Impacts of Social Environments on Neighborhood Depression Incidence: Fully Accounting for Spatial Effects

1. Introduction

Neighborhood variations in depression are an important aspect of the overall mental health burden, and typically co-morbid with other mental illnesses such as psychosis. Such variations, or those in common mental disorders more generally, have been explored relatively little, but available studies [1,2] show major influences of area socio-demographic composition, ethnic mix, and contextual environmental factors, such as area crime, urbanicity, and neighborhood social cohesion. Meta-analysis [3] shows significant impacts of neighborhood on depressive symptoms, controlling for individual attributes (such as age, income, and ethnicity). Neighborhood influences are relevant to spatial variations in other mental health outcomes [4,5].

We focus in the present study on such place effects, namely the impacts of socio-behavioral environments. In particular, we consider the role in explaining neighborhood depression of cohesion and crime [6,7], and how these impacts operate spatially. It has been argued [8] that existing studies of depression over-emphasize the relevance of socio-economic composition and neglect the role of contextual neighborhood characteristics, both objective and subjective (e.g. actual and perceived crime levels). Thus [8, page 43] states “both failure to control for nesting of people within geographic space [..], and an almost exclusive focus on neighborhood socioeconomic status [..] limit understanding of the ways in which neighborhood characteristics influence depressive symptoms”.

1.1. Full Accounting for Spatial Effects

As well as taking fuller account of health impacts linked to socio-environmental contexts, we also consider associated issues in appropriate modelling of spatial impacts. Models for spatial variations in mental health, such as those based on Bayesian disease mapping [9,10,11], follow the standard disease mapping approach [12,13], namely: spatially homogenous effects of area predictors; predictors defined without regard to possible spillover effects; and a spatially structured residual random effect to account for unmodelled spatial dependencies.

We instead here assess gains (statistical and substantive) from considering impacts of spatial structure and spatial dependence more comprehensively. This is done in three major ways.

First, we consider spatial spillovers in environmental influences. The importance of spillover is acknowledged in causal approaches to assessing impacts of neighborhood exposures [14]. It is also recognized in spatial linear regression methodologies, including regression on spatially lagged predictors, and known as spatial Durbin regression [15]. An example of these two approaches to overlap is provided in studies of pollution impacts by Giffin et al [16] using a causal method, and by Sarrias and Molina-Varas [17], using spatial Durbin regression.

Spillover may also characterize impacts of social environments (e.g. cohesion, crime), access environments (e.g. exercise access) and so on, as these are not defined spatially by the arbitrary administrative boundaries of the neighborhoods typically used in ecological disease studies. For example, residents living nearer neighborhood boundaries are likely to experience spillover effects of environmental levels in adjacent areas. Regarding neighborhood crime rates, Graif [18] mention that “despite the rapidly growing ecological-level evidence indicating geographic spillover effects on crime, research on neighborhood effects [..] has treated neighborhoods as if they were isolated islands, independent of their surroundings”.

On conceptual grounds spatial spillover would be anticipated as most relevant for contextual environmental influences. A compositional effect is inherently an effect of the socio-demography of the local population, whereas a contextual effect from environmental influences is spatially unbounded in impact, spilling over administrative boundaries.

Second, we explicitly consider context heterogeneity (spatial non-stationarity), particularly regarding environmental influences. As discussed in [19] p.112 “much of the observed spatial correlation in residuals that is frequently observed [..] applied to spatial data results from applying a global model to a non-stationary process”. Non-stationarity can often be summarized in terms of spatial regimes, defined by Vidoli and Benedetti [20] as “an aggregation of neighboring units that are homogeneous in functional terms or that share the same relationship between a dependent variable and some covariates”. These have been infrequently considered in spatial epidemiology, though see Sridharan et al [21] for an exception.

Non-stationarity in neighborhood depression has been analyzed using frequentist geographically weighted (GWR) regression [22,23]. Drawbacks with the usual application of GWR are mentioned by [24]. Here we use Bayesian approaches [25] to spatially varying coefficients (SVC) that have been shown to have advantages in analyzing spatial data. For example, these include full posterior summaries regarding spatial random effects and can also be applied with variances themselves treated as non-stationary.

Thus, we allow for variations in the precision of spatial random effects, both for spatially varying coefficients and the spatial residual. The study [26] considered such heteroscedasticity just for spatial residuals, because spatial effects may be inherently less variable in some parts of the overall region (domain) being considered. We extend this approach here to heteroscedasticity in SVC effects.

