Ethnic Residential Segregation: A Multilevel, Multigroup, Multiscale Approach Exemplified by London in 2011

The aim is to develop an explicit model-based approach in which segregation is summarized by a variance term around a mean. Put simply, if there is no variance beyond chance, each area will have the same underlying ethnic experience. Systematic underlying differences in the relative rates between places are shown by a large variance term, beyond what we could have observed from chance alone. As such, this method builds on the recent analysis of school segregation (Leckie and Goldstein 2015; Leckie et al. 2012) and indeed on the long-forgotten paper of Kish (1954), who examined residential differentiation. The new development here is that whereas Leckie and colleagues and Kish used a binomial model in which the outcome is a proportion, we use a Poisson model in which the outcome is a count of the number of people. The Poisson is highly suitable for low absolute counts because many OAs have small numbers in many of the ethnic categories. Moreover, the Poisson formulation allows the comparison of each and every ethnicity, and no particular ethnicity has to be chosen as the reference category as in the binomial logit model; rather, the comparison is with a theoretically even distribution. Our methodology also builds on the pioneering study by Moellering and Tobler (1972), who decomposed variation for predetermined spatial aggregates but only for continuously measured, ratio-scale data. In contrast, we have rates based on a varying numerator and denominator. The resultant model allows for the simultaneous analysis of multiple ethnic groups at multiple scales in an explicit modeling framework that allows us to put confidence intervals (Bayesian credible intervals) around the estimates. This overall model also allows us to see the extent to which members of each ethnic group co-locate with other groups at each scale.

Although one may presume that a total population—a census—does not require an inferential modeling approach (as strongly insisted by Gorard 2007), this is definitely not the case here. An important characteristic of these fine-grained data is their uneven absolute size. Although the mean count is some 22 people of a particular ethnicity in an OA, it ranges from 0 to 736 with a median of just 5 and a lower quartile of just 1, indicating the extent of the clustering of different ethnic groups, as shown in descriptive analyses of London (Johnston et al. 2014, 2015). The approximate standard error of the log of a relative rate based on an observed to expected ratio is inversely proportional to the square root of the observed count (Breslow and Day 1987: equation 2.9). Consequently, we can anticipate a great deal of stochastic variation and do not want to be misled by this natural variation. Indeed, common descriptive indices such as the D Dissimilarity Index are known to suffer from the upward bias of the null—showing systematic segregation when there is none—when small counts are analyzed (Allen et al. 2015; Carrington and Troske 1997).⁶ In contrast, we consider the observed counts as an outcome of a stochastic process that could produce different results under the same circumstances. It is this underlying process, or relative risk, that is of interest, and the actual observed values give only an imprecise estimate of this. We need a method that estimates the differences in the “true” ethnicity rates shorn of random noise.

To introduce the model, we begin with a two-level model in which individuals are nested within OAs, and we do so for just two ethnic groups: Black Africans and Black Caribbeans. We will later extend this to multiple scales and multiple groups. This basic model is specified as follows:

where O _ij is the long stacked vector of the observed count for individuals i in OAs j. This vector has two observations: the count of Black Africans and Caribbeans for each and every OA. The other observed variables are the expected counts (E _ij ) for each ethnic group if their numbers were distributed evenly according to the total population size of the OA. In addition, two separately coded dummy variables (African _ij ; Caribbean _ij ) identify which stacked count represents which ethnicity.

The counts are modeled in a Poisson regression model,⁷ where the observed counts are seen as coming from a Poisson distribution with a mean rate of occurrence given by π _ij . However, it is the natural log of the underlying rate that is modeled, and this is achieved by the use of an offset, which is the log of the expected count with a coefficient constrained to 1 (McCullagh and Nelder 1989). Thus, the expected value is effectively treated as a nuisance and allows us to model the underlying relative rate with the response simply being the log of the observed counts rates (difference in logs being equivalent to division in raw values). Moreover, using the log also ensures that we cannot estimate a negative relative rate.

