Article Text

Download PDFPDF

Does income inequality get under the skin? A multilevel analysis of depression, anxiety and mental disorders in São Paulo, Brazil
  1. Alexandre Dias Porto Chiavegatto Filho1,2,
  2. Ichiro Kawachi2,
  3. Yuan Pang Wang3,
  4. Maria Carmen Viana3,
  5. Laura Helena Silveira Guerra Andrade3
  1. 1Department of Epidemiology, School of Public Health, University of São Paulo, São Paulo; Brazil
  2. 2Department of Social and Behavioral Sciences, Harvard School of Public Health, Harvard University, Boston, Massachusetts, USA
  3. 3Institute of Psychiatry, University of São Paulo Medical School, São Paulo, Brazil
  1. Correspondence to Dr Alexandre Dias Porto Chiavegatto Filho, Department of Epidemiology, School of Public Health, University of São Paulo, Av. Dr Arnaldo, 715, São Paulo, SP 01246-904, Brazil, alexdiasporto{at}


Objective Test the original income inequality theory, by analysing its association with depression, anxiety and any mental disorders.

Methods We analysed a sample of 3542 individuals aged 18 years and older selected through a stratified, multistage area probability sample of households from the São Paulo Metropolitan Area. Mental disorder symptoms were assessed using the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) criteria. Bayesian multilevel logistic models were performed.

Results Living in areas with medium and high-income inequality was statistically associated with increased risk of depression, relative to low-inequality areas (OR 1.76; 95% CI 1.21 to 2.55, and 1.53; 95% CI 1.07 to 2.19, respectively). The same was not true for anxiety (OR 1.25; 95% CI 0.90 to 1.73, and OR 1.07; 95% CI 0.79 to 1.46). In the case of any mental disorder, results were mixed.

Conclusions In general, our findings were consistent with the income inequality theory, that is, people living in places with higher income inequality had an overall higher odd of mental disorders, albeit not always statistically significant. The fact that depression, but not anxiety, was statistically significant could indicate a pathway by which inequality influences health.


Statistics from


An extensive literature on the social determinants of health suggests that neighbourhood environment is an important contributor to the health of its residents. Most of the evidence has been provided by the study of the socioeconomic characteristics of area of residence (eg, neighbourhood poverty), which has been shown to affect health outcomes including high blood pressure, obesity, low birth weight, respiratory function and overall cause-specific, mortality, among others.1 ,2

Over and above the influence of neighbourhood poverty, research has also examined the potential impact of income inequality in health. In this case, the maldistribution of income in an area, that is, the income difference between its residents, could have an independent adverse effect on health, above and beyond the effect of absolute income deprivation. This theory, known as the ‘relative income hypothesis’, originally set forth by Wilkinson,3 has been tested in a large variety of settings. A meta-analysis performed by Kondo et al,4 that included 29 multilevel studies and over 60 million participants, concluded that income inequality was associated with a modest excess risk of premature mortality and poor self-rated health.

Various hypotheses have been suggested to explain the effect of income inequality on health.5 According to the original theory proposed by Wilkinson,3 status competition and class differentiation present in a highly unequal society lead to psychological distress. Individuals on the lower end of the socioeconomic ladder will develop a sense of inferiority and shame, affecting a variety of health behaviours and outcomes.6

This hypothesis has been somewhat supported by studies that analysed underlying causes of death. A study of US states found that income inequality was associated with higher mortality rates for causes that are associated with stress and social dysfunction such as high blood pressure, metabolic syndrome and excess alcohol consumption.7 A similar result was found by Chiavegatto Filho et al8 when analysing neighbourhoods of the city of São Paulo, Brazil.

Nevertheless, the specific pathway by which income inequality is expected to affect health, that is, by the development of psychological distress, has not been confirmed. Studies that tested this hypothesis have provided results that are, at best, inconclusive. An analysis by Henderson et al9 of US residents found that, after controlling for individual-level variables and state median income, depressive symptoms were not associated with state income inequality. Another study performed in the USA, of 60 metropolitan areas, found no association between local income inequality and depressive or anxiety disorders.10 A recent study of deprived areas of Glasgow found an association between mental well-being and perceived relative standard of living.11 An analysis of a Swedish sample found that the association between anxiety and neighbourhood proportion of individuals with low income disappeared after individual factors were account for.12

Ecological studies that analysed larger regions, on the other hand, have found more consistent results. A comparison of 12 countries found a statistically significant association between income inequality and prevalence of mental illness.13 An analysis of 45 US states performed by Messias et al14 found a significant positive correlation between income inequality and prevalence of depression.

According to Subramanian and Kawachi,15 multilevel analysis is recommended for income inequality and health studies, as a robust method of dealing simultaneously with the contextual effects of income inequality while at the same time addressing the compositional effects of individual income on area-level income inequality. The inclusion of individual characteristics besides income, such as sex, age, education and marital status is also important as these are well-known correlates of individual health outcomes, independent of local income inequality.16

The analysis of Brazil brings important insights to the income inequality literature because, despite recent improvements, the country still has one of the highest levels of inequality in the world, considerably higher than other developed countries,17 where most of the studies have been performed so far. There is also interest in analysing small geographic scales such as municipalities or metropolitan areas, as results found for smaller areas have been frequently failed to corroborate the ‘relative income hypothesis’.18 The use of alternative geographic scales may help to pin down which social comparisons are relevant for individual health. One possibility is that although individuals most frequently compare their economic situation to others who are closest to them (in geographic space), the differences may not be large because of residential segregation—that is, the poor tend to cluster close to others who are poor. On the other hand, social comparisons at broader levels of aggregation (eg, a resident of a favela comparing their living standards to the average of the metropolitan area) could generate starker contrasts, and hence feelings of frustration, envy and shame.

For the present study, we analysed the results from the São Paulo Megacity Mental Health Survey (SPMHS), the Brazilian segment of the World Mental Health (WMH) Survey Initiative, a multicentric study coordinated by the WHO and performed in 28 countries. São Paulo was found to have one of the highest prevalence of mental disorders (10% of residents suffered from recently active severe mental disorder),19 which provides interest in identifying its determinants.

The objective of the present study was to test the association of individual symptoms of mental disorders and area-level income inequality, after controlling for individual characteristics. We aim to test this relationship first for the presence of any mental disorder in the last 12 months, and then for anxiety and depression separately (the most prevalent categories of disorders identified by the study).


The SPMHS is a representative sample survey of individuals aged 18 years and older living in the São Paulo Metropolitan Area (SPMA), which is composed by the city of São Paulo (11 104 715 residents in 2007) and its 38 surrounding municipalities (8 844 543 residents).19 Interviews were performed from May 2005 to May 2007 by professional interviewers who received standardised training. The SPMHS procedures for recruitment, obtaining informed consent and protecting human participants were approved by the Research and Ethics Committee of the University of São Paulo Medical School. A detailed description of the sampling and weighting methods is presented elsewhere.20

Respondents were selected from a stratified multistage clustered area probability sample of the non-institutionalised civilian population living in private households in the SPMA. Six stages of selection were employed to target 5000 households from the two geographic strata that compose the SPMA: the City of São Paulo and the remaining 38 municipalities. The Primary Sampling Unit (PSU) was either an individual city or, in the case of the Municipality of São Paulo, one of its boroughs. The second stage of selection was the census units (CU), the smallest unit for which census data are available. In the third stage, CUs were clustered according to geographic proximity, situated within each PSU. In the fourth stage, one CU was randomly selected within each cluster. In stage 5, one block from within each CU was randomly selected using a large-scale road map of the area. In the final stage, the interviewer obtained a household listing of all residents from a household informant. For each household, one resident was randomly selected by the use of a Kish table.

Initially, 7700 households were selected to achieve the planned sample of 5000 participants, allowing for a 35% non-response rate. The initial sample was composed of 5037 individuals (response rate of 81.3%). Of these, 37 (0.73%) were excluded because of incomplete or missing information on place of residence. For the present study, we included only participants with five or more years of residence in the current address to control for health selection bias.21 A total of 3542 individuals were included in the analysis. Area of residence was categorised by municipalities (average of 232 751 residents in 2007), or, in the case of residence in city of São Paulo, by its 31 administrative regions (known as subprefeituras, average of 355 467 residents), totalling 69 areas included in the models.

The questionnaire was administered in face-to-face interviews conducted by professional interviewers, assisted and supervised by the academic research team. The instrument used for all participants was the Composite International Diagnostic Interview (WMH-CIDI), a fully structured lay interview that assesses mental disorders according to the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) criteria, translated and adapted to the Brazilian-Portuguese language.22 DSM-IV was the most recent version of DSM at the time the multicentric study was standardised. Sections about sociodemographical information were also included in the questionnaire. DSM-IV disorders were identified by taking into account the previous 12 months. Individuals were selected as having any mental disorder based on DSM-IV assessment of at least one core disorder: major depression, mania, panic disorder, specific phobia, social phobia, agoraphobia, generalised anxiety disorders (GAD), adult separation anxiety (ASA), substance use disorders, intermittent explosive disorder (IED), attention-deficit/hyperactivity disorder, dysthymia, bipolar disorder I and II, oppositional-defiant disorder, obsessive-compulsive disorder (OCD), conduct disorder and post-traumatic stress disorder (PTSD), as well as suicidal behaviour. Anxiety disorders were based on the DSM-IV diagnosis of at least one of the following: panic disorder, GAD, agoraphobia without panic disorder, specific phobia, social phobia, PTSD, OCD and ASA. Depression was based on the presence of a major depression episode according to DSM-IV criteria. Any mental disorder meant that the individual presented at least one recently active DSM-IV/CIDI disorder under study.

We first controlled for individual demographic (sex and age) and socioeconomic (education, income and marital status) characteristics. Education was defined, according to the Brazilian system, into basic (8 years or less), high school (9–11 years) and college (12 years or more). Marital status was defined as single (never married), married and previously married (widowed/divorced). Individual income and income inequality (measured by the Gini coefficient of area of residence) were categorised into low, medium and high by tertiles, due to non-linearity. The Gini coefficient is a measure of statistical dispersion, that ranges from 0.0 (perfect equality, with every household earning exactly the same) to 1.0 (absolute inequality, with a single household earning the locality’s entire income). Mathematically, it is equivalent to half the average absolute difference between the incomes of any two households randomly sampled for a population, and then normalised to the mean. For the present study it was calculated using the results from the 2010 Census.23 The analysis was performed separately for any mental disorder (at least one recently active disorder under study), and for depression and anxiety (the two most prevalent categories of disorders identified by the study).

We estimated the logistic models by using a multilevel methodology, with individuals (level 1) and municipalities/subprefeituras (level 2). Several benefits have been associated with the multilevel (or hierarchical) approach, such as allowing the simultaneous examination of the effects of group-level and individual-level predictors, and accounting for the non-independence of observations within groups.24 We applied Bayesian inference to estimate the parameters by using Markov Chain Monte Carlo, as it has been shown to decrease bias in binary models,25 and because it allows to test the goodness-of-fit of models by comparing the deviance information criterion (DIC).

The models were estimated equally for the three dependent variables (depression, anxiety and any mental disorder). First, we fitted a null model (M0) to test area-level variance. Then we included the demographic variables (M1), followed by a model with the inclusion of socioeconomic characteristics (M2) and then income inequality (M3). The coefficients were presented in terms of ORs to facilitate interpretation.

The models were estimated by using MLwiN V.2.25 software. We first calculated maximum-likelihood estimates for starting values of the distribution. The first 500 simulations were discarded as ‘burn-ins,’ followed by 10 000 further estimations.26 We also calculated the intraclass correlation coefficient using the latent variable method to analyse the proportion of the total variance attributed to the area level.27


Table 1 presents the descriptive analysis of the sample, where 55.99% were women, 42.32% were 40–59 years old, 53.87% had only basic education and 51.78% were married. The average values for low, medium and high individual income were, respectively, US$ 1851, U$ 5045 and US$ 20 040, and for income inequality (Gini coefficient), 0.18, 0.23 and 0.34, respectively. In comparison with the SPMA population, the study sample was more female (56% vs 53%) and older (16% vs 11% of individuals 60 years old and older). From the sample of 3542 individuals, 958 (27.05%) were identified as having any mental disorder, 370 (10.45%) depression and 541 (15.26%) anxiety in the last 12 months. Simple logistic regressions showed that women had statistically significant (p<0.05) higher odds of presenting any of the three categories of mental disorders. Medium-income inequality was statistically associated with higher odds of any mental disorders and depression, but not anxiety. High-income inequality was only associated with higher odds of depression.

Table 1

Descriptive characteristics for the complete sample and individuals diagnosed with any mental disorder, depression and anxiety: São Paulo Megacity Mental Health Survey, 2005–2007

Multilevel models for any mental disorders (ie, at least one recently active DSM-IV/CIDI disorder under study) are presented in table 2. The null model (M0, without the inclusion of control variables) did not present a significant area-level variance (p>0.05). The inclusion of demographic variables (sex and age) decreased the DIC from 4130.75 to 4059.58, indicating better fit. Men and individuals 60 years old or older had lower odds of any mental disorder, in relation to women and individuals aged 18–39 years old (0.58; 95% CI 0.49 to 0.67, and 0.55; 95% CI 0.43 to 0.71, respectively). The inclusion of individual income, education and marital status did not bring significant change to the model (M2). In the last model (M3), we introduced area-level income inequality. The direction was as expected by the ‘relative income hypothesis’, with higher inequality being associated with higher odds of mental disorder. This association was statistically significant for medium (1.32; 95% CI 1.03 to 1.68), and marginally significant for high-inequality areas (1.24; 95% CI 0.97 to 1.57), relative to low-inequality areas.

Table 2

Multilevel models for any mental disorder: São Paulo Megacity Mental Health Survey, 2005–2007

Table 3 presents the results for depression. Again, men and older individuals (60 years old or older) had lower odds of depression compared with women and younger individuals. Higher income inequality was also associated with increased risk of depression; both medium-inequality areas and high-inequality areas were statistically associated with increased risk of depression, relative to low-inequality areas (1.76; 95% CI 1.21 to 2.55, and 1.53; 95% CI 1.07 to 2.19, respectively).

Table 3

Multilevel models for depression: São Paulo Megacity Mental Health Survey, 2005–2007

Table 4 presents the results for anxiety disorders. Men presented lower odds of having anxiety, but in this case individuals 40–59 years old had statistically higher odds in M1 (1.31; 95% CI 1.07 to 1.60), and close to significant in the other models. Income inequality had the same direction as the previous two analyses, that is, higher inequality was associated with higher odds of presenting the disorders, but differently from the previous two, this association was not statistically significant (1.25; 95% CI 0.90 to 1.73 for medium, and 1.07; 95% CI 0.79 to 1.46 for high).

Table 4

Multilevel models for anxiety disorder: São Paulo Megacity Mental Health Survey, 2005–2007


After controlling for sex, age, income, education and marital status, higher local-income inequality was statistically significantly (p<0.05) associated with higher odds of presenting depression, but not anxiety. For both of the specific categories of disorders, as well as for any mental disorder (at least one recently active DSM-IV/CIDI disorder under study), our findings are consistent with the income inequality theory, that is, people living in places with higher income inequality had overall higher odds of mental disorders, albeit not always statistically significant.

The use of a multilevel approach was important to control for the diminishing marginal effects of income on health (the ‘concavity effect’), where individuals with very high income will benefit less from an increase in income. According to Subramanian and Kawachi,15 a multilevel analysis is necessary to distinguish the concavity effect from the independent contextual (or ‘real’) effect in income inequality analyses.

Studies regarding mental disorders are important to test the pathway proposed by the original ‘relative income hypothesis’, which states that local inequality will lead to adverse social comparisons associated with psychological distress, leading to a higher propensity of diseases and mortality. According to the theory, income inequality will ‘get under the skin’,6 affecting mental health and eroding social trust.28 ,29 Nevertheless, despite an increase in studies that analysed income inequality in the last decade, income inequality is still considered ‘an under acknowledged source of mental illness and distress’.30

We found that, while depression and anxiety presented the same direction in its association with inequality (ie, the higher the inequality, the higher the odds of presenting the disorder), depression was statistically significant, but anxiety was not. This could be an indicator of how income inequality affects the presence of mental disorders, as depression has been frequently conceptualised as a ‘backward-looking’ emotion, and anxiety as ‘forward-looking’.31 In the particular case of high-income inequality, living in these areas may lead poor individuals to develop feelings of failure, both in life and work, associated with depression, when they compare themselves with their richer neighbours. This is in accordance with the income inequality hypothesis.3

Another explanation for the effect of income inequality on depression is through the mediating effect of social capital.32 According to a study by Oishi et al33 which analysed income inequality and happiness over a 37-year period in the USA, individuals were less happy in years with more societal income inequality than in years with less societal income inequality, an association that was found to be mediated by perceived fairness and trust. Walker et al34 proposed that feelings of withdraw and shame experienced by individuals of lower social position are a powerful mediator between relative deprivation and health outcomes such as depression. Higher income inequality has also been associated with a self-enhancing bias,35 where individuals will be more likely to view themselves as superiors to others. Individuals with low social rank living in these unequal societies could develop negative cognitions (such as a constant feeling of defeat) that mediate the association between relative income and mental health.36

According to a meta-analysis of income inequality studies, most of the articles that analysed countries and states have found a consistent association between income inequality and health, while studies that analysed smaller units of aggregation (eg, counties or neighbourhoods) have been less consistent.15 One explanation for this discrepancy may be due to the fact that social comparisons within smaller geographic scales may fail to generate as pronounced a sense of injustice, frustration, shame/envy as the sense of inequality generated by social comparisons across broader geographic scales.37 The present study contributes to the mental health literature by analysing a highly unequal metropolitan area.8 Individually, the 69 areas analysed by this study had lower income inequality than the Metropolitan Area, due to clustering of residents with similar income,38 but there was still significant area-level variance (the Gini coefficients ranged from 0.10 to 0.53). Our results indicate that the social comparisons that lead to psychological distress may also be relevant for smaller areas.