Under the conditional autoregressive (CAR) spatial random effects approach of Besag et al [27] commonly used in the standard model, spatial variability, as measured by a precision parameter, is assumed constant over the domain. This assumption may be restrictive for a large domain with many small areas. We accordingly use multiple precision parameters, expressing different intensities of spatial dependence in different sub-regions. We apply a penalized complexity prior [28] to varying precision parameters to prevent overfitting.

Heteroscedasticity in CAR spatial random effects has been considered before e.g. [29] but existing approaches may involve many weakly identified parameters. Spatial volatility models are quite widely applied in spatial econometric applications using simultaneous autoregressive (SAR) models applied to continuous responses e.g. [30] under a linear normal assumption. However, these methods do not transfer straightforwardly to geographic disease applications, typically with count responses.

1.2. Depression Incidence in English Neighborhoods

We consider impacts of social-contextual environments in a study of depression incidence across 6856 England neighborhoods. We develop a measure of neighborhood cohesion using cross-scale modelling methods [31,32], since observed data on cohesion indicators is not available at neighborhood level, but for 296 larger areas (English local authorities). We then use regression methods to assess impacts of environment and socio-demographic characteristics on depression incidence. We find crime and cohesion as major influences on area incident depression, outweighing those of composition (neighborhood socio-demography).

We show improved model performance by explicitly considering context heterogeneity and heteroscedasticity (namely the two forms of spatial non-stationarity) and spillover effects in socio-behavioral environments. Moreover, allowing for non-stationarity and spillover enhances the total impacts of these environments on depression incidence.

2. Data and Methods

2.1. Outcome Data

We seek to explain neighborhood variations in the incidence of clinically diagnosed depression (among people over 18) for N=6856 Census defined neighborhoods, known as Middle Level Super Output Areas (MSOAs). These provide entire coverage of England, one of the UK nations. These data are obtained from the Quality Outcomes Framework (QOF), which records chronic disease prevalence in UK primary care [33,34]. There were 714,592 new depression cases across England in the year considered (2023/24), with an incidence rate (among adults) of 16.3 per 1,000. Expected MSOA cases are based on age rates for depression from the Global Burden of Disease site (https://www.healthdata.org/research-analysis/gbd) applied to MSOA populations.

There are considerable spatial inequalities in incidence, going beyond what variations in socio-demographic composition would appear to be able to explain: across the nine standard regions composing England, the incidence rates vary from 12 per 1000 (Eastern region) to 23.7 per 1000 (North West England). At MSOA level, standardized incidence ratios (new cases divided by expected new cases) vary from 0.13 to 3.97, with 5th and 95th percentiles of 0.46 and 1.73, a nearly 4-fold contrast.

2.2. Potential Predictors for Regression Models

The choice of predictors for regression to explain such inequalities is based on evidence from other studies [35,36]. The first predictor is MSOA socio-economic status (SES). This index is based on UK Census 2021 indicators (see Appendix 1). Area SES, and inverse status measures such as area deprivation, have been shown to influence levels of neighborhood mental ill-health [35].

Secondly, social fragmentation [37] is a measure of transience, household tenure insecurity, and a non-family household structure, which has been shown as a significant influence on mental ill-health [35,37]. This index is also a composite based on Census data and has been taken as a proxy for neighborhood cohesion [38].

However, in the present study we develop an index of neighborhood cohesion based more directly on concepts underlying cohesion [39,40,41]. The methods used are set out in section 2.3, and the derivation of this index precedes the regression analysis.

Many studies also show enhanced risk of mental ill health, including depression [42], in more urban areas, though measures of urbanicity vary. Here we adopt a measure based on concepts of urban form [43].

Crime levels are also associated with enhanced rates of mental ill-health [44], and crime may reinforce lower social cohesion. Violent crime and perceived lack of safety inhibit normal activities and may impact older people especially [8]. The study [8] mentions that for older people “characteristics of the environment [such as observed and perceived crime] assume greater salience” and criticizes previous studies of neighborhood effects, characterized by a “dominant trend of defining neighborhoods according to the socioeconomic status of its residents.”.

In the UK, ethnic mix is a potentially important influence on mental health outcomes. There is considerable evidence of higher psychosis among people of black and mixed ethnicity, but ethnic differentials in depression are less well researched. The study by Williams et al [45] finds excess depression risk compared to white British groups, though the ethnicity effect is intertwined with that of socio-economic disadvantage and urbanicity.

