the U.K.

This discourse of citizenship intersects with visions of a multicultural society. In an address to the Hansard Society given on January 17, 2005, the Chief Inspector of Schools in England, David Bell, stated his view that:

citizenship education can be a positive force for good […] – promoting acceptance of different faiths and cultures as well as alternative lifestyles. Pupils can learn when to draw lines: how to say no to racial and religious intolerance; how to stand up to injustice; how to bring about change in policies that are unacceptable (Bell 2005, p. 18).

This especially is important if, whereas multicultural appreciation might be gleaned from the shared, day-to-day experiences of a class of pupils drawn from a mix of ethnic and cultural backgrounds, the actual practice is of various ethnocultural groups attending different schools from each other, preferring those where their particular group is more dominant. To quote a provocative (and contested—see The Observer 2005) speech by Trevor Phillips, Chair of the (British) Commission for Racial Equality in which he warned that Britain is “sleepwalking to segregation”:

[there are some] white communities so fixated by the belief that their every ill is caused by their Asian neighbors that they withdraw their children wholesale from local schools.

He later continues:

the passion being spent on arguments about whether we need more or fewer faith schools is, in my view, misspent. We really need to worry about whether we are heading for USA-style semi-voluntary segregation in the mainstream system (Phillips 2005).

Phillips cites empirical evidence suggesting ethnic segregation between English and Welsh schools exceeds that between residential localities (Burgess and Wilson 2005; Johnston et al. 2004, 2005). This increase may be a consequence of (constrained) parental choice in regards to which school their children attend—a choice that the government sets out to extend in its recent White Paper, subtitled “More choice for parents and pupils” (HM Government 2005). The White Paper outlines a quasi-market based system of schooling allowing successful schools to expand and take over failing ones; permits universities, charitable bodies, and businesses to form trusts to run “independent state schools” and set their own admissions criteria; and states that “the local authority must move from being a provider of education to being its local commissioner and the champion of parent choice.”

Although there has long been an element of affording preference to school allocations (by asking parents which school they would like to send their children to but without guaranteeing that choice), most English local education authorities have used allocation rules dominated by the aim of sending pupils to the nearest schools to their homes. However, at least since the 1988 Education Reform Act giving much greater power to parents in the selection of schools for their children, the rhetoric of choice has become increasingly loud in government policies for education (West et al. 1998). The apparent “marketization” of education therefore has been the focus of much research (see Dale 1997). One group of large-scale quantitative studies has argued that the introduction of greater parental choice has resulted in a fall in interschool segregation according to family poverty—as indexed by the number of students qualifying for free school meals—although these findings have been questioned on technical grounds (Taylor 2001; Taylor and Gorard 2001; Gorard et al. 2001; Goldstein and Noden 2003).

A paper by Parsons et al. (2000) showed considerable numbers of students attending comprehensive secondary schools other than those nearest to their homes. A similar situation is found in our study region, too. In 2002, in Birmingham, only 25% of pupils attended their nearest secondary school (estimated using Thiessen polygons to model the “catchments” of schools in a desktop GIS: see Longley et al. 2005). However, the aggregate figure conceals variation both by ethnicity and by a geodemographic classification of the census zones (Output Areas, OAs) containing the home addresses of pupils. For example, Table 1 shows that 41% of Bangladeshi pupils attended their nearest secondary school, while only 15% of pupils described as Black Caribbean did. Table 2 shows that 48% of pupils from areas described as “Terraced Blue Collar” attended their nearest school, compared with 14% of pupils from “Transient Communities” neighborhoods. Combining the ethnic and geodemographic information together in Table 3 it is shown that 54% of pupils described as of “Black Other” ethnicity and living in “Afro-Caribbean Communities” attend their nearest school whereas, intriguingly, only 13% of Black Caribbean pupils living in “Afro-Caribbean Communities” appear to.

Table 1 The proportion of Birmingham pupils attending their nearest school and average distance traveled to school, by ethnic category, and ranked by the proportion of the group attending their nearest school

Table 2 The proportion of Birmingham pupils attending their nearest school and average distance traveled to school, by geodemographic classification, and ranked by the proportion of the group attending their nearest school

Table 3 The 20 highest and 20 lowest ranked ethnic and geodemographic cross-tabulations in regard to the proportion of Birmingham pupils attending their nearest school

There are two primary sources of data presented in Table 3. The first is the 2001 Area Classification of U.K. Census OAs, freely available from National Statistics’ Neighborhood Statistics Service (NeSS⁴). OAs are the smallest area units for which census data are available and were built from clusters of contiguous and socially homogeneous (in terms of tenure of household and dwelling type) unit postcodes. There are 3,127 OAs in Birmingham, with an average count of 312 persons (125 households). The geodemographic classification of these and all other OAs in the United Kingdom was conducted by a team at the School of Geography, University of Leeds which produced, using k-means cluster analysis (see Berry and Linoff 1997), a nested hierarchy of 7 (super-groups), 21 (groups) and 52 (sub-groups). The clustering was based on a selection of 41 census variables to represent five domains: demographic structure; household composition; housing; socioeconomic; and employment (see Vickers et al. 2005 for further detail). Note that Tables 2 and 3 are at the Group level and include the names given to the clusters. These are available from the project website⁵ but not from NeSS where

as part of reviewing the classification against the National Statistics Code of Practice, the National Statistician decided that such names could be seen as “labelling” or stereotyping people resident in output areas within each cluster. Given the small population size of output areas, it was decided that this was not appropriate for a National Statistics product.

It is therefore important to emphasize that the names are only indicative and should be considered in the context of more detailed cluster summaries provided both at NeSS and at Leeds.⁶

The second dataset gives a residential unit postcode (ZIP+4 equivalent) and an ethnic code for each pupil attending a state-funded school in Birmingham. It is taken from the Pupil Level Annual School Census returns (PLASC), released for research by the Department for Education and Skills (DfES).⁷ The ethnicity of each student is recorded by staff at the pupil’s enrollment but is open to parental alteration.

While the PLASC data cover every pupil in a state-funded primary school (102,300 pupils) and secondary school (64,959) in Birmingham, the analysis presented here concentrates only on the second group. Furthermore, we have excluded from the analysis pupils for whom either their home postcode or ethnic coding is not known, who live in a census OA of unknown geodemographic type or who live near the edge of Birmingham’s metropolitan district and for whom their apparently closest school (within Birmingham) may not actually be so.⁸ Finally, of those pupils remaining, any living in OAs containing less than nine other pupils were removed from the analysis to avoid small number effects when calculating the proportion of pupils per OA of a particular ethnic group. As a result of the data cleaning our analyses are based on data for 53,274 pupils, representing 78 schools, 2,189 OAs and 19 geodemographic groups.

Tables 1–3 suggest some interesting differences in the distances traveled to school by pupils of different geodemographic and ethnic types but some caution is required. First, they are based only on the straight-line distances between home and nearest/actual school attended and not the actual distance traveled, which will be more circuitous.⁹ It would be possible to estimate true distances using road network analysis, although to do so generally presumes that pupils travel by private automobile, a presumption that is almost certainly false. (Pooley et al. 2005 cite Department for Transport data published in 2001 showing that 43% of 11–16 years old in Britain walk to school, 32% travel by bus, 19% take a car, and 2% cycle.)

Second, the distances traveled are not solely due to choice. While parents can express a preference as to which school their child attends, ultimately each school has only a certain number of places available and, if oversubscribed, will operate selection criteria (for example, offering places to siblings). Faith schools—those supported by religious groups—may also adopt selective practices as, of course, do single-gender schools. The admissions criteria for each (nonprivate) secondary school in Birmingham are documented at http://www.bgfl.org/services/admissions.

With particular regard to Table 3 and our earlier discussion of the limitations of conventional geodemographic analysis, are the differences between the geodemographic and ethnic groups actually significant or simply “due to chance”? To answer the question the analysis has been transplanted into the multilevel framework shown in Fig. 1. This is a logit model that regresses the binary response (either pupils do attend their nearest school or they do not) against a series of dummy variables—one for each of the eight ethnic categories shown, with the category of “Chinese” being used as the comparator (i.e., it is present in the dataset but has no dummy variable associated with it). Note that the response variable is actually whether pupils do not attend their nearest school (coded 1) and that this is a simple, hierarchical model with three levels: the pupils (subscript i) live in census OAs (j) that are assigned to geodemographic clusters (k). The structure of the model avoids assuming the pupils are independent in geodemographic terms. They are not, because pupils living in the same OA as each other necessarily belong to the same geodemographic group, together with pupils from other OAs.

Fig. 1

MLwiN screenshot showing the structure of multilevel logit Model 1. Note: This model estimates the likelihood a pupil in Birmingham does not attend their nearest school, with three levels (pupil, i; census zone, j and geodemographic group, k), eight ethnicity classes (dummy variables) and measuring variance at the geodemographic level

As with any regression model, we are interested in the coefficients and measures of standard error assigned to each of the predictor variables. However, unlike a standard model, the intercept term (β ₀) is permitted to vary at the most aggregate level of the hierarchy—the geodemographic classes. In short (in Fig. 1) we are interested in the variance of v _0k which estimates how much the likelihood of a pupil attending a nearest school varies by neighborhood type, having controlled for the differing likelihood between ethnic groups. The model is fitted using version 2.02 of MLwiN and a Markov Chain Monte Carlo (MCMC) simulation procedure with a burn-in length of 5000 and a monitoring chain of 50,000 (see Browne 2004; Rasbash et al. 2004; Snijders and Bosker 1999 for further details).

Reassuringly, the multilevel analysis—summarized by Fig. 2 (and again, later, in Table 4)—confirms the previous results in Tables 1 and 2. With regard to the ethnic component of the model (and remembering that we are now focusing on the binary opposite to Tables 1 and 2—the likelihood that pupils do not attend their nearest secondary school, relative to the Chinese group), the regression coefficients have the same rank ordering as in Table 1, with the exceptions of the “Black Other” and Pakistani groups for which the positions are reversed (but with no statistical significance). The Black Caribbean group remains as the least likely to attend their nearest secondary school; the Bangladeshi group remains as the most likely.

Fig. 2

MLwiN screenshot showing the coefficients fitted to Model 1 (see text for explanation and discussion)

With regard to the geodemographic component, for which we are interested in the variance of the random intercept v _0k , significant difference between the groups is found at the 95% confidence level. Figure 3 shows the rank order of v _0k for the geodemographic groups. There is broad agreement with Table 2 with, for example at the lower end of the rank ordering, pupils from “Terraced Blue Collar” and “Public Housing” least likely to not attend their nearest school (i.e., they are most likely to attend their nearest school). At the other end, there may seem to be disagreement with Table 2—the “Settled in the City” group appears more likely to not attend their nearest school than the “Transient Communities” group. This is deceptive, however, insofar as we need also to consider the 95% confidence intervals that are shown above and below the mean of v _0k. Looking at these, there is no significant difference between the estimated likelihood of a “Settled in the City” or “Transient Communities” pupil not attending his/her nearest secondary school, having controlled for ethnicity effects; but, there is a significant difference between a “Transient Communities” and “Terraced Blue Collar” pupil, for example.

Fig. 3

Measuring residuals at the neighborhood level to identify the geodemographic clusters where pupils are most likely not to attend their nearest school, having controlled for ethnicity effects. The 95% confidence intervals are also shown. Note: The geodemographic ranks from left to right are: 1—2b (Settled in the City); 2—7b (Afro-Caribbean Communities); 3—6b (Least Divergent); 4—2a (Transient Communities); 5—4d (Thriving Suburbs); 6—4a (Prospering Younger Families); 7—7a (Asian Communities); 8—3c (Accessible Countryside); 9—4b (Prospering Older Families); 10—6a (Settled Households); 11—4c (Prospering Semis); 12—6c (Young Families in Terraced Homes); 13—6d (Aspiring Households); 14—5a (Senior Communities); 15—1b (Younger Blue Collar); 16—1c (Older Blue Collar); 17—5b (Older Workers); 18—5c (Public Housing); and 19—1a (Terraced Blue Collar)

Modeling Ethnic Exposure as an Indicator of School Choice

An important component of the segregation debate for British schools is whether pupils for whom their ethnic group has relatively low prevalence within their residential locality consequently attend less local schools but ones where their ethnic group is more dominant (therefore contributing to a process of increased segregation from neighborhoods to schools).¹⁰ Formally, in regard to our multilevel model structure, we ask a slightly different question: does the likelihood that a pupil of a particular ethnic category attends his or her least distance secondary school decrease as the proportion of pupils in their census OA not of the same ethnic category increases?