Another important issue in the income inequality literature is determining which individuals (rich or poor) are more affected by income inequality.39 ,40 Previous studies have shown that the association between individual income and health outcomes may be more important for high-income neighbourhoods.41 ,42 For example, based on a pure social comparison effect, one might hypothesise that the affluent would do better (ie, have higher health status) by living in a community where everyone else is poor.43 On the other hand, the pollution effects of income inequality (eg, spillover violence) predict that they will fare worse. Our present study was not able to examine these cross-level interactions, mainly because of the overall low correlation found between individual income and mental disorders. Further studies are warranted to better understand cross-level interaction effects.

We note some important limitations of our study. First, it was based on a cross-sectional survey, so the temporal association between inequality and mental disorders was not established. We tried to avoid the bias of assuming that context will have an immediate effect by including only individuals living in the same address for 5 years or more, but to determine a temporal association between mental disorders and inequality will require a longitudinal approach. Second, the survey used for collecting the data (WMH-CIDI), although following the same pattern for every country, may present variations due to cultural differences in reluctance to admit emotional or substance-abuse problems,44 so international comparisons should be handled with care. Third, the methodology did not perform clinical diagnosis of the individuals, but assessed mental health disorders by self-reported symptoms according to the WMH-CIDI. Fourth, the causal factor that mediates the association between income inequality and mental disorders is still unclear. Previous studies have suggested that it could mediated by the decrease of social capital in unequal areas,32 but this remains unconfirmed. Fifth, the sampling strategy was based on area probability, which means that individual characteristics were not necessarily representative of the population of the SPMA. Seventh, despite having a relatively acceptable response rate (81.3%), the presence of response bias cannot be discarded.


Although not definite, our results suggest that mental disorders, especially depression, are associated with area-level income inequality. We propose, in accordance with the ‘relative income hypothesis’, that this may be an important pathway for the association between income inequality and morbidity and mortality. Recent results from the Global Burden of Disease Study have shown that mental disorders are a fast growing cause of global disability.45 From 1990 to 2010, major depressive disorder increased from 15th to 11th in the rank of global disability-adjusted life years. Economic studies have also shown that income inequality is on the rise for most countries, especially for the most developed.46 The association found in the present study between depressive disorder symptoms and income inequality may indicate a pathway by which countries could start to address this fast-growing burden.

What is already known on this subject

  • Income inequality has been associated with a number of health outcomes, but the pathway by which it affects health is still unclear. The original relative income theory proposed that the mediation was through psychological effects and mental disorders.

What this study adds

  • Higher area-level income inequality was statistically associated with depression, but not with anxiety. The fact that depression is considered a ‘backward-looking’ emotion is in accordance with the original theory, which proposes that high-income inequality will lead to psychological distress due to feelings of work/life failure and negative social comparisons.



  • Contributors ADPCF, IK and LHSGA designed the analysis. ADPCF drafted the article. IK, YPW, MCV and LHSGA revised it critically and interpreted the data. ADPCF, IK, YPW, MCV and LHSGA approved the final version to be published.

  • Funding The present analysis was funded by FAPESP (grant: 12/09717-2). The São Paulo Megacity Mental Health Survey was funded by the State of Sao Paulo Research Foundation, Brazil (FAPESP Grant 03/00204-3, URL: Instrument development was supported by the Foundation for Science and Technology of Vitoria, Espirito Santo, Brazil (Fundo de Apoio à Ciência e Tecnologia do Município de Vitória—FACITEC 002/2003). The subproject on violence and trauma was supported by the Secretaria de Segurança Pública of the State of Sao Paulo, Brazil. The Sao Paulo Megacity Mental Health Survey is carried out in conjunction with the WHO World Mental Health (WMH) Survey Initiative. The authors thank the WMH staff for assistance with instrumentation, fieldwork and data analysis.

  • Competing interests None.

  • Ethics approval Research and Ethics Committee of the University of Sao Paulo Medical School.

  • Provenance and peer review Not commissioned; externally peer reviewed.

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.