2.3. Developing the Index of Neighborhood Social Cohesion: Cross-Scale Analysis

Spatial disaggregation is needed to estimate MSOA neighborhood cohesion and involves three indicators, which are observed only at a higher spatial scale (296 local authorities). The indicators used are central criteria for cohesion [40], page 289 based on responses to questions in the Community Life Survey (CLS) of England [46], regarding belongingness, trust, and social interaction, specifically: how strongly do you feel you belong to your immediate neighborhood?; how often do you chat to your neighbors?; thinking about the people who live in this neighborhood, to what extent do you believe they can be trusted? We take as positive evidence of cohesion the responses: belonging ‘very strongly’ or ‘fairly strongly’; chatting more than once a month; and the answer that many of the people can be trusted.

The disaggregation estimation uses cross-scale methods [31,32], with the likelihood being for observed CLS data for England local authorities (e.g. on neighborhood belonginess), but based on a model predicting outcomes for MSOAs within local authorities. Bayesian methods are used via the BUGS program [47]; see Appendix 2 for a statistical specification. The neighborhood (MSOA) regression model includes neighborhood characteristics, based on evidence from the CLS regarding factors that especially impact on cohesion, and a neighborhood spatial random effect. The neighborhood cohesion index is then the score of the leading component from principal component analysis of the predicted proportions of MSOA populations with positive views of belongingness, trust and interaction.

The characteristics used in the MSOA regression for the cross-scale models are area socio-economic status, proportions non-white, and urbanicity. For example, official commentary on CLS findings [48] reports that “adults living in rural areas were more likely to feel a sense of belonging to their neighborhood (69%) than adults living in urban areas (59%)”, while regarding interaction, “compared with the England average (69%), the proportion of adults who chatted to their neighbors at least once a month was higher among adults from the White British ethnic group (72%)”.

2.4. Initial Sifting of Predictors

We carry out initial regressions at MSOA level using a standard disease mapping approach within the R INLA program: spatially homogenous effects of area predictors, and a spatial (conditional autoregressive) random effect [11,12,24] to account for unmodelled spatial dependencies.

This approach is used because of collinearities between the predictors: between fragmentation and cohesion, and between cohesion and urbanicity. To anticipate later results, fragmentation and urbanicity are excluded from later analysis due to collinearities causing regression findings at odds with accumulated evidence.

2.5. Regression Strategy

With a reduced set of predictors, we consider four successive regression frameworks. The first uses the standard disease mapping approach with a negative binomial likelihood in INLA, namely spatially homogenous predictor effects, no account of spatial spillover, and a spatial (conditional autoregressive) residual random effect.

The second introduces spatial lag effects for the environmental influences, crime and neighborhood cohesion, where spillover seems most likely. We assume a spatial interaction matrix defined by adjacency, with $w_{ij}$=1 if MSOAs i and j are adjacent, and otherwise $w_{ij}$=0. Then the spatial lag calculation involves taking the average of the measure (say x) in neighborhoods adjacent to neighborhood i, namely ${{\displaystyle \sum }}_{j\ne i}{w}_{ij}{x}_j.$This spatial lag calculation is provided by the R package spdep.

The third framework extends the second by taking spatially varying effects of the environmental variables, following the usual Bayesian spatially varying coefficient strategy [25]. This allows for possible non-stationary impacts of local environment variables, which may impact on their average effect.

The fourth is most general, allowing for region-specific heteroscedasticity both in the spatially varying regression coefficients and in the spatial residual. The SVC strategy is applied to both local and spillover coefficients. We use the nine English standard regions to assess heteroscedasticity. This is implemented using the fbesag option within the R program INLA [26]; see Appendix 3.

2.6. Spatial Regimes

Demonstrable non-stationarity from the regression stage may be summarized using the notion of spatial regimes. The R spatially constrained cluster analysis package ClustGeo, implements hierarchical clustering of area attributes, but including spatial constraints, so that spatial clustering in depression risk, neighborhood characteristics and predictor effects is allowed for. A mixing parameter is chosen to obtain both high intra-cluster homogeneity in neighborhood characteristics, but also high spatial proximity within clusters [48].

2.7. Measuring Goodness of Fit

We use three goodness of fit measures to compare these regression frameworks: the deviance information criterion (DIC) [50], the widely applicable information criterion (WAIC) [51], and the average absolute deviation (AAD) between actual and fitted incidence counts.

3. Results

3.1. Findings from Cross-Scale Modelling Regarding Neighborhood Cohesion

Table 1 shows the results of the three regressions which form part of the cross-scale modelling of neighborhood cohesion. It is evident that less urban environments are significantly associated with a greater sense of neighborhood belongingness, with trusting neighbors, and with neighborhood social interaction. Higher neighborhood SES is also associated with higher levels of belongingness and trust. Higher proportions of non-white groups tend to reduce trust and neighborhood interactions.