The proportion is a measure of the pupil’s exposure to ethnicities other than his/her own living in the same census neighborhood; reciprocally, it is also a measure of the level of ethnic concentration of the pupil’s ethnic group within the neighborhood (since: proportion not of the same ethnicity as the pupil + the proportion who are = all pupils in the neighborhood). It is incorporated into the model by multiplying the dummy variable for each ethnic category by the proportion of pupils in the census OA not of that ethnicity. The result is a series of interaction terms that conflate a pupil-level variable (ethnicity) with a census OA-level variable (proportion). While such a procedure would normally raise concerns about spatial autocorrelation and underestimation of the standard errors of the coefficients, the multilevel model structure ameliorates these.

The results of the model are summarized in Table 4, as Model 2. Note that we have now measured variance not only at the geodemographic level but also at the school and OA levels. The model structure has four levels (pupils, schools, OAs, and geodemographic groups); these are no longer hierarchical but cross-classified (since the schools pupils attend are not necessarily in the OAs they reside in). Also shown in Table 4 are the results of fitting a third model to the data. The basis of this model (Model 3) is the same as Model 2 but now includes additional exploratory variables not derived solely on the basis of ethnicity. These include whether the pupil receives a free school meal (a measure of economic disadvantage), the straight-line distance from home to the nearest school, and some attributes of the school attended: whether it is all male, all female, has a selected intake, is a faith school, number of pupils, and average GCSE (General Certificate of Secondary Education) results (a national qualification obtained by most students when they are aged about 16).

Table 4 Coefficients obtained for multilevel models 1–4 (refer to text for detail and discussion)

Included in Table 4 is the Deviance Information Criterion (DIC), which is a generalization of the Akaike Information Criterion (AIC).¹¹ The DIC diagnostic is a composite measure of the fit and complexity of a particular model and can be used to choose between models. The lower the DIC value the better. In this way, both Models 2 and 3 offer improvement over Model 1. Model 2 is marginally the better because Model 3 (which is not parsimonious) is penalized by the greater number of insignificant variables in it. Unsurprisingly, given the inclusion of school attributes, Model 3 has decreased variance at the school level. Overall, however, there is little evidence that the attributes of the schools are especially significant in the model, other than where the school is selective—particularly all-male schools to which pupils necessarily travel further to attend.

Looking at the fixed interaction terms in Models 3 or 4, these are found to be significant for the White and “Black Other” groups. Recall that these terms show the apparent effect that increasing exposure to other ethnic groups in the neighborhood has on the pupil’s own likelihood of not attending the nearest school, having controlled for some of the attributes of schools. Adding the coefficients for these terms to those obtained for the dummy (ethnicity) variables and the intercept (ignoring variance around the intercept at the geodemographic level for the time being) predicts the likelihood that a White or “Black Other” pupil attends his/her nearest school; these likelihoods can then be plotted against the corresponding level of exposure to other ethnic groups, as in Fig. 4.

Fig. 4

Likelihood that White or “Black Other” pupils do not attend their nearest school given their exposure to other ethnic groups in the neighborhood. Note: These are the fixed effects from Model 3 of the interaction terms for White and “Black Other” pupils, having controlled for other pupil and school attributes

White pupils form the majority of all pupils in 1,452 of the 2,189 census OAs in our study region (66%), and constitute the largest ethnic group in a further 124 (6%). In contrast, the “Black Other” group never dominates. Consequently, whereas Fig. 4 indicates a clear linear trend for White pupils (always more likely not to attend their nearest school than to do so but with that likelihood increasing with increasing exposure to other ethnic groups across the range 0 to 1), for “Black Other” groups the range is more limited (they always are exposed to other ethnic groups); only when they constitute a proportion of 0.1 or less of the pupils in an OA are they less likely to attend their nearest school.