The two intercepts in the model are (1) β₁, which gives the log average rate across all OAs for Black Africans; and (2) the overall log rate for the Black Caribbean population, β₂. We anticipate that both of these estimates, when exponentiated, will give the all-London rate for the mean area as 1.⁸ There are allowed-to-vary differences for each OA for each ethnicity (u _1j ; u _2j ); a positive value indicates a log rate that is higher than expected given an even distribution or equivalently a relative concentration of that ethnic group. These differences are assumed to come from a joint normal distribution, such that the variance σ _u1 ² and σ _u2 ² give the differences between OAs for Black Africans and Caribbeans, respectively. These are our measures of comparative segregation. A useful term is the higher-level OA covariance term, o _u12; when standardized by the product of the square root of the variances, this gives the correlation between the differences for the two groups. A negative value indicates that each group is antagonistically located relative to the other; a positive value represents a tendency to co-locate.

At the lowest individual level, the variances are constrained to be equal to the underlying rate for each ethnic group as befits an exact Poisson distribution. Consequently, the model separates the two sources of variation: the variation due to “true” between-OA variation, and that due to stochastic Poisson variability. Equivalently, the lower level of the model is used to model the natural variation of a Poisson variable, whereas the higher level is used to model the extra-Poisson variation of the “true” rates to give a measure of comparative segregation.

The apparent problem, of course, is that we do not have individual data but only aggregate counts for OAs because of confidentiality requirements. However, we can use the device of a pseudo-level in which the OAs are both the is and js in the model. Consequently, there is exactly the same set of units at Level 1 and Level 2, and each Level 2 unit has exactly one Level 1 unit. This views the aggregate counts at Level 2 as consisting of replicated responses for individuals at Level 1. This device allows for extra Poisson variation in the same manner as Browne et al. (2005) achieved for overdispersed binomial multilevel models.

Owen and Jones (2015) discussed a number of ways of turning these variances, which are on a log scale, into a more readily interpretable form. They found that the most appealing is the median rate ratio (MRR) given that this facilitates comparisons between standardized rates. The MRR can be conceptualized as the increased rate (on average; hence, the median) if one compares the rates of two MSOAs chosen at random from the distribution with the estimated variance. If there is no segregation, then the MRR would be 1; a value of 2 would indicate substantial segregation with the randomly chosen area, with the higher rate having twice the rate of the lower area.⁹ The calculation of the MRR is a simple transformation of the variance, and the same operation could be used to derive the 95 % credible intervals (CIs) around each MRR value for significance testing purposes.¹⁰

The normality assumption of the higher-level differences is obviously a key assumption for the validity of the variance in summarizing the differences in the relative rates. Although inference in multilevel models is typically robust to moderate departures from normality (McCulloch and Neuhaus 2011), severe skewness or outliers can pose problems. Whether these are present can be informally assessed with a normal probability plot. In practice, we have found that the normality assumption is generally met, no doubt due to using the log of the underlying rate. Indeed, in their study of London schools, Leckie and Goldstein (2015) found that parameter estimates obtained through MCMC procedures were not unduly sensitive to the inclusion/exclusion of religiously exclusive outlying schools.¹¹

The model can readily be extended to work at more scales, and this is the specification for two ethnicities and the three scales of the individual, the OA, and the MSOA that form a strict three-level hierarchy.

where O _ijk is the long stacked vector of the observed counts for both Black Africans and Caribbeans in cell (type of person) i for OA j in MSOA k. Of the two intercepts, β₁ gives the log average rate across all MSOAs and OAs for Africans, and the London-wide log rate for Caribbeans is β₂. There are ethnic-specific differences at the MSOA level (v _1k ; v _2k ) and for OAs within MSOAs (u _1k ; v _1k ). These differences at each of the higher levels are assumed to come from a joint normal distribution; thus, σ _v1 ² gives the segregation for Africans at the MSOA level, and we can test whether this is different from the variance for Caribbeans, σ _v2 ² . The higher-level covariance term, when standardized, will give the correlation between the differences at that level between each pair of ethnic groups—that is, the extent to which ethnic groups co-locate at that scale.

