(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Association of environmental and socioeconomic indicators with serious mental illness diagnoses identified from general practitioner practice data in England: A spatial Bayesian modelling study [1]

['Joana Cruz', 'Department Of Environment', 'Geography', 'University Of York', 'Wentworth Way', 'York', 'United Kingdom', 'Guangquan Li', 'Department Of Mathematics', 'Physics']

Date: 2022-07 

Our study provides further evidence on the significance of socioeconomic associations in patterns of SMI but emphasises the additional importance of considering environmental characteristics alongside socioeconomic variables in understanding these patterns. In this study, we did not observe a significant association between green space and SMI prevalence, but we did identify an apparent association between green spaces with a lake and SMI prevalence. Deprivation, higher concentrations of air pollution, and higher proportion of ethnic minorities were associated with higher SMI prevalence, supporting a social-ecological approach to public health prevention. It also provides evidence of the significance of spatial analysis in revealing the importance of place and context in influencing area-based patterns of SMI.

Mean SMI prevalence at LSOA level in major conurbations mirrored the national associations with a few exceptions. In Birmingham, higher average SMI prevalence at LSOA level was positively associated with proximity to an urban green space with a lake (0.992 [0.99 to 0.998]). In Liverpool and Manchester, lower SMI prevalence was positively associated with road traffic noise ≥75 dB (1.012 [1.003 to 1.022]). In Birmingham, Liverpool, and Manchester, there was a positive association of SMI prevalence with distance to flood zone 3 (land within flood zone 3 has ≥1% chance of flooding annually from rivers or ≥0.5% chance of flooding annually from the sea, when flood defences are ignored): Birmingham: 1.012 [1.000 to 1.023]; Liverpool and Manchester: 1.016 [1.006 to 1.026]. In contrast, in Leeds, there was a negative association between SMI prevalence and distance to flood zone 3 (0.959 [0.944 to 0.975]). A limitation of this study was because we used a cross-sectional approach, we are unable to make causal inferences about our findings or investigate the temporal relationship between outcome and risk factors. Another limitation was that individuals who are exclusively treated under specialist mental health care and not seen in primary care at all were not included in this analysis.

Across England, the environmental characteristics associated with higher SMI prevalence at LSOA level were distance to public green space with a lake (prevalence ratio [95% credible interval]): 1.002 [1.001 to 1.003]), annual mean concentration of PM 2.5 (1.014 [1.01 to 1.019]), and closeness to roads with noise levels above 75 dB (0.993 [0.992 to 0.995]). Higher SMI prevalence was also associated with a higher percentage of people above 24 years old (1.002 [1.002 to 1.003]), a higher percentage of ethnic minorities (1.002 [1.001 to 1.002]), and more deprived areas.

We carried out a retrospective analysis of routinely collected adult population (≥18 years) data at General Practitioner Practice (GPP) level. We used data from the Quality and Outcomes Framework (QOF) on the prevalence of a diagnosis of SMI (schizophrenia, bipolar affective disorder and other psychoses, and other patients on lithium therapy) at the level of GPP over the financial year April 2014 to March 2018. The number of GPPs included ranged between 7,492 (April 2017 to March 2018) to 7,997 (April 2014 to March 2015) and the number of patients ranged from 56,413,719 (April 2014 to March 2015) to 58,270,354 (April 2017 to March 2018). Data at GPP level were converted to the geographic hierarchy unit Lower Layer Super Output Area (LSOA) level for analysis. LSOAs are a geographic unit for reporting small area statistics and have an average population of around 1,500 people. We employed a Bayesian spatial regression model to explore the association of SMI prevalence in England and its major conurbations (greater London, Birmingham, Liverpool and Manchester, Leeds, and Newcastle) with environmental characteristics (green and blue space, flood risk areas, and air and noise pollution) and socioeconomic characteristics (age, ethnicity, and index of multiple deprivation (IMD)). We incorporated spatial random effects in our modelling to account for variation at multiple scales.

The evidence is sparse regarding the associations between serious mental illnesses (SMIs) prevalence and environmental factors in adulthood as well as the geographic distribution and variability of these associations. In this study, we evaluated the association between availability and proximity of green and blue space with SMI prevalence in England as a whole and in its major conurbations (Greater London, Birmingham, Liverpool and Manchester, Leeds, and Newcastle).

