The influence of neighbourhood formality status and socio-economic position on self-rated health among adult men and women: a multilevel, cross sectional, population study from Aleppo, Syria

Residential neighbourhoods in Aleppo were defined as either ‘formal’ or ‘informal’. There were 87 neighbourhoods in the formal zone and 29 neighbourhoods in the informal zone, as described in the municipal registry. The population of Aleppo was split between the formal and informal zones in a ratio of approximately 2:1. The number of households within a neighbourhood varied from 97 to 18,100.

Data were from the Aleppo Household Survey (AHS), which was conducted by the Syrian Centre for Tobacco Studies (SCTS) in 2004. The survey combined a health interview with a health examination, providing a ‘baseline map’ of the main health issues and exposures affecting communities in Aleppo. Details of the survey, population and sampling have been described elsewhere [8, 15] but relevant details on the structure of the dataset are given here. Ethical approval for the survey study was obtained from Institutional Review Boards of the university of Memphis in the USA and the SCTS [15].

The survey design employed multistage random sampling. The sampling unit was the household with one person aged 18–65 randomly selected from each household to participate. Households were clustered in neighbourhoods within two zones: formal and informal. Sampling was within zone, with a target of 1000 households from each. Firstly, within each zone, neighbourhoods were randomly selected with probability proportional to the population size of the neighbourhood. A fixed proportion of households from each neighbourhood was then selected by first choosing a random starting point followed by the systematic sampling of every fifth household until the neighbourhood target sample size was achieved (0.6% of households per neighbourhood in the formal zone and 0.9% in the Informal zone). The resultant sample of households would be spatially closer together than one would expect by chance if simple random sampling within administrative neighbourhood boundary had been employed.

Table 1 includes a description of the outcome variable and compositional (individual-level) and contextual (neighbourhood-level) variables included in the analysis. In the absence of neighbourhood data from the Census in Aleppo, we aggregated the individual level socioeconomic indicators from the survey data to the neighbourhood level [16].

Multilevel models were fitted in MLwiN software (version 2.02) [17] to account for the hierarchical structure of the data. Models were fitted with and without design based weights. Design based weights were calculated to adjust for the unequal probability of selection of individuals in the sample. Neighbourhood and individual level weights were calculated separately and incorporated into the multilevel analysis at the appropriate level [18]. Design weights were calculated for the whole sample of 2038 individuals, based on Census data for the combined male and female population of Aleppo as no information was available by gender. For most variables, inferential conclusions were very similar for the weighted and un-weighted analyses. However, there were some discrepancies for categorical variables when cell frequencies were small; in such cases un-weighted findings are reported [19]. We report un-weighted multilevel binary logistic regression models with excellent SRH as the reference category. Multilevel models were fitted using Markov Chain Monte Carlo (MCMC) estimation [20] with orthogonal parameterisation and parameter expansion [21]. We compared successive models using the deviance information criterion [22]. Inferences were drawn from chains of length 50,000 after a burn-in of 2,000 with the means and 95% CIs from the posterior distribution given in the results.

The variance between neighbourhoods was expressed as both an intra-class correlation coefficient (ICC) and a median odds ratio (MOR). The ICC was calculated assuming a linear threshold model [23] and can be interpreted as the percentage of the unexplained variation in the probability of poor/normal SRH, which can be attributed to differences between neighbourhoods. The MOR [24] is an alternative way of expressing neighbourhood level variation that translates the variance to the odds ratio (OR) scale, making the MOR comparable with the ORs of individual or neighbourhood level variables. An MOR of one indicates no residual variation between neighbourhoods in the probability of poor/normal SRH.

The analysis proceeded in four stages, separately for males and females. Firstly, we fitted an empty variance components model examining the amount of variation in the probability of poor/normal SRH that could be attributed to the individual and neighbourhood levels of analysis.

The second stage expanded the empty MLM by including age, BMI, smoking status, level of physical activity, marital status, social support variables and perception of neighbourhood variables as fixed effects. Formality was also added at this stage as a fixed effect. In addition to main effects, all plausible interactions between pairs of individual level variables and between individual level variables and formality were explored. The significance of fixed effects terms in the model were tested by considering whether the odds ratio differed significantly from one. To assess whether there was an association between individual level SES and SRH after adjusting for other variables, each individual level SES variable was then added in turn and simultaneously. In the third stage, we allowed any statistically significant fixed effect for individual level SES to vary randomly across neighbourhoods, enabling us to determine whether individual SES variables predicted SRH in a manner that differed across neighbourhoods. Finally, we entered the aggregated neighbourhood level SES variables one at a time and assessed their inclusion on the regression coefficient for individual level SES and on the between neighbourhood variation.

Analyses were also performed with and without very small and very large neighbourhoods to assess the influence of neighbourhood size on the fixed effect estimates. Neighbourhood size was also included as an explanatory variable in an attempt to determine if there was anything ‘different’ about small neighbourhoods. MCMC diagnostics were used to assess parameter estimation and multilevel residuals were examined to assess model fit.

Of 2370 people invited, a total of 2,038 (86%) people took part in the survey. Survey respondents were clustered in 45 neighbourhoods; 1017 respondents from 27 formal neighbourhoods and 1021 respondents from 18 informal neighbourhoods. The sample included 921 (45.2%) men and 1117 (54.8%) women. For females, the median achieved number of respondents per neighbourhood was 22, range 2 to 95 (inter-quartile range 9 to 35); for males, the median was 17, range 1 to 72 (inter-quartile range 10 to 27).

Table 2 shows a simple summary of the sample characteristics by formality and gender.

Overall, women were more likely to describe their health as poor (11.5% of those reporting poor SRH were women compared to 6.2% of men). They were also more likely to report that they were illiterate (26.6% of women compared with 13.9% of men) and more likely to be unemployed (86.3% of women reported they were unemployed compared with 10.3% of men).

Tables 3 and 4 present the results of fitting the empty and final variance component models for females and males separately. The final model results are presented as odds ratios together with 95% CIs (also adjusted for age, BMI, smoking status and level of physical exercise). For the empty models the ICCs suggest there is some evidence of variation in the probability of poor/normal SRH between neighbourhoods. For females 6% of the variation in the probability of both poor and normal SRH (versus excellent) is attributable to neighbourhoods and for males 7% of the poor SRH and 4% of the normal.

For females, of the five separate individual level SES variables, only ‘number of household items owned’ was significantly associated with SRH after adjustment for all other variables in the model. For an additional item in the home, females were less likely to have poor or normal SRH versus excellent. The individual level ORs for item ownership were attenuated only very slightly when neighbourhood level average item ownership was included in the models, thus suggesting a contextual effect of neighbourhood average item ownership independent of individual level ownership. Furthermore, after adjusting for neighbourhood SES a woman residing in the informal zone was less likely to report poor or normal SRH than excellent suggