Gac Sanit. 2022;36(6):534–539 Original Neighbourhood income inequalities in mental health in Barcelona 2001-2016: a Bayesian smoothed estimate Xavier Bartoll-Rocaa,b,∗ , Marc Marí-Dell’Olmoa,b,c , Mercè Gotsensa,b , Laia Palènciaa,b,c , Katherine Péreza,b,c , Elia Díeza,b,c , Carme Borrella,b,c,d a Agència de Salut Pública de Barcelona (ASPB), Barcelona, Spain b Institut d’Investigació Biomèdica Sant Pau (IIB-Sant Pau), Barcelona, Spain c CIBER de Epidemiología y Salud Pública (CIBERESP), Madrid, Spain d Departament de Ciències Experimentals i de la Salut, Facultat de Ciències de la Salut i de la Vida, Universitat Pompeu Fabra, Barcelona, Spain a r t i c l e i n f o a b s t r a c t Article history: Objective: Obtaining reliable health estimates at the small area level (such as neighbourhoods) using Received 25 November 2021 survey data usually poses the problem of small sample sizes. To overcome this limitation, we explored Accepted 21 March 2022 smoothing techniques in order to estimate poor mental health prevalence at the neighbourhood level Available online 26 May 2022 and analyse its profile by income in Barcelona city (Spain). Method: A Bayesian smoothing model with a logit-normal transformation was applied to four repeated Keywords: cross-sectional waves of the Barcelona health survey for 2001, 2006, 2011 and 2016. Mental health status Mental health was identified from the 12-item General Health Questionnaire. Income inequalities were analysed with Small-area analysis neighbourhood income in quantiles for each year and trends in the pooled analysis. Results: The prevalence of poor mental health ranged from 14.6% in 2001 to 18.9% in 2016. The yearly difference between neighbourhoods was 12.4% in 2001, 16.7% in 2006, 14.2% in 2011, and 20.0% in 2016. The odds ratio and 95% credible interval (95%CI) of experiencing poor mental health was 1.40 times higher (95%CI: 1.02-1.91) in less advantaged neighbourhoods than in more advantaged neighbourhoods in 2001, 1.61 times higher (95%CI: 1.01-2.59) in 2006 and 2.31 times higher (95%CI: 1.57-3.40) in 2016. Conclusions: This study shows that the Bayesian smoothed techniques allows detection of inequalities in health in neighbourhoods and monitoring of interventions against them. In Barcelona, mental health problems are more prevalent in low-income neighbourhoods and raised in 2016. © 2022 SESPAS. Published by Elsevier España, S.L.U. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Desigualdades de renta por barrios en salud mental en Barcelona 2001-2016: una estimación de alisado bayesiano r e s u m e n Palabras clave: Objetivo: Obtener estimadores de salud en áreas pequeñas (como los barrios) utilizando datos de encues- Salud mental tas supone hacer frente al problema de insuficiente tamaño muestral. Para superar esta limitación Análisis de áreas pequenas exploramos técnicas de alisado con el fin de estimar la prevalencia de mala salud mental a nivel de barrio y analizar su patrón por renta en la ciudad de Barcelona (España). Método: Se aplicó un modelo de alisado bayesiano con transformación logística-normal a cuatro muestras transversales repetidas de la Encuesta de Salud de Barcelona para los años 2001, 2006, 2011 y 2016. La salud mental fue identificada con el Cuestionario General de Salud de 12-items. Las desigualdades de ingreso se analizaron por cuantiles de la renta por barrio para cada año y las tendencias en el análisis conjunto. Resultados: La prevalencia de mala salud mental oscila entre el 14,6% en 2001 y el 18,9% en 2016. La diferencia entre barrios fue del 12,4% en 2001, del 16,7% en 2006, del 14,2% en 2011 y del 20,0% en 2016. La odds ratio y el intervalo creíble al 95% (IC95%) de experimentar mala salud mental fue 1,4 veces superior (IC95%: 1,02-1,91) en los barrios menos aventajados respecto de los más aventajados en 2001, de 1,61 (IC95%: 1,01-2,59) en 2006 y de 2,31 (IC95%: 1,57-3,40) en 2016. Conclusiones: Este estudio muestra que las técnicas de alisado bayesiano permiten la detección de desigualdades en salud a nivel de barrios para su monitorización e intervención con el fin de reducirlas. En Barcelona, los problemas de salud mental son más prevalentes en los barrios de menor renta y se incrementaron en 2016. © 2022 SESPAS. Publicado por Elsevier España, S.L.U. Este es un artı́culo Open Access bajo la licencia CC BY-NC-ND (http://creativecommons.org/licenses/by-nc-nd/4.0/). ∗ Corresponding author. E-mail address: xbartoll@aspb.cat (X. Bartoll-Roca). https://doi.org/10.1016/j.gaceta.2022.03.007 0213-9111/© 2022 SESPAS. Published by Elsevier España, S.L.U. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc- nd/4.0/). X. Bartoll-Roca et al. / Gac Sanit. 2022;36(6):534–539 535 Introduction the population from the official register stratified by district accord- ing to age and sex quotas. Non-respondents were substituted by the Monitoring indicators at the small area level has been a grow- population of the same sex and age group until the target sample ing area of research in the last few decades, allowing the detection was reached. Interviews were conducted face-to-face by trained of widening socioeconomic inequalities.1 In particular, the use of professionals with computer-assisted personal interviewing. The official registers has enabled identification of socioeconomic health sample size was computed to achieve a precision of ±2% for the inequalities in mortality in small areas across European and Span- whole city and ±6% for the 10 districts. ish cities.2,3 Estimation of health inequalities at the small areal Mental health was measured with the 12-item version of the level is of a major interest from a policy perspective as it allows General Health Questionnaire (GHQ-12). The questionnaire focuses improvement and monitoring of community health interventions. on the risk of mental disorders, occurring in the last 30 days, and is Interest in this topic has increased since the economic crisis of a widely used screening instrument to detect risk of current, diag- 2008, with systematic reviews reporting worsening in a variety nosable psychiatric disorders. It has been shown to be a valid and of health outcomes and especially in mental health trends.4,5 In reliable screening tool for the Spanish population.17 The twelve Spain, since the 2008 economic crisis, mental health has clearly questions were answered on a 4-point Likert scale (from 1 to 4). deteriorated especially among men in the lowest socioeconomic Following recommendations, we used a 2-point scoring method, position and the long-term unemployed.6 Previous evidence indi- rating a problem as absent (1 or 2) or present (3 or 4). Participants cates that the urban environment is associated with worse mental scoring 3 or more current problems were classified as having poor health, although identification of a causal link is hampered by the mental health.18 complex interactions between urban environments and individual Neighbourhood income level for 2016 was taken from estimated characteristics.7–9 However, there is scarce evidence of trends in official reports consisting of a disposable family income index inequalities at the small area level using survey data. Administra- imputed through a set of economic neighbourhood dimensions.16 tive delimitation of small areas does not generally respect an even To avoid confusion from income variations by year, all compu- distribution of population sizes so these small areas, as neighbour- tations were performed with income from 2016. Quintiles were hoods, can register too few cases than statistically recommended10 . introduced from the most advantaged (as the reference category in In this case, indirect estimates for small areas can be obtained with the interval of 119-214.7, where 100 stands for the mean income spatial information under the principle that “the closer, the more of the city) to the least advantaged (from 49.9 to 72.5). A sensitivity similar”.11 Information from contiguous areas can be borrowed to analysis was performed with yearly income. smooth the estimate, so that a higher weight is assigned to the esti- To improve the estimates, we included the centred distribu- mate of the closest areas, decreasing the problem of small sample tions of age and sex of the neighbourhood as a sociodemographic size. The most common practice is to use random area effects to variable. account for between-area variation.12 When smoothing techniques are used from surveys, small area Statistical model for smoothed estimates estimation has to deal with missing values, and take into account the sampling design used in the estimates. Fortunately, recent First, normalised post-stratification weights were computed approaches based on Bayesian smoothing techniques have been through the official population register by age and sex for each developed to account for sampling design.13,14 Any model that does neighbourhood. The correction of the sampling weights aimed to not take sampling weights into account will be rendered biased. In make the distribution of the sample across neighbourhoods as simi- the case of binary outcomes, sampling weights can be incorporated lar as possible to the population. The direct estimates of poor mental into the model by several possible transformations of the (binary health by area k, corrected with post-stratification weights (Pk ,) health) outcome variable.14 In this study, we used the logit-normal are asymptotic estimators of the true area prevalences, but it is transformation to obtain smoothed estimates of mental health at not guaranteed that the probability be constrained in the interval the neighbourhood level. Our investigation extends previous work (0,1), in particular in areas with small sample size. To overcome on small area estimation using survey data in Spain by incorpo- this limitation, a convenient logit-normal transformation consists rating for the first time a complex survey design into small area of defining the total counts of poor mental health of area k, yk , as the estimation.15 logistic of Pk or logit(Pk ).13 The interested reader can found further We restricted our analysis to Barcelona city. The city lies in the statistical details in the Supplementary Material Appendix. In this North-East coast of Spain and has 1.7 million inhabitants. To our way, the sampling design is incorporated into the estimation. To knowledge, it is the only city in Spain with data on mental health smooth the prevalences with information from the neighbourhood obtained regularly every 5 years between 2001 and 2016 through areas, we use Bayesian hierarchical models. In particular, we use the Barcelona Health Survey (BHS). The BHS collects information the model proposed by Besag-York-Mollie19 (BYM) which takes in from a representative sample of each of the 10 administrative to account two types of random effects: spatial (uk ) and heteroge- districts of the city but not for the 73 neighbourhoods. These neigh- nous (vk ). Accordingly, the logit(Pk ) can be modelled as the linear bourhoods group populations of unequal size, the largest with a combination of these two random effects as: population around 60 000 and the smallest with 500, with sharp socioeconomic differences.16 logit (Pk ) = ˇo + ˇ1 XEk + ˇ2 XSk + uk + vk (1) The objective of this study was, first, to estimate the preva- where XEk stands for (centred) area average age and Xsk for the lences of poor mental health in Barcelona at the neighbourhood (centred) proportion of women at the area level. These covariates level from 2001 to 2016 and, second, to analyse the association of were introduced to obtain better estimates and meaningful fitted mental health with neighbourhood income and its trend during this and posterior distributions for the intercept and its 95% Credibility period. Intervals (95%CI). Subscripts in equation (1) indicating yearly computations were omitted for simplicity. Prior distributions Method are assigned to the random effects, and hyper prior distribu- Population and variables tions to the parameters of the prior distributions. The structural spatial residual is modelled as an intrinsic conditional autore- Data were drawn from the 5-yearly BHS for 2001, 2006, 2011 gressive structure (iCAR) whose precision matrix (the inverse of and 2016. The BHS sample design consists of random selection of the variance–covariance matrix) is proportional to the adjacent 536 X. Bartoll-Roca et al. / Gac Sanit. 2022;36(6):534–539 neighbourhoods. The heterogenous effect is assumed to have Table 1 Summary statistics for the population and dispersion measures of direct and a normal distribution with mean 0 and variance  u . A uniform smoothed estimates of poor mental health by year. distribution U(0,∞) was assigned to the standard deviations of both random effects specified in the BYM. A normal vague prior 2001 2006 2011 2016 distribution was assigned to the parameters of the intercept and Population 1,507,346 1,610,774 1,620,292 1,608,710 coefficients. The model is scaled in the spatial variance in order to unify the interpretation of the prior and make the results Sample size 7519 2405 3210 3219 transferable between applications. Poor mental health Unlike other possible transformations apart from the logit- Total cases 1072 369 477 583 normal transformation, a different approach is the effective sample Smoothed prevalence (%) 14.6 17.5 15.7 18.9 size method. Briefly, the method consists of computing the effec- Dispersion measures tive sample (m∗k ) as the sample size that is needed to match the Total range (%) variance from a random sample with that under a complex sam- Direct estimates 0-100 0-100 0-100 0-100 ple design. Therefore, the effective sample size, rather than the Smoothed 9.9-22.3 11.2-27.9 10.6-24.8 11.1-31.1 actual sample size, acknowledges the variable information that Interquartile range (%) each individual supply under complex sampling.13,20 Accordingly, Direct estimates 10.5-17.9 5.9-20.4 7.3-18.8 8.8-24.8 the effective number of cases can be computed as yk∗ = m∗k · Pk . In Smoothed 12.7-16.2 14.9-19.6 13.9-16.9 16.1-21.2 this paper, we show smoothed prevalence estimates applying the Coefficient of variation logit-normal transformation and used the effective sample method Direct estimates 92.2 84.9 105.5 60.0 Smoothed 17.1 19.6 17.9 21.6 as a sensitivity analysis. Statistical model for the association of income and trend with poor mental health To analyse the association of mental health with neighbourhood income for each year, we simply added to equation (1) the quintiles of neighbourhood income IQk introduced as dummy variables with the reference category as the most advantaged neighbourhood (in bold the vector of coefficients) according to: log it (Pk ) = ˇo + ˇ1 XEK + ˇ2 XSk + ˇ3 IQk + uk + vk (2) where the subscript for time is omitted for simplicity. To assess the profile of inequalities in mental health by income, Figure 1. Boxplot of year-specific smoothed neighbourhood estimated prevalence we used a spatio-temporal model with the pooled data across years. of poor mental health. The temporal dimension is introduced with a common trend and random area effects to allow for deviations from trend in equation what extent these limitations may affect the results, a sensitivity (2). An interaction term with time trend, as a continuous variable, analysis has been carried out using the BYM2 model.21 and the neighbourhood income accounts for deviations from trend To fit all models, we used the Bayesian framework with the by income quantiles, and the complete specification follows the Integrated Nested Laplace Approximation computation INLA.22 All expression: computations were performed with R programme version 4.0.04. log it (Pk ) = ˇo + ˇ1 XEk + ˇ2 XSk + ˇ3 IQk Results + (ˇ4 + dk )t + ˇ5 IQk t + uk + vk (3) Prevalence estimates of poor mental health where ˇ4 stands for the coefficient of the common trend, dk stands for the area random effect assumed to be distribute as a Normal (0, Due to the sparse number cases in some areas, the direct esti-  d ) and ˇ5 the interaction term between income and time. Analo- mates of poor mental health shown in Table 1 ranged between of gous to the first model, a uniform distribution U (0,∞) is assigned 0% and 100% for each year, which is problematic. In contrast, the to the standard deviations of the random effects (including that smoothed estimates showed a smaller range, between 9.9% and of time) and a normal vague prior distribution to the parameters 31.1%. The reduction in the dispersion of the direct estimates are of the intercept and coefficients. The model is also scaled in the also shown in the narrower interquartile range and the reduction spatial variance. We report selection model measures: the Akaike of the coefficient of variation in the order of between one-fifth and information criteria (AIC), the deviance information criteria and the almost 3 times lower. The distribution was more concentrated in conditional predictive ordinate, with smaller values meaning bet- 2001 and 2011, although for 2001 this could be due to the larger ter fit. Other possible specifications of time that relax the linearity sample size. The interquartile difference ranged from 3% in 2011 hypothesis in (3) consists of also smoothing time dynamically with to 5.1% in 2016. The maximum difference in poor mental health the prior and or subsequent observation and even its interaction between neighbourhoods was fairly constant and was greater than with the spatial and random terms. We assessed the performance 12% except in 2016, when it increased: 12.4% in 2001, 16.7% in of some of these models with the selection model criteria but the 2006, 14.2% in 2011 and 20.0% in 2016. For these years, the preva- improvement in these measures were negligible compared with lence of poor mental health and 95%CI were 14.6% (13.1%,15.6%), our benchmark model so they were excluded (results available on 17.5% (15.3%,19.3%), 15.7% (13.6%,17.6%), and 18.9% (16.5%,20.8%), request). respectively. Estimation based on the classical BYM model has been criticised Year-specific smoothed estimates of poor mental health with due to the problem of non-identifiability between the spatial ran- the logit-normal transformation are shown in Figure 1. Larger dis- dom effect and the heterogeneous effect. In order to evaluate to persion was observed in 2006 and 2016 with poor mental health X. Bartoll-Roca et al. / Gac Sanit. 2022;36(6):534–539 537 Table 2 Odds ratios and 95% credible intervals of poor mental health and by year and the pooled estimate. 2001 2006 2011 2016 Pooled Age 1.04 (0.99-1.09) 1.01 (0.97-1.06) 0.98 (0.94-1.03) 1.00 (0.97-1.04) 1.01 (0.99-1.03) Men Reference Reference Reference Reference Reference Women 1.64 (0.54-4.99) 2.29 (0.96-5.68) 1.21 (0.41-3.61) 1.17 (0.44-3.17) 1.57 (1.03-2.39) Income IQ 1 Reference Reference Reference Reference Reference IQ 2 1.04 (0.77-1.39) 1.03 (0.63-1.66) 0.85 (0.55-1.33) 1.33 (0.90-1.97) 0.99 (0.80-1.22) IQ 3 1.17 (0.86-1.60) 1.29 (0.78-2.12) 0.97 (0.61-1.53) 1.36 (0.89-2.07) 1.14 (0.90-1.43) IQ 4 1.29 (0.95-1.75) 1.33 (0.81-2.18) 1.05 (0.64-1.68) 2.10 (1.42-3.08)a 1.19 (0.95-1.48) IQ 5 1.40 (1.02-1.91)a 1.61 (1.01-2.59)a 1.04 (0.62-1.69) 2.31 (1.57-3.40)a 1.34 (1.07-1.66)a Time (t) 1.00 (0.91-1.10) IQ 2*t 1.07 (0.93-1.22) IQ 3*t 1.05 (0.91-1.20) IQ 4*t 1.15 (1.01-1.31)a IQ 5*t 1.14 (0.99-1.30) AIC 92.2 160.1 150.9 132.6 571.1 DIC 94.3 162.6 154.4 132.8 553.8 CPO −3.75 −3.33 −3.32 −3.53 −5.03 AIC: Akaike information criteria; CPO: conditional predictive ordinate; DIC: deviance information criterion; IQ: income quantile from the most advantaged IQ1 to the least advantaged IQ5. a Significant at p-value 0.05 or lower. tending to increase except in 2011 when it returned to the 2001 estimates successfully reduced variability and provided a reliable level. A similar profile was obtained in the sensitivity analysis with picture of mental health inequalities. We observed inequalities the effect size method (see Fig. I in online Appendix). No specific in poor mental health across neighbourhoods (except for 2011), spatial pattern was observed for the smoothed estimates before which increased in 2016. the economic crisis of 2008, but for 2011 and 2016 worse men- A profile of inequalities that decreased during the economic tal health tended to persist in the low-income neighbourhoods crisis in 2011 but widened again in 2016 was also found in of the north-east and south-west of the city (see Fig. II in online England.23 The decrease in inequalities during the economic cri- Appendix). sis depended on whether the dominant response was a coping response or a stressful one.24 For instance, a pioneering study Association and trends of neighbourhood income with poor investigating the relationship between economic conditions and mental health mortality rates during the period 1972-1991 argued that a larger proportion of the population may acquire healthier habits, such Income inequalities for the least advantaged neighbourhoods as more hours of sleep, more physical exercise, less demanding for all years, except 2011, are shown in Table 2. The odds ratio (OR) work and more time for leisure, leading to a reduction in health and 95%CI of experiencing poor mental health was 1.40 (95%CI: inequalities.25 Conversely, the Spanish labour reform of 2012 1.02-1.91) times higher in less advantaged compared with more may have helped to prolong the adverse effects of the economic advantaged neighbourhoods in 2001, with OR 1.61 (95%CI: 1.01- downturn through higher unemployment rates, worse working 2.59) times higher in 2006 and OR 2.31 (95%CI: 1.57-3.40) times conditions, and greater job insecurity, leading to poor prospects higher in 2016 for the fifth poorer income quintile IQ5 and OR 2.10 for younger people and disproportionately affecting the population (95%CI: 1.42-3.08) for IQ4. A visual representation of this profile of with a lower socioeconomic position.26 income inequalities is reported in Fig. III in online Appendix, where As stressed by the recent above-mentioned literature,9,27 indi- a steeper gradient was observed in 2006 and 2016. vidual determinants of mental health probably have stronger The issue of whether income inequalities by quartiles has nar- mechanisms than neighbourhood settings. However, the relentless rowed or expanded with time is captured with the interaction effects of the economic crisis of 2008 have exacerbated divergence terms between income quartiles and time in the pooled analysis in in neighbourhood vulnerability. It has been suggested that mecha- Table 2. The interaction was positive and significant for less advan- nisms of divergence may be greater in countries, such as in Spain, taged neighbourhoods, quartile IQ4, with OR 1.15 (95%CI 1.01-1.31) with a higher proportion of home-ownership and with a tight rental and almost significant for IQ5 OR 1.14 (95%CI 0.99-1.30) but not housing market. In this situation, it is more difficult for low- to for the other neighbourhood quartiles. However, the lower value middle-income groups and first home buyers to obtain a mortgage obtained by the sum of the goodness of the fit measures of the than in the past, while the financial effort of renting increases, so partial (yearly) models compared to the value of the pooled model that this population is forced to move to more deprived areas.28 with trend indicates the poorer performance of this last model, so In particular, evictions have been a major problem in Barcelona. that a conclusion on a widening inequality gap over time should be From the first registers in 2013 to 2016, 12,333 evictions have been avoided. Similar results were obtained in the sensitivity analysis recorded, which may have contributed to increasing inequalities in considering income for each year (data not shown). The sensibil- the 2016 wave, as evicted persons report poor mental health up to ity analysis applying BYM2 gave also equivalent results (Appendix) 6 times more frequently than the general population.29 but being the interaction term in pooled analysis significant also Our result agrees with recent cross-sectional studies report- for the IQ5. ing weak evidence in showing increasing inequalities in the risk of psychosocial problems among neighbourhoods.30 Although the Discussion Neighbourhood Law was withdrawn in 2011 during the economic crisis, Barcelona has continued to carry out the municipal pro- In this study, we obtained small area estimates (at the neigh- gramme Health in the neighbourhoods aimed at reducing health bourhood level) for poor mental health incorporating a complex inequalities. The programme reached 13,600 people among 25 of survey design through the logit-normal transformation. Smoothed the most disadvantaged neighbourhoods in the city. Participants in 538 X. Bartoll-Roca et al. / Gac Sanit. 2022;36(6):534–539 these community interventions ranked mental health problems as Transparency declaration among the first priorities needing to be tackled.31 Our study cannot be interpreted as a direct evaluation of this programme but it does The corresponding author on behalf of the other authors guar- provide indicators on the geographical scale in which the inter- antee the accuracy, transparency and honesty of the data and ventions are carried out and allows interventions to be focused information contained in the study, that no relevant information on the neighbourhoods most requiring them without neglecting has been omitted and that all discrepancies between authors have the importance of individual effects. Because some authors have been adequately resolved and described. found only weak evidence of area-based interventions to avoid the adverse “neighbourhood effect” on mental health,32–34 a mix Authorship contributions of individual-level targeted and area policy is likely to be needed. As stated, the association observed cannot be interpreted as X. Bartoll-Roca and M. Marí-Dell’Olmo contributed to the plan- causal, due to the repeated cross-sectional design rather than lon- ning, conception, design and implementation of the study. M. gitudinal or instrumental variable techniques, and to the lack of a Marí-Dell’Olmo contributed to the finalisation and analysis of the proper measure of the neighbourhood environment. A limitation of data. X. Bartoll-Roca drafted the manuscript. M. Marí-Dell’Olmo, our estimated models to further explore is the potential confound- M. Gotsens and Laia Palència contributed to methodological revi- ing between fixed and random effects. Spatial confounding occurs sion of the manuscript. K. Pérez and E. Díez contributed to critical when covariates have a spatial pattern and are collinear with spatial discussion, and C. Borrell to comments on the neighbourhood envi- random effects. To overcome this problem, it is proposed to spec- ronment and to the final version of the manuscript. ify random effects as orthogonal to fixed effects.35 However, while adding spatially correlated random effects does not adjust the fixed effects for the missing covariates, it does smooth the fitted values. Funding Another limitation of our study is that we were unable to stratify by sex due to the small sample size in some areas. In the context of None. this small sample size, the results should be interpreted with cau- tion if the adjacent neighbourhoods differ in their socioeconomic Conflicts of interest and environmental settings. In this case, the smoothed estimates could deviate from real values. None. Availability of databases and material for replication Appendix A. Supplementary data Data available to persons on request to xbartoll@aspb.cat from Supplementary data associated with this article can be found, in the Institut d’Investigació Biomèdica Sant Pau, Barcelona (Spain), in the online version, at doi:10.1016/j.gaceta.2022.03.007. accordance with the data policy applicable to the Barcelona health survey. Statistical codes to the Bayesian estimates are provided in the References online Appendix. 1. OECD. Making cities work for all: data and actions for inclusive growth. Paris: OECD; 2006. 2. Borrell C, Rodríguez-Sanz M, Pasarín MI, et al. Socioeconomic inequalities in mortality in 16 European cities. Scand J Public Health. 2014;42:245–54. 3. Marí-Dell’Olmo M, Gotsens M, Palència L, et al. Trends in socioeconomic inequal- ities in mortality in small areas of 33 Spanish cities. BMC Public Health. What is known about the topic? 2016;16:663. 4. Parmar D, Stavropoulou C, Ioannidis JPA. Health outcomes during the 2008 Bayesian smoothing techniques can provide reliable esti- financial crisis in Europe: systematic literature review. BMJ. 2016;354:i4588. mates of the prevalence of poor mental health at the small 5. Frasquilho D, Matos MG, Salonna F, et al. Mental health outcomes in times area level. Common challenges to obtaining reliable estimates of economic recession: a systematic literature review. BMC Public Health. 2016;16:115. from health surveys are low sample sizes in small areas and 6. Urbanos-Garrido RM, López-Valcárcel BG. The influence of the economic crisis complex designs on the association between unemployment and health: an empirical analysis for Spain. Eur J Health Econ. 2015;16:175–84. What does this study add to the literature? 7. Truong KD, Ma S. A systematic review of relations between neighborhoods and mental health. J Ment Health Policy Econ. 2006;9:137–54. 8. Mair C, Diez Roux AV, Galea S. Are neighbourhood characteristics associated with The smoothed estimates substantially reduced variabil- depressive symptoms? A review of evidence. J Epidemiol Community Health. ity compared with direct estimates and allowed detection 2008;62:940–6. of socioeconomic inequalities between neighbourhoods that 9. Baranyi G, Sieber S, Cullati S, et al. The longitudinal associations of perceived remained during the study period with some oscillations neighborhood disorder and lack of social cohesion with depression among adults aged 50 years or older: an individual-participant-data meta-analysis from 16 What are the implications of the results? high-income countries. Am J Epidemiol. 2020;189:343–53. 10. Rao Molina JNK, Molina I. Small area estimation. 2nd ed. New Jersey: Wiley Series in Survey Methodology; 2015. Territorial inequalities can be monitored with survey data 11. Tobler WR. A computer movie simulating urban growth in the Detroit region. in a small sample context. A deepening of individual-level and Econ Geogr. 1970;46:234–40. area-based targeted interventions are recommended to reduce 12. Fay RE, Herriot RA. Estimates of income for small places: an application of James- these inequalities. Stein procedures to census data. J Am Stat Assoc. 1979;74:269–77. 13. Mercer L, Wakefield J, Chen C, et al. A comparison of spatial smoothing methods for small area estimation with sampling weights. Spat Stat. 2014;8:69–85. 14. Watjou K, Faes C, Lawson A, et al. Spatial small area smoothing models for handling survey data with nonresponse. Stat Med. 2017;36:3708–45. 15. Montoya I, Esnaola S, Calvo M, et al. Estimación de indicadores de salud en Editor in charge áreas pequeñas a partir de datos de la Encuesta de Salud de Euskadi. Gac Sanit. 2019;33:289–92. 16. Martí C, Casas E, Jorge Juan A, et al. Distribució territorial de la renda familiar Miguel Ángel Negrín Hernández. per càpita a Barcelona. Barcelona: Ajuntament de Barcelona; 2007. X. Bartoll-Roca et al. / Gac Sanit. 2022;36(6):534–539 539 17. Sánchez-López MDP, Dresch V. The 12-item general health questionnaire (GHQ- 27. Gruebner O, Rapp MA, Adli M, et al. Cities and mental health. Deutsches 12): reliability, external validity and factor structure in the Spanish population. Arzteblatt International. 2017;114:121–7. Psicothema. 2008;20:839–43. 28. Zwiers M, Bolt G, Van Ham M, et al. The global financial crisis and neighborhood 18. Goldberg D. Manual of General Health Questionnaire. Windsor, UK; 1978. decline. Urban Geogr. 2016;37:664–84. 19. Besag J, York J, Mollié A. Bayesian image restoration, with two applications in 29. Vásquez-Vera H, Rodríguez-Sanz M, Palència L, et al. Foreclosure and health in spatial statistics. Ann Inst Stat Math. 1991;43:1–20. Southern Europe: results from the Platform for People Affected by Mortgages. J 20. Chen C, Wakefield J, Lumely T. The use of sampling weights in Bayesian hier- Urban Heal. 2016;93:312–30. archical models for small area estimation. Spat Spatiotemporal Epidemiol. 30. Loureiro A, Santana P, Nunes C, et al. The role of individual and neighborhood 2014;11:33–43. characteristics on mental health after a period of economic crisis in the lis- 21. Simpson D, Rue H, Riebler H, et al. Penalising model component complex- bon region (Portugal): a multilevel analysis. Int J Environ Res Public Health. ity: a principled, practical approach to constructing priors. Statistical Science. 2019;16:2647. 2017;1:1–28. 31. Daban F, Pasarín MI, Borrell C, et al. Barcelona Salut als Barris: twelve years’ 22. Rue H, Martino S, Chopin N. Approximate Bayesian inference for latent Gaussian experience of tackling social health inequalities through community-based models by using integrated nested Laplace approximations. J R Stat Soc Ser B Stat interventions. Gac Sanit. 2021;35:282–8. Methodol. 2009;71:319–92. 32. Andersson R, Musterd S. Area-based policies: a critical appraisal. Tijdschr voor 23. Thomson RM, Niedzwiedz CL, Katikireddi SV. Trends in gender and socioe- Econ en Soc Geogr. 2005;96:377–89. conomic inequalities in mental health following the Great Recession and 33. Thomson H. A dose of realism for healthy urban policy: lessons from area-based subsequent austerity policies: a repeat cross-sectional analysis of the Health initiatives in the UK. J Epidemiol Community Health. 2008;16:2647. Surveys for England. BMJ Open. 2018;8:e022924. 34. Moore THM, Kesten JM, López-López JA, et al. The effects of changes to the built 24. Catalano R, Goldman-Mellor S, Saxton K, et al. The health effects of economic environment on the mental health and well-being of adults: systematic review. decline. Annu Rev Public Health. 2011;32:431–50. Health and Place. 2018;53:237–57. 25. Ruhm CJ. Are recessions good for your health? Q J Econ. 2000;115:617–50. 35. Reich BJ, Hodges JS, Zadnik V. Effects of residual smoothing on the posterior of 26. Bartoll-Roca X, Palència L, Gotsens M, et al. Socioeconomic inequalities in self- the fixed effects in disease-mapping models. Biometrics. 2006;4:1197–206. assessed health and mental health in Barcelona, 2001-2016. Gac Sanit. 2021 Mar 23;S0213–9111, 00051-0. doi: 10.1016/j.gaceta.2021.02.009. Online ahead of print.