Funding: JC, MJA, PAC, RJ, SLP and PCLW were supported by UK Research and Innovation Closing the Gap Network+ (ES/S004459/1). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. URL: https://www.york.ac.uk/healthsciences/closing-the-gap/ . PAC and RJ are part funded by the National Institute for Health Research (NIHR) Yorkshire and Humber Applied Research Collaboration https://www.arc-yh.nihr.ac.uk/ .

Second, we evaluated, at Lower Layer Super Output Area (LSOA) level, the association between SMI prevalence and green and blue space, man-made stressors (noise and air pollution), natural stressors (flood risk), alongside socioeconomic factors (age, ethnicity, and deprivation), and compared the associations identified across England as a whole with those identified in each of the 5 major conurbations identified by Office National of Statistics [ 21 , 22 ]: Greater London, Birmingham, Liverpool and Manchester, Leeds, and Newcastle. We hypothesised that SMI prevalence would be lower in LSOAs with greater areas of green space and woodland, with shorter distance to green and blue space, greater distance from flooding zones, with lower pollution (air and noise) and in less deprived areas, with relatively older populations, and a higher percentage of ethnical minorities populations.

Firstly, we applied a fine-resolution spatial analysis at 2 levels—England as a whole, and its major conurbations—to allow for the identification of geographic patterns and any variability in the associations in different locations. We selected major conurbations because previous research suggests that prevalence of SMI in these areas is likely to be high [ 22 ], and they are more likely to experience poor environmental characteristics such as high noise, poor air quality, and limited availability of green and blue spaces [ 23 ].

This paper evaluates the association between SMI prevalence and environmental characteristics in England. To guide our analysis, we adapted the framework developed by Zhang and colleagues [ 20 ] that combines the availability of green spaces, socioeconomic characteristics, with the context of neighbourhood, district boundaries, and urbanity, and combined the framework developed by Dzhambov and colleagues [ 21 ] (availability of blue spaces) with exposure to environmental stressors classified into man-made (air and noise pollution) and natural stressors (flood risk) according to their origin.

Studies have linked road traffic noise with negative effects on the working memory and verbal domains in people with schizophrenia [ 12 ]. The World Health Organisation (WHO) recommends that average traffic noise should be below 53 dB, with adverse health effects if above this value, and noise becomes harmful when it exceeds 75 dB [ 13 ]. An association between air pollutants, such as particulate matter with less than 2.5 μm diameter (PM 2.5 ), particulate matter with less than 10 μm diameter (PM 10 ), ozone (O 3 ), nitrogen oxides (NO x ), sulphur dioxide (SO 2 ), and health outcomes has been described often. For example, psychotic and mood disorders have been linked with long-term exposure to PM 2.5 and NO 2 [ 6 , 14 , 15 ], O 3 [ 16 ], and seasonal peaks in NO 2 [ 14 ]. The socioeconomic context of neighbourhoods also affects mental health [ 17 ]. Deprived neighbourhoods (high crime and education deprivation) have been associated with higher incidence of schizophrenia [ 18 , 19 ].

Serious mental illness (SMI), which includes schizophrenia, bipolar affective disorder, or psychosis, affects 335 million people worldwide [ 1 ] and is responsible for a significant health care burden. In England, the economic cost of SMI was estimated as £2.82 billion in 2019 [ 2 ]. People with SMI experience reduced life expectancy compared with the general population, e.g., for people diagnosed with schizophrenia, life expectancy is reduced by 13.6 (men) to 15.9 years (women) [ 3 ]. Recently, there has been a focus on the role of the environment on the risk of developing SMI. Research has shown that exposure to some air pollutants (e.g., NO x , NO 2 ) during childhood is associated with increased prevalence of schizophrenia [ 4 – 6 ], while proximity to green spaces, blue spaces, and natural areas is associated with reduced rates of schizophrenia and other SMI [ 7 – 9 ]. But the evidence is sparse regarding the associations between environment and SMI in adulthood [ 10 ], as well as the potential links between these associations and the geographic distribution of SMI, the contextual factors that may affect these patterns [ 11 ].

Methods

We investigated the association between SMI mean prevalence, socioeconomic and environmental variables, by applying a Bayesian spatial regression model with random effects. This is a cross-sectional analysis of routinely collected and publicly available primary care data. The setting is in General Practitioner Practice (GPP) in England who submitted data between 2014/2015 and 2017/2018 to the NHS Quality and Outcomes Framework (QOF). The participants are aged 18+ and registered in those GPPs.