Turning to the geodemographic level of Model 3 (in Table 4), the variance between groups is approximately one fifth of that between OAs or between schools but remains significant. Figure 5 shows that it is pupils from the “Asian Communities” and “Aspiring Households” neighborhoods that are more likely not to attend their nearest school than the fixed parameters of Model 4 otherwise predict.

Fig. 5

Residual variation at the neighborhood level. Above each rank position are the mean effect and 95% confidence interval for the geodemographic groups having also allowed for variance at the census OA scale (Model 3). To the right of each (and no longer in rank order) are shown the equivalent values obtained if variance at the OA scale is not modeled. Note: The geodemographic ranks from left to right are: 1—7a (Asian Communities); 2—6d (Aspiring Households); 3—2a (Transient Communities); 4—1c (Older Blue Collar); 5—1a (Terraced Blue Collar); 6—4d (Thriving Suburbs); 7—6a (Settled Households); 8—7b (Afro-Caribbean Communities); 9—5b (Older Workers); 10—5a (Senior Communities); 11—4b (Prospering Older Families); 12—5c (Public Housing); 13—2b (Settled in the City); 14—4c (Prospering Semis); 15—1b (Younger Blue Collar); 16—6c (Young Families in Terraced Homes); 17—3c (Accessible Countryside); 18—6b (Least Divergent); 19—4a (Prospering Younger Families)

Figure 5 also shows the effect, at the geodemographic level, of not modeling variance at the OA level. This is similar to geodemographic applications that look only at the differences between geodemographic groups but ignore heterogeneity within the groups. Generally the rank ordering does not change if variance within the geodemographic clusters is ignored; the tendency of pupils in “Asian Communities” and “Aspiring Households” neighborhoods to not attend their nearest schools is still shown to be underestimated by the fixed parameters of the model. But note that the difference between these two neighborhood groups and the rest is underplayed by the more traditional geodemographic approach. Conversely, we obtain a better model of geodemographic differences if we first accept that there is variance at the OA level and see what remains over and above it. That the model is better is reflected in the DIC diagnostic: 37211 for Model 3 (in Table 4), rising to 47272 if variance at the OA level is ignored. However, we should not conclude that geodemographics is a conservative and therefore “safe” form of identifying differences between neighborhood groups: looking at Fig. 5 it is possible to identify occasions when a geodemographic approach is likely to identify differences between neighborhoods that are not, in fact, statistically significant (compare ranks 3 and 15 modeled with and without OA variance, for example).

Measuring “Neighborhood Effects”

Geodemographic analyses are sometimes presented as evidence of neighborhood effects (for example, Webber and Longley 2003), although not always with a clear explanation of how these are defined or caused. Dietz (2002), drawing on the work of Manski (1993, 2000), identifies four types of neighborhood effect. A first (actually, Manski’s second) is a correlated effect—that individuals in a neighborhood tend to have similar characteristics. It is this that geodemographics most obviously measures. However (as Dietz carefully notes) there are numerous social, economic, cultural, and other processes that lead certain “types” of people to be living in particular places. These processes of sifting and sorting may come to structure the neighborhood but are exogenous to it. To describe the resulting correlations as neighborhood effects therefore gives a misleading impression of causation (an observation that Smith and Easterlow 2005 also make in relationship to health geographies: see below). That is not to say that correlation effects are never due to neighborhood-level inputs. Examples of where they are could include the consequences of a spatially targeted urban renewal program or the effects of poor design and architecture on the lives of residents of a housing estate. It is just to say that correlation effects are not in themselves evidence of neighborhood effects.

The correlation effects described above can all be described in terms of a functional relationship: where X then Y, with X being either exogenous or endogenous to the neighborhood where it leads to Y. It is only when X is endogenous and therefore contained in the neighborhood that the relationship with Y might be described as a neighborhood effect but even then the criterion seems insufficient. Still, there is a functional relationship between Y and X, expressing what is sometime described (by O’Sullivan and Unwin 2003, for example) as a first order relationship, or as spatial heterogeneity. When “a lot of X” leads to “a lot of Y” in a place then there is certainly a geography that may be interesting to explain and, in any case, needs the application of a spatially relevant modeling technique such as multilevel modeling to handle the nonindependence of the observations and/or the residuals in or of the model. Yet, a more persuasive conceptualization of neighborhood effect is the second order relationship where the amount of Y in the locality is actually significantly more (or less) than that predicted by X alone (especially when X is not a single variable but a multivariate matrix), indicating spatial dependence.

The second order relationship suggests the possibility of spatial and social interaction effects within the neighborhood. These might be catalysts, where the change in aggregate neighborhood behavior is due (at least in part) to a change in one or more individual’s behavior (the individual case affects the aggregate). Alternatively they could be reactions, where the actions of the individual are a response to the characteristics of their neighbors (the aggregate affects the individual). These catalysts and reactions are, respectively, the endogenous and exogenous effects attributed to Manski by Dietz (2002) but these are terms that we avoid to prevent confusing effects that are either endogenous or exogenous to the individual with those that are the same to the neighborhood—that is, to retain a sense of scale. The reactive effects are also sometimes called compositional effects but again the terminology risks confusion unless they are understood to be place-specific and locally contingent reactions to the neighborhood’s composition (if it is more generally true that a particular composition, X, leads to behavior Y, then a first order relationship is being described). Finally, we are cautious about the language of “contextual effects” because the word context could refer to: the composition of the neighborhood as the setting for individual or group behavior; the neighborhood’s relationship to other nearby neighborhoods (and how they impact upon each other—the fourth neighborhood effect identified by Dietz); or to regional or national effects impacting upon the neighborhood and its population.

Traditional methods of geodemographic analysis that examine the prevalence of a particular consumer characteristic or social phenomena in any one cluster group relative to all others (and index accordingly) cannot disentangle these various effects, most particularly because they cannot easily separate first order relationships from second order ones. If, for example, there is a relationship between Y and X, and more of X is present in geodemographic cluster k than any other, then it is not surprising to find more of Y in k too. It may be important to know that there is a lot of Y in k but it is not evidence of a neighborhood effect. Worse, traditional geodemographic practices obscure the first order relationship. Because the cluster groups are not internally homogeneous it is never entirely clear what it is about the socioeconomic and demographic composition of k that causes, helps explain, or is most directly associated with Y. It follows that it is difficult to determine that there is more of Y in k than might have been expected on the evidence of X.

Does our multilevel framework offer improvement? Model 3 certainly suggests geodemographic differences between the school choices of pupils, having first controlled for pupil and school level attributes, and having established the relationship that White and “Black Other” pupils tend to travel further to schools as their exposure to ethnic groups other than their own increases within their neighborhood. We know that there is significantly greater likelihood not to attend the nearest secondary school for pupils living in “Asian Communities” neighborhoods (which is interesting, given Trevor Phillips’s comments presented earlier) but we have no clear idea of whether this is true of all pupils in these neighborhoods or only some. One reason it may be true of only some pupils is that, despite its name, the “Asian Communities” group is actually ethnically diverse in Birmingham: 9,311 of the pupils in these neighborhoods are Pakistani (35%); 7,914 are White (30%); 2,630 Indian (10%); 2,192 are recorded as “Other” (8%); 2,033 are Bangladeshi (8%); 1,946 Black Caribbean (7%); 297 Black African (1%); 76 Chinese; and 73 “Black Other.”

In fact, our final model—Model 4 in Table 4, above—suggests that the response to increasing exposure to ethnic groups other than their own for pupils living in “Asian Communities” neighborhoods does differ from the response for pupils living in the other geodemographic groups. For all Black Caribbean pupils, they are more likely to attend their nearest secondary school as exposure to other ethnic groups increases but that trend is more the case for pupils who do not live in “Asian Communities” neighborhoods than those that do. For White pupils it seems to make no difference: they are increasingly likely to not attend their nearest secondary school as exposure to other ethnic groups increases, regardless of whether the pupil lives in an “Asian Communities” neighborhood or not. For Indian pupils living outside of “Asian Communities” neighborhoods, exposure to other ethnic group seems to make no difference but, for those within “Asian Communities” neighborhoods, as exposure increases so does the likelihood of not attending their nearest school.

It is notable, however, that Model 4 has a marginally worse DIC score than Models 2 and 3 (because it is more complex), that the unexplained variance at the geodemographic level has not changed significantly, and that it is still the likelihood that pupils in “Asian Communities” neighborhoods do not attend their nearest secondary school that is most unpredicted by the fixed parameters of Model 4. Putting this together, Fig. 6 shows the likelihood that Indian pupils will not attend their nearest school given whether they live in an “Asian Communities” neighborhood or not and given their exposure to other ethnic groups in their neighborhood. Adding in the unexplained geodemographic variance increases the probability that Indian pupils living in “Asian Communities” neighborhoods will not attend their nearest secondary school by over 0.2 (about a third more than the predicted likelihood when geodemographic variance is excluded).

Fig. 6

Predicted likelihood (from Model 4) that Indian pupils will not attend their nearest secondary school: (a) Fixed effects only (“baseline” probability that an Indian pupil will not attend his/her nearest school + amount probability increases given proportion of pupils in the neighborhood who are not Indian). (b) The same fixed effects plus the variation at the geodemographic level. In both (a) and (b) the upper line is for Indian pupils living in “Asian Communities” while the lower line is for pupils in all other neighborhood groups

Have we evidence of a neighborhood effect? Perhaps. Smith and Easterlow (2005), in a critique of multilevel analysis used to measure health inequalities, argue that it is never possible to prove a neighborhood effect because it is never known for certain that an additional predictor variable might “explain away” the apparent neighborhood effect (that is, reduce an apparently second order relationship to a first order one). Logically they are correct, although their observation can be generalized: it is never possible to know for sure in any piece of work (quantitative or qualitative) that you are not missing that extra piece of information, data, variable, anecdote, experience, memory, writing, etc., that would change the interpretation or explanation of the phenomenon being studied.

Against that rather self-defeating logic we could reach for a number of philosophical perspectives including critical realism (Danermark et al. 2002), pragmatism (Menand 1997), and inference to the best explanation (Lipton 2004). Here, however, we are satisfied to concede that our analyses are not proof of a neighborhood effect. The reasons are threefold. First, we have not explicitly modeled the spatial configuration of schools around each pupil’s residential address. Some places and neighborhood types will have more choice than others, given constraints such as distance and school admissions policies. Second, we have not included any census-based indicators that might explain some of the variance we know to exist at the OA level in our models.

Finally, we have not directly measured what is likely to be a key determinant of school choice in Birmingham if—as we suggest—there is an ethnic component to the choice. We have not included the ethnic composition of schools at the time the choice is made. This apparently simple observation might imply a relatively minor change to our models (an additional predictor variable) but is deceptive. In fact, what we now need to model is a process, with longitudinal data. The task is to examine the composition of schools at time t ₀ and infer their influence on the patterns of travel at t ₁. Each of these spatiotemporal elements can be developed using the PLASC data within a multilevel framework and is an area of ongoing research. Nonetheless, there is a caveat. While we could go on adding multiple variables at multiple levels of analysis, to do so risks the same accusations of naive empiricism that have been raised against the more inductive geodemographic practices. What we actually advocate, therefore, is a more deductive approach, grounded in economic and social theory to inform the selection of the variables and levels of the model to be tested.

Conclusion

In this paper we have presented a critique of geodemographics as a method of spatial demographic analysis for social research. Our primary concern has been that the sorts of social patterns and trends that are discerned by conventional geodemographic analyses may not be secure in statistical terms and provide limited robust evidence of the sorts of neighborhood effects that geodemographics is sometimes said to reveal.

Despite this, our comments should not be read as an unconditional dismissal of geodemographic practices or their rising popularity in commerce and public service delivery. We accept entirely that the application of geodemographic typologies to guide resource allocation or to target prospective customers is of proven value to businesses and public sector institutions. We also understand that there is merit in a relatively simple and comprehensible method of exploratory data analysis that can help to identify economic, demographic, and cultural cleavages across the socio-spatial landscape, and provide a start to explaining why those cleavages exist and/or how they can be managed. Our concern is not that geodemographics is used as a “first pass” method of data exploration or inductive knowledge generation but that its use can shift into areas of prediction, explanation, or social monitoring that are rather less defensible, primarily because the internal heterogeneity of the cluster groupings makes the reasons why geodemographic patterns are found in datasets hard to discover (and therefore to manage).