Classic measures of segregation (like D) are aspatial and depend only on the numerical values in each observation unit (e.g., OAs in London), taking no account of the situation in surrounding areas or the spatial patterning in the rates. Swapping the units spatially so that all the areas with large Caribbean populations are contiguous would produce no change in such an index but may imply much greater segregation. Researchers in recent years have shown considerable interest in developing spatially sensitive measures from two broad viewpoints. From the spatial econometrics perspective (Paelinck and Klaassen 1979), Wong (1998) developed a family of local spatial segregation indices that take account of neighborhood joins, thus taking into account the population characteristics of a wider area (defined a priori as in touching boundaries). Wong (2003) extended these measures to work at multiple scales, but those measures were not set in an inferential framework and were based on the observed (and therefore potentially unreliable) local rates. From the spatial smoothing perspective (Fotheringham et al. 2002), Reardon and his coauthors (2004, 2009) also developed spatially sensitive measures. They used a spatially weighted version of the information theory index where the weights are determined a priori by some function of the spatial distance between areas. Lee et al. (2008; see also Östh et al. 2014, 2015) used this approach to analyze multiple scales by defining circles of different radii and moving these around the map. They provided an analysis at scales from a 500 meter radius (a pedestrian-based neighborhood) to a 4,000 meter radius (which they call a “macro-local environment”) of nearly 20 square miles.¹² Their determination of the degree of segregation involved no modeling or inferential framework to deal with unreliable rates resulting from small counts.

Both sets of approaches take into account local spatial autocorrelation or dependency. It may be thought that multilevel models of the type that we are developing here are aspatial, and Elffers (2003) has argued that the between-area variance is invariant to the spatial arrangement of areas. This is undoubtedly true of the standard two-level model, but it is not true in general, for two reasons. First, it is possible to include spatial weights in a multilevel model for the higher areal levels (see Jones and Subramanian 2014b) and thereby estimate both unstructured and spatially structured segregation. Such explicitly spatial multilevel modeling is undergoing rapid development.¹³ Second, the hierarchical model with more than two levels has an implicit spatial dependence, and we now consider this in more detail.

We begin by stressing that multilevel models analyze the within- and between-differences (Bell and Jones 2015), and the variance σ _u1 ² in the second formulation no longer summarizes differences between OAs as in the original model but now represents the differences between OAs after taking account of the differences between MSOAs in which they are located (Subramanian et al. 2001). To illustrate the procedure, Fig. 1 uses two higher-scale examples. In each of the three diagrams, the solid line shows the relevant London-wide relative rate of 1—the theoretical even distribution of the population who are, say, Bangladeshi. The city is split at the larger spatial scale into two MSOAs, A and B, each of which is divided at the smaller scale into three OAs. In panel (a) A and B differ substantially in their Bangladeshi relative rates but differ little across their constituent OAs within each MSOA. Segregation is substantial at the larger scale but, holding its extent at that level constant, insubstantial at the smaller scale. Bangladeshis are concentrated in B; but in both A and B, there is little within-MSOA, between-OA variation. Panel (b), on the other hand, shows little difference between MSOA A and B but substantial variation within each; and panel (c) shows substantial variation at both scales. In panel (a), therefore, segregation displays macro-scale variability only; in panel (b), it displays only micro-scale variability; and in panel (c), there is substantial segregation at both scales. Because the multilevel approach measures segregation at one scale net of the others, it does not inevitably mean that the finer scale is necessarily the most segregated.

The key notion in the model is that the highest-level difference is a random, allowed-to-vary departure from a general relationship, and each level’s residual is an allowed-to-vary random departure from the higher-level departure. Consequently, we can calculate a variance partitioning coefficient (VPC: Goldstein 2011; Jones and Subramanian 2014a), which decomposes the total variance into the multiple scales. Moreover, this VPC gives the proportion of the variance between—or the degree of similarity or correlation within—scales, equivalent to the well-known intraclass correlation coefficient (Kish 1954).

Conceptually, in the three-level example, σ _v1 ² is the between-MSOA variance for ethnic group 1, which is our measure of segregation at this level; σ _u1 ² is the within-MSOA, between-OA variance for ethnic group 1, which is a measure of segregation at the OA level net of differences at the MSOA level; and σ _v1 ²  + σ _u1 ² is the between-OA variance for ethnic group 1, which is equivalent to the measure of variance for that scale in the initial two-level model.