Spatial level Data were analysed at LSOA level. LSOAs are small areas designed to be of similar size, with an average of approximately 1,500 residents or 650 households. They were produced by the Office for National Statistics for the reporting of small area statistics, like the Census. The 32,482 LSOA units across England have a mean population of 1,500 individuals and no less than 1,000 people. Distance to environmental characteristics (e.g., green space, green space with a lake) were calculated as the Euclidean distance from the LSOA-weighted population centroid that reflects the spatial population distribution from the 2011 United Kingdom Census.

Response variable: Mean serious mental illness prevalence The response variable was the mean prevalence of SMI as defined by the QOF indicator MH001—people with a diagnosis of SMI: schizophrenia, bipolar or other affective disorders, and other patients on lithium therapy [24–28]. The QOF is an incentivized voluntary process for all GPP in England and was introduced as part of the GPP contract in 2004, detailing practice achievement results. The QOF contains 4 domains: Clinical, Public Health, Public Health—Additional Services, and Quality Improvement. Each domain consists of a set of achievement measures, known as indicators, against which practices score points according to their level of achievement. GPP are incentivised as part of the QOF payments to maintain this register which makes the recording of the indicator likely to be an accurate point prevalence estimate. Individuals who are exclusively treated under specialist mental health care and not seen in primary care at all were not included in this analysis. The QOF includes on average 97% of the active GPP in England and the number of patients ranged between 56,413,719 (April 2014 to March 2015) [25] and of 58,270,354 patients for financial year April 2017 to March 2018 [28] (S2 and S3 Tables). We used QOF data on SMI prevalence for the period April 2014 to March 2018, reported at the GPP level [24–28]. For each GPP, there are also data on the LSOA of origin of its registered patients [29]. The average SMI prevalence in an LSOA is a weighted average of the prevalence in the GPP where the inhabitants of that LSOA are registered; the weights are the proportion of patients from that LSOA registered in each of the GPP [30]. The mean prevalence was then taken for each LSOA for the period between April 2014 to March 2018. We chose to analyse the mean prevalence instead of annual data to provide more power to the response variable. There were 53 LSOAs that did not have values for the SMI prevalence in 2017/2018. Their outcomes were treated as missing and were imputed based on the covariate values of these LSOAs and the estimated random effects of the middle super output area (MSOA), District, and Clinical Commissioning Group (CCG) (i.e., groups of GPP which come together in each area to commission the best services for their patients and population) within which each of these LSOAs resides (see Statistical Analysis for more detail and definitions of MSOA and CCG).

Environmental characteristics To assess the relationship between SMI prevalence and environment, we considered variables that have been associated with health: green and blue space, flood risk areas, and air and noise pollution. We derived the following variables in relation to green space: area of public green space per LSOA (ha) [31], distance to the nearest point of access of public green space (km) [31], and woodland area (ha) in each LSOA [32] (see S4 Table). Green and blue spaces are often associated with one another, and in this study, we included green spaces with water features (lakes and rivers) by measuring distance from the LSOA population-weighted centroid to the nearest public green space with a lake [33] and distance to a public green space with a river [34] (S4 Table). To calculate flood risk areas, we used the zoning with the highest probability of occurrence designated by the UK Environment Agency as Flood Zone 3 (i.e., land within this zone has ≥1% or 0.5% chance of flooding annually, from rivers and the sea, respectively) [35] (S4 Table). We measured noise pollution exposure as the distance from the LSOA population-weighted centroid to the nearest source of automobile noise ≥75 dB (S4 Table). We used the Department for Environment, Food and Rural Affairs (Defra) dataset [36], which provides the annual average road noise levels for the 16-hour period between 7 AM and 11 PM, for 2017, in the following noise classes: 55 to 59, 60 to 64, 65 to 69, 70 to 74, >75 dB [36]. These data are only available for roads within areas with a population of at least 100,000 people and along major traffic routes. Therefore, not all of England has a noise map. In order to use this variable, we made the assumption that the areas not covered by this assessment did not have automobile noise ≥75 dB. For air pollution, we used Defra’s 1 × 1 km gridded modelled annual mean PM 2.5 data for 2014 [37] (S4 Table). Defra makes use of the Automatic Urban and Rural Network, with 138 sites operating in 2014 to monitor and model at national scale PM 2.5 roadside concentration. The reason for choosing this pollutant over any other was due to the existent literature that supports an association between PM 2.5 and development of psychoses [6,14,15].

Social, demographic, and economic factors Socioeconomic variables were all measured at the LSOA level. We included ethnicity in our model since studies report that minority ethnic groups have higher incidence risk of SMI [19,34,35]. Ethnicity and age were both sourced from 2011 UK Census [38]. Ethnic minorities were measured as a percentage of the population in the following groups as identified by the 2011 UK Census [38]: Asian (Asian or British Asian), black (black, African, Caribbean, or black British), mixed (mixed or multiple ethnic groups; other ethnic groups). Adults (> = 18 years old) were split into 4 age groups: 18 to 24, 25 to 44, 45 to 64, ≥65 years old (S4 Table). We measured the percentage of the population in each age group. To assess the association of socioeconomic variables with SMI prevalence, we used the scores of 4 domains (crime, barriers to housing and services, employment deprivation, and income deprivation) and 2 subdomains (living environment—indoors; education, skills, and training—adult skills) from the Index of Multiple Deprivation (IMD) 2015 [39] (S4 Table). The smaller the score, the less deprived the LSOA is. Each set of scores was transformed into quintiles with the first quintile being the least deprived category.

Geographical variables For geographical variables, we used geographic regions and settlement categories. The region indicator for the 9 regions in England—London, the North East, North West, Yorkshire, East Midlands, West Midlands, South East, East of England, and the South West—was included as a categorical covariate, as opposed to as a set of region-level random effects, due to a small number of regions. As discussed in Statistical Analysis, our model captures spatial variability in data via random effects specified at finer spatial resolution levels. Thus, the fixed effect specification on region is sufficient to account for regional differences. We used the following settlement categories: rural town and fringe; rural town and fringe in a sparse setting; rural village and dispersed; rural village and dispersed in a sparse setting; urban city and town; urban city and town in a sparse setting; urban major conurbation; urban minor conurbation, as defined by the Office of National Statistics [40,41] (S4 Table).

Statistical analysis There was no prospective protocol and the analysis plan was as follows. To investigate the association between SMI mean prevalence, socioeconomic and environmental variables, a Bayesian spatial regression model with random effects was constructed on the log-transformed mean SMI prevalence. Our model captures complex spatial dependency structures at different spatial resolution levels using spatial random effects. The Bayesian implementation of our spatial model enables us to flexibly construct and fit realistic models to describe the variability in SMI prevalence, to assess robustness of our conclusions to various plausible model assumptions, to incorporate uncertainty associated with the data and with the model parameters. Log transformation was applied to achieve normality for the distribution of the outcome values. Let y i denote the log mean SMI prevalence of LSOA i (i = 1, …, N with N = 32482 LSOAs). Eq 1, referred to as the full model hereafter, models this outcome value y i as a function of the risk factors and a collection of random effect terms. (1) In Eq 1, β 0 is the intercept. The term x ik is the value of the kth risk factor in LSOA i so the regression coefficient, β k , is the log prevalence ratio (PR) [42], measuring the effect of that risk factor on the outcome of interest, SMI prevalence. Also included in Eq 1 are 3 spatial random effect terms, v MSOA[i] , m district[i] , and g CCG[i] , specified at the MSOA (there are 6,791 in the study region), Local Authority District (District; 326), and CCG (207) levels, respectively. Each MSOA is formed based on a group of contiguous LSOAs. Among all MSOAs in England and Wales, the mean population size is 7,200 with the minimum of 5,000. Districts are administered by either single tier (e.g., Unitary Authority, the metropolitan district, and the London borough) or 2-tier local authorities (e.g., county and the local authority district) in various parts of England. CCGs are groups of GPP that come together in each area to commission the best services for their patients and population. These 3 sets of random effects were included in the model to capture the residual variability at the 3 geographical levels that was not accounted for by the inclusion of the observable covariates. Such residual variability can arise due to unmeasured/unobservable risk factors. Finally, e i is the independent error term in the regression model and for all LSOAs. To fully specify the model in the Bayesian framework, prior distributions were assigned to the model parameters, which are the regression coefficients, the spatial random effects, and the random effect and the error precisions. The prior specifications are given as follows. For each regression coefficient, a vague normal prior with mean 0 and a variance of 1,000 (i.e., N (0, 1,000)) was assigned. The use of N (0, 1,000), in particular the large variance chosen, reflects the assumption that little is known about the association between each covariate and SMI prevalence. Therefore, the information used to estimate the regression coefficients largely comes from the data. For each set of spatial random effects, the Besag–York–Mollié (BYM) spatial prior model [43] was used. The BYM model is formulated as a sum of 2 sets of random effects, a set of spatially structured random effects and a set of spatially unstructured random effects. The spatially structured random effects are modelled via the intrinsic conditional autoregressive (ICAR) model. The ICAR model assumes that the random effects from 2 nearby spatial units at the same spatial resolution level (e.g., 2 MSOAs) are more like each other compared to the situation where these 2 spatial units are far apart. To operationalise the above idea of similarity in space, at each spatial level, we defined spatial proximity via contiguity whereby 2 areas (e.g., MSOAs) are neighbours to each other if they share a common boundary and they are not neighbours otherwise. These spatially structured random effects capture the residual variability that displays a spatial pattern. For the spatially unstructured random effects in the BYM model, the exchangeable model was used. This exchangeable specification on the random effects assumes that the effects from the unobserved/unmeasured covariates on SMI prevalence vary from one area to another but such varying effects do not display a spatial pattern. We also considered different versions of the full model, each with a different specification of the random effect component. Results on model comparison are summarised in S5 Table. Finally, a Gamma distribution, Gamma(1, 0.00005), was used as a vague prior on the error precision, , and on each of the random effect precisions associated with the BYM specification. It is worth emphasising the following 2 points on the spatial modelling. First, under the ICAR specification, while spatial contiguity defines a local neighbourhood structure, spatial smoothing under the ICAR model is not restricted to an area’s immediate neighbours but spans and propagates throughout the small areas at that spatial level [44]. Second, estimation of the spatial random effects depends not only on the spatial prior model used but also on the observed small area SMI prevalence. A strength of the Bayesian approach is that we utilise both sources of information, prior and data, to estimate model parameters. To gauge the contribution of each model component, we also fitted 2 models: the covariates only model and the random effects only model, the expressions of which are given in Eqs 2 and 3, respectively (Table 1). All terms are specified in the same way as for the full model. (2) (3) PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 1. Model comparison via DIC and WAIC. https://doi.org/10.1371/journal.pmed.1004043.t001 Model comparison was performed via deviance information criterion [45] (DIC) and Watanabe–Akaike information criterion (WAIC) [46]. Both criteria evaluate models based on goodness of fit (how well a model describes the observed data) and model complexity. A smaller DIC or WAIC value indicates a better model (Table 1). The analysis was placed within the Bayesian framework. This not only offers the flexibility to incorporate random effects at multiple spatial scales but also allows us to consider different plausible assumptions on the dependence structure of the random effects. The latter is important in terms of assessing potential sensitivity of our findings regarding the risk factor effects to different modelling assumptions. Parameter estimation for all models was carried out through the integrated nested Laplace approximation (INLA) approach via the R package R-INLA [47]. INLA, a well-established technique to implement Bayesian spatial models [48], has shown to be computationally efficient to handle the large number of spatial units (32482 LSOAs in England) and the complexity of our spatial model. Briefly, INLA obtains posterior estimation of parameters via the nested Laplace approximation, the defining feature of the method to enable fast computation for fitting complex models to large spatial datasets [49]. The Kullback–Leibler divergence (D KL ), a standard output from INLA, is a diagnostic to measure the accuracy of the INLA approximation [50]. For the full model that we shall report in the Results section, the D KL diagnostic values for all regression coefficients and all random effects were small, indicating a reliable fitting from INLA (D KL mean = 9.666e−05, D KL min = 0.000, D KL max = 1.123e−03). S1 Fig shows the posterior distributions of the random effect standard deviations. All distributions are unimodal and well behaved where the distribution is not being pushed towards 0, indicating good estimations of these parameters. For a covariate effect, we report the posterior mean and the 95% credible interval (formed using the 2.5th and the 97.5th percentiles of the posterior distribution) of the PR, i.e., exp(β k ) with β k being the regression coefficient in Eq 1. The posterior mean gives a point estimate of the covariate effect and the 95% credible interval, hereafter referred to as 95% CI, provides an interval estimate within which the “true” effect lies. An interval estimate that does not contain 1 indicates a high level of certainty (over 95% chance) that an association between the covariate in question and SMI prevalence exists—the value 1 indicates no association.

[END]
---
[1] Url: https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1004043

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/