The solution, we suggest, is to regard geodemographic typologies as less an analytical tool and more a framework providing structure for analysis. Consider Ashby and Longley’s (2005) study of how geodemographic analysis (specifically the Mosaic U.K. classification) can be used—successfully—to guide resource allocation for local policing (based on a study region of North and East Devon located in South West England). They show that total crime incidents are three times more likely to occur in “Council Flats” neighborhoods than any other—a worrying statistic that undoubtedly is relevant to policing and that implies a link between local authority housing tenure and the likelihood of being a victim of crime. But, if the link is true, it cannot be proven by the geodemographic analysis, because the “Council Flats” group actually contains a mixture of tenures. And if it is not true, the geodemographic analysis offers little alternative explanation as to what other factors are associated with high crime rates.¹²

If, instead, the geodemographic classification provided the structure for a multilevel analysis that included, among others, census measures of housing tenure, then not only might the link be verified (or otherwise), it would also be possible to: (a) identify other neighborhoods not of the “Council Flats” group that also have high crime rates but not obviously so because they are “averaged away” at the more aggregate geodemographic scale; and (b) identify neighborhoods where the crime rate is significantly higher or lower than that expected based on tenure (or other predictor variables) and whether these are characteristically of particular geodemographic types. Both (a) and (b) have implications for resource allocation for local policing, and for crime prevention and management.

In this paper we have adopted a statistical geodemographic framework to examine whether pupils of differing ethnic and neighborhood groups appear to exercise school choice differently (or are constrained to do so) insofar as this choice is expressed by them attending their nearest secondary school or not. A related issue, but not one we explicitly address, is whether pupils are attending schools that are more representative of their ethnic group, therefore increasing segregation at the school relative to the neighborhood level. There is evidence that they do. For example, White pupils who live in neighborhoods where their ethnic group constitutes 20% or less of all pupils and who do not attend their nearest school are, on average, in schools where the increase in the percentage of the pupils in the school vis-à-vis the percentage in the neighborhoods who are white is 15% (there is no increase for those who do attend their nearest school). Of the same White pupils, for those living in “Asian Communities” neighborhoods the difference is 35% (11% for those who do attend their nearest school).

Perhaps implicit to our analyses is the conception of schools as being of a homogeneous type, implying that it is usually rational to attend the nearest school (and not to do so is a reaction to the ethnic composition of neighborhoods). Such a conception of education simplifies the analytical framework but is not entirely satisfactory, especially given government policy encouraging schools to specialize in particular subject areas or vocations. That said, Renzull and Evans (2005) draw on theories of racial composition to consider the role of school choice and of charter schools, in bolstering “a return to school segregation” within the United States. A charter school is “a nonsectarian public school of choice that operates with freedom from many of the regulations that apply to traditional public schools [...] Charter schools are public schools of choice, meaning teachers and students choose them” (http://www.uscharterschools.org). They are also the fastest growing educational innovation in the United States. Analyzing national datasets collected by the National Center of Education Statistics, Renzull and Evans (2005, p. 413) come to a stark conclusion: “charter schools provide a public school option for white flight without the drawbacks of residential mobility.”

In the Education White Paper, the Prime Minister expresses his view that:

while parents can express a choice of school, there are not yet enough good schools in urban areas; such restrictions are greatest for poor and middle class families who cannot afford to opt for private education or to live next to a good school, if they are dissatisfied with what the state offers (HM Government 2005, p. 4).

He may be right; nevertheless, the White Paper was contentious for many, including the ruling party’s own MPs—even the Deputy Prime Minister was reported as having expressed reservations! From our perspective, we can understand the social reasons for wanting to extend the rights to free school transport to children from poorer families to a selection of nearest schools, for example. However, an associated risk is that increased choice within the education system could further the processes of ethnic segregation that have raised much concern within Britain.