The three MSOA indicators obtained from the cross-scale analysis are combined into a single cohesion score via principal component analysis. The first component accounts for 87% of variation in the three separate indicators. This score is positive for areas with high cohesion.

The strong association with urban level apparent from Table 1 also shows if MSOAs are classified according to whether or not they are in the highest quintile for the resulting cohesion scores. Table 2 shows percentages of neighborhoods with high cohesion across the nine English region and according to a three-way urban settlement classification [52].

Across all regions, highest cohesion is most apparent in smaller rural settlements. By contrast, urban areas have low cohesion. Thus, only 3% of London neighborhoods exhibit high cohesion, and only 10% of all urban neighborhoods show high cohesion.

There is in fact a correlation of -0.87 between the urbanicity score and the cohesion score. Such collinearity is relevant to choice of predictors in subsequent regressions involving varying depression incidence as the outcome.

3.2. Multicollinearity

Initial regression of depression incidence following the standard disease mapping approach included the following neighborhood predictors: area SES; percent non-white; social fragmentation; urbanicity; social cohesion score; and crime level. This regression showed negative impacts of urbanicity and fragmentation, at odds with substantive evidence and reflecting collinearity: namely correlations of -0.87 between urbanicity and cohesion; of 0.61 between urbanicity and fragmentation; -0.54 between fragmentation and cohesion; and of 0.59 between crime and urbanicity. We therefore retain a subset of predictors with effects in this initial regression that are in line with substantive evidence: area SES; percent non-white; cohesion and crime.

3.3. Neighborhood Variations in Depression Incidence: Regression Sequence

Table 3 shows the results of the regression sequence proposed in section 2.5 with these four predictors – extending cohesion and crime effects to include spatial lags. The predictors are all on a [0,1] scale, and the coefficients in Table 3 represent logged relative risks of depression incidence.

Table 3 also shows the improved fit obtained under the second and subsequent models (denoted M2, M3 and M4), allowing for spillover and non-stationarity, as compared to the standard model (M1). The improved fit is especially notable for model M4 with region specific spatial precisions combined with spillover effects.

Regarding environmental impacts, comparing spillover and nonstationary models to the standard model, spillover and local effects of both socio-behavioral environments are shown to be cumulative. Both the local and spillover effects for cohesion in M2, M3 and M4 are negative, namely greater cohesion reduces depression incidence (a protective effect). Both the local and spillover effects of crime are positive, namely higher crime increases depression incidence. This cumulation of effects enhances the overall role of these environments in explaining variations in depression incidence.

3.4. Comparing Relative Depression Risks

Accordingly, we consider implications for effects of cohesion and crime under different regression models. For an MSOA with high local crime, say 0.9 on the crime scale, and high spillover crime, also 0.9, the anticipated relative risk can be compared with an MSOA with low local crime, say 0.1, and low spillover crime, also 0.1. Other predictors are assumed to have value 0.5.

Model M2 adds spatial lag effects to the standard disease mapping approach. The spillover effect is most pronounced for crime, with coefficient of 0.306 and 95% interval (0.139, 0.473). This exceeds the local effect, which has a coefficient of 0.243. The major spillover effect of crime is maintained in model M3 and M4.

Considering the coefficients in M2, the relative depression risk for high crime areas is 1.24, as compared to 0.80 in low crime areas, a relative risk ratio of 1.55. By contrast, under the standard model, M1, the contrast in these respective relative risks is attenuated: 1.10 vs 0.88, or a ratio of 1.26.

Similarly, under M2, relative depression incidence for high cohesion areas is 0.849, as compared to relative incidence of 1.157 in low cohesion areas. So, the protective effect (reduced risk) can be expressed as a ratio of these two extremes, namely 0.73. In the standard model, this protective effect is closer to the default 1, namely 0.80.

3.5. Models with Non-Stationarity

It is notable that nonstationary models, both M3 and M4, show most evidence of non-stationarity for cohesion effects, as measured by the standard deviation of varying neighborhood coefficients around the average coefficient. The standard deviation in model M4 of varying cohesion coefficients is 0.215 (around the average coefficient of -0.219). Variation in the crime coefficients is much smaller, with a standard deviation of 0.016. In model M3 these two standard deviations are respectively 0.223 and 0.012.

Figure 1 and Figure 2 show the varying coefficient estimates from model M4 for cohesion and crime respectively. The strongest – most negatively signed and hence most protective – effects of cohesion are in the northern half of England. By contrast, the strongest – most adverse – effects of crime on depression risk seem to be in the southern half of England. In the South East region outside London, mostly relatively prosperous and less urbanized, the average local cohesion coefficient is just -0.05, compared to the average -0.219. These patterns suggest that crime effects are stronger where cohesion effects are weaker – that cohesion moderates the impact of crime.

Figure 1 and Figure 2 demonstrate simply the effects of environmental variables on health outcomes are differentiated over broad regional settings. Table 4 accordingly shows estimated depression relative risks, and average impacts of environmental risk factors, according to settlement type and standard region. The quite sharp differences in risk factor impacts shown in Figure 1 and Figure 2 also show in this Table. Cohesion is the strongest influence on depression in the North East region and adjacent Yorkshire and Humberside. In South East England, varying crime levels are the predominant influence on varying depression.

3.6. Implications for Spatial Regimes

These patterns recall the notion of spatial regimes discussed above. We accordingly apply clustGeo with 12 clusters and mixing parameter 0.3, as this shows a moderate 13% loss in attribute homogeneity and a 15% loss in spatial homogeneity [48, page 1815].

Table 5 shows, for example, clusters 4 and 12 to be mostly in southern England (South East and South West), to show a low level of urbanicity, and the local crime effect (enhancing depression risk) to outweigh the local cohesion effect. By contrast, clusters concentrated in the North and Midlands, such as clusters 6, 9 and cluster 10, show the protective cohesion effect outweighing the adverse crime effect. Cluster 6, concentrated in two regions of north east England (the North East region, Yorkshire and Humberside), shows the strongest protective effect of neighborhood cohesion.

4. Discussion

There is extensive literature on the canonical disease mapping approach for disease counts, set out first by Besag et al [24], and discussed in later reviews [11,12,52]. This specifies spatially constant effects of known neighborhood risk factors, with a spatially correlated residual to represent unmodelled spatial dependencies.

The number of studies adapting this canonical approach to consider spatially varying regression coefficients or spatial volatility is, by contrast, relatively small, certainly much fewer than the number of frequentist studies applying GWR and linear model spatial volatility – for reviews of the latter see Comber et al [53] and Otto et al [29].

The lack of consideration of spatial heterogeneity in neighborhood disease studies is noted by Sridharan et al [20] using a non-Bayesian form of spatial model. This neglect of possible non-stationarity may lead to a potentially oversimplified view of risk factor effects and rules out the potential for identifying spatial regimes.

Similarly, the great majority of disease mapping studies do not evaluate the extent of spillover in environmental risk factors, where environmental risks (crime, pollution, food access, neighborhood social cohesion, etc.) are not defined by the arbitrary administrative boundaries most commonly used in disease mapping. However, acknowledging spillover is central to causal approaches to assessing neighborhood exposures. Spatial interactions are a form of network and relevant causal approaches acknowledging spillover (known as interference in the causal literature) are set out by Forastiere et al [54]. From a substantive viewpoint, the notion of extended neighborhoods, suggested by Graif [17] in relation to spillover effects in neighborhood crime, is also relevant to environmental impacts on disease.

Without considering these two respects, non-stationarity and spillover, the canonical disease model can be considered as potentially simplistic and may underestimate environmental impacts. This is certainly the case in the present study of neighborhood depression incidence, where we assess impacts of contextual neighborhood environments.

We have focused especially on contextual impacts on depression, similarly to the approach in [8], who criticize the over-emphasis on the deprivation-health link. Our findings reinforce existing evidence [7,55,56] on impacts of cohesion and crime on depression, with the spatial regime findings showing the interplay of these influences: typically, one is weaker while the other is stronger.

In terms of modelling effectiveness, we find a considerably improved fit through acknowledging non-stationarity and spillover in environmental impacts. In particular, we have shown considerable variability in impacts of social environments, especially neighborhood cohesion. Spillover effects are also considerable, especially for neighborhood crime.

5. Conclusions

We have focused in this study on contextual effects in spatial epidemiology, namely the impacts of socio-behavioral environments on neighborhood mental health. In particular, we have considered variations in neighborhood depression incidence according to levels of cohesion and crime, and aimed to provide a nuanced assessment of how these impacts operate spatially, moving beyond the simplifying assumptions of the standard Bayesian disease mapping approach. We have shown much improved fit through acknowledging non-stationarity and spillover, apparent especially in variable impacts of neighborhood cohesion, and spillover effects for neighborhood crime.