Consequently, the proportion of the total variance due to differences between MSOAs, the intra-MSOA correlation is given bywhere σ _e1 ² is the within-MSOA, within-OA, between-people variance for group 1.¹⁴

The proportion of the variance due to differences between OAs, the intra-OA correlation is given by

Finally, we can calculate the similarity of OAs within the same MSOAs:

The hierarchical structure is therefore defining the local neighborhood structure, and we are implicitly modeling spatial dependence. The degree of segregation is not invariant to swapping because we are specifying that a set of OAs belongs within—is hierarchically nested in—a specific MSOA. The inherently spatial nature of this dependence is shown in Fig. 2. Cells (C) are sorted so that they are nested in OAs (O) and MSOAs (M), and it can then be seen that intra-OA correlation (ρ₁) assesses the degree of correlation in the same MSOA and same OA, while the intra-MSOA correlation (ρ₂) gives the correlation for those in the same MSOA but different OAs.¹⁵

The final elaboration of the model is to extend it to more than two ethnicities and to more than three scales. This is trivial in terms of specification but increases estimation time substantially.

In the Poisson model, quasi-likelihood empirical Bayes (EB) procedures have been found to overestimate the higher-level variance (Jones and Subramanian 2014b). Consequently, we use full Bayes (FB) procedures for all the models in this article specified with minimally informative prior distributions.¹⁶ The FB approach allows the calculation of (potentially asymmetric) CIs, which indicate the degree of support enjoyed by different values of our segregation measures, which are estimated without having to rely on asymptotic normality assumptions that are unlikely to hold in applications with a relatively small number of units (as in the case of the Boroughs) and for variance terms that cannot be less than 0.

An important by-product of the MCMC estimation is the deviance information criterion (DIC; Spiegelhalter et al. 2002), which yields an estimate of the badness of fit of the model penalized by model complexity, which in turn is estimated by the degrees of freedom consumed in the fit (pD). A difference in DIC of 10 between two models implies very little support for the model with the higher value of DIC.

We estimate the models with the MLwiN 2.31 software (Rasbash et al. 2014). The estimates are based on an initial quasi-likelihood estimation, a discarded burn-in of 50,000 simulations to get away from potentially biased results, and a further 100,000 monitoring simulations according to Draper’s (2008) good-practice recommendations. We find it beneficial to use hierarchical centering to obtain less-correlated chains—that is, more informative chains (Browne 2012). The trace of the estimates is evaluated for convergence (shown by lack of trend), and the models were run so that the effective sample size (ESS) of the estimates for each parameter was at least equivalent to 500 independent estimates to characterize the degree of support for parameter values. The estimation takes several days on a standard desktop PC.

The first set of interpretations looks at each ethnic group separately, exploring differences in the level of segregation across the four scales and establishing whether these differences are statistically substantial using the CIs, as commended in what Cummings (2014) termed the “new statistics,” which abjures the term statistically significant. Thus, the first block shows that the White British are more segregated at the Borough than at the MSOA scale, but the overlap between the CIs for the two measures suggests no substantial difference. Those levels of segregation are, however, much larger than those at the smaller two scales, and substantially so. White British people are substantially segregated into particular boroughs and MSOAs within London; within each of those units, however, small-scale variation around the local average is minimal. In general, therefore, each London Borough is relatively homogeneous across its smaller-scale areas in the White British share of the local population: whatever the percentage White British overall (and in most cases, it was either high or low), there is little variation around that figure across its constituent neighborhoods.

To aid comparison, we reexpress the variances as MRRs (Table 3). For interpretation, we can classify these ratios according to well-known effect sizes, as Cohen (1988) recommended originally for odds ratios. Accordingly, values greater than 4.3 indicate very large ratios: MRRs between 2.5 and 4.3 and between 1.5 and 2.5 are considered medium and small, respectively; and MRRs less than 1.5 are treated as low.¹⁷ The pattern is very clear, as summarized in Table 4: the overwhelming number of MRRs have either low (below 1.5) or small (1.5 to 2.5) values. The only exceptions are (1) five Borough rates, four of which are medium (for Indian, Pakistani, Black Caribbean, and Arab), one large (for Bangladeshi); and (2) one OA rate that is medium (again, for Bangladeshi).

The across-scale differences for each of the ethnic groups are shown in Table 3. For 8 of the 13 groups (White Irish, White Other, Indian, Pakistani, Bangladeshi, Black African, Black Caribbean, and Arab), segregation is highest at the Borough scale and second highest at the OA scale, with much lower levels at the MSOA and LSOA scales. Overlapping CIs indicate that those differences are not statistically substantial in several cases; however, the CIs for Borough and OA scales overlap for Pakistanis, but not for Indians and Bangladeshis.

No other group has a pattern similar to the White British: that of continued declining segregation with decreasing scale. The three mixed-ethnicity groups—like the eight identified previously—have their highest segregation levels at the Borough and OA scales, too, but larger for the latter than the former (although in the case of the White-Black Caribbean mixed group, the CIs overlap). The Chinese stand out with much greater modeled segregation at the OA level than any of the other three; within each Borough, a small number of neighborhoods have relatively numerous Chinese residents, but they are not substantially concentrated in particular Boroughs.

In general, therefore, these modeled variances suggest that for most of London’s ethnic groups in 2011, segregation was both a macro- and a micro-scale phenomenon (the Borough and OA, respectively) but not also at the meso-scale (MSOA and LSOA). Most migrant groups are concentrated into particular boroughs, and within them there is significant small-scale local variation given that they are clustered in some parts of the boroughs but not others. The White British are the main exception to this: they, too, are concentrated in particular boroughs, within which there is no local spatial variation. They are represented by panel (a) in Fig. 1, whereas most of the others fall in type (c) (also shown in Fig. 1), and the Chinese are the main exception as type (b) (Fig. 1). The majority White British are concentrated in large blocks of territory, represented here at the Borough scale, in which the subdivisions are homogeneously White, whereas the minority ethnic groups are also concentrated in (a smaller number of) certain boroughs and additionally in certain small blocks within those boroughs.

In these comparisons, the segregation levels (variances) are rank ordered to identify which are the most- and least-segregated ethnic groups at each scale (Table 5). Although there are differences in detail, the general pattern is very clear: the most-segregated groups at all four scales are those with self-assessed Asian (especially South Asian) and Black ethnicities, whereas the least segregated at every scale, too—are the White British, Irish, and Others. The levels of segregation for those claiming a mixed ethnic identity tend to be less than those for the nonwhite group that they partially identify with, but more than for the white populations.

Although the 13 groups can be arranged along continua as in Table 5, the differences between adjacent groups are rarely significantly different, especially at the larger spatial scales. At the Borough scale, for example, the CIs overlap between every adjacent pair, but there are differences between nonadjacent pairs.¹⁸ The first substantial difference (denoted by the underline) along the continuum is between the Bangladeshis (variance (V) = 3.443; CI = 2.025, 5.835) and the Indians (V = 1.179; CI = 0.705, 1.905). The next substantial difference is between Indians and the mixed White–Black Caribbean group (V = 0.294; CI = 0.174, 0.483). No group below that on the continuum differs substantially from the White–Black Caribbeans, suggesting that the 13 groups can be divided into three according to their segregation level at that scale, with the boundaries between the three groups shown by lines in the ranking: (1) three of the four South Asian groups (the most segregated at that scale); (2) the Indians, the two Black groups, and the White Others (less segregated); and (3) the remaining groups, comprising the White British and Irish, the Chinese, and the three mixed groups (the least segregated).

Similar splits are reported in the other three columns of Table 5 for the smaller scales. They show greater variety: more clusters of ethnic groups that differ substantially from their neighbors in their degree of segregation, although segregation levels are generally low at the MSOA and LSOA scales. The greatest degree of substantial variation is at the smallest scale (the OA), which is divided into nine segments in each of which the top-ranked ethnic group has a significantly smaller level of segregation than that at the bottom of the segment above it and where each of the four least segregated groups has a significantly smaller modeled level of segregation from that immediately above it on the continuum. The Bangladeshis are the most segregated at all four scales, and the Arabs and Pakistanis are also highly segregated across the four; at the other extreme, the White British are either the least- or the second least–segregated group (again, as with all of the other comparisons, with segregation at the higher scales held constant).

These findings are in line with those of other descriptive, single-scale analyses of ethnic segregation in London (e.g., Johnston et al. 2015), but also extend them. There is no established theory suggesting which groups should be most or least segregated, let alone of any variations across scales. In general, however, the more recent arrivals are expected to be more segregated than the longer-established groups; those culturally more distinct from the host society are expected be more segregated than those that are less so (most Black Caribbeans are Christians, for example); and those claiming mixed ethnicities are expected to be less segregated than the minority group with which they partially identify but more segregated than the dominant White groups (their mixed identity being an indicator of cultural, and possibly economic, assimilation). The findings reported in Tables 2, 3, and 5 sustain that interpretation. Thus, the predominantly Muslim Bangladeshis and Pakistanis are among the most segregated groups at every scale, for example, but the more heterogeneous other South Asian group (Indians, comprising Muslims and Sikhs as well as the majority Hindus¹⁹) is less so; and the White and mixed groups are among the least segregated.

In addition, however, the decomposition provided by the modeled variances—the assessed segregation level at each scale is net of that identified at the higher scales—provides information not available from other studies. This is exemplified in two particular cases. The Black Caribbeans are long-established in London—large-scale immigration having been initiated in the late 1940s—and the group has not grown over the most recent decade with few new arrivals (Jivraj and Simpson 2015). They are relatively highly segregated at the macro-scale (i.e., Borough), reflecting the parts of London in which they initially settled, but much less so at the micro-scale (i.e., OA), almost certainly indicative of economic and social mobility over the last few decades; while concentrated in particular parts of London (indicative of inertia in residential decision-making at the macro-scale), they are not strongly clustered within particular smaller areas there—a patterning that distinguishes them from several other more recent and still-expanding (Johnston et al. 2013) immigrant groups, including the Black Africans.

By way of contrast, the relative size of the segregation measures for the Chinese is the inverse of that for the Black Caribbeans. They rank ninth among the 13 groups for their degree of segregation at the Borough scale, for example, but fourth at the OA scale (and the CIs in Table 3 indicate a statistically significant difference between the two). Across London as a whole, therefore, the Chinese are relatively widely distributed—certainly more so than the other Asian ethnic groups. (The Chinese MRR at that scale is significantly smaller than those for the Bangladeshis, Pakistanis, and Indians: see Table 2.) Within those areas where they are relatively concentrated, however, they are more clustered at the most local scale (at the OA level) than all but three other groups, including Indians. As with Black Caribbeans, therefore, these multiscale estimates net of any segregation at higher scales not only confirm our general appreciation of which groups are more or less segregated; they also add an important indication of the statistical significance of those differences and demonstrate that relative levels of segregation vary by scale, indicative, among other factors, of the groups’ length of settlement in the city and its degree of economic and cultural assimilation into the wider society.

This final analysis explores the degree to which the various ethnic groups are segregated into the same areas, at the four spatial scales. This involves standardizing the covariance for each set of differences for each ethnicity at the relevant scale. The resultant correlation coefficients are reported in Table 6. Only those exceeding ±0.4 are shown, with positive coefficients in italics and negative coefficients in bold. The first block shows the correlations at the Borough (below the diagonal) and MSOA (above the diagonal) scales; the lower block does the same, respectively, for the LSOA and OA scales. One clear conclusion stands out: the sparseness of the matrices (the relatively small number of coefficients >±0.4) indicates that most distributions are relatively independent of each other, with few (especially strong) common patterns. This is particularly the case at the smaller scales: only 21 of the 78 coefficients are above that threshold in the Borough analyses; 25 at the MSOA scale; 13 at the LSOA scale; and just three at the OA. A finding of few positive correlations indicates that at all four scales, each of the groups has a distinct residential distribution, separate from that of most if not all of the 12 others.

The only substantial negative correlations in the four matrices apply to the White British population: at the Borough and MSOA scales, areas where there are many more White British residents than average tend to have fewer than average members of several of the Asian and Black ethnic groups. Among the positive correlations, some of the largest refer to the Black Africans and Black Caribbeans, plus those claiming a mixed White–Black African/Black Caribbean identity; these four groups cluster together in above average proportions at all scales. The only other clear pattern of clustering together—notably at the Borough and MSOA scales—is of Indians and Pakistanis, many of whom can be found not only in the same (western) sector of London (Johnston et al. 2014) but also clustered in major segments of those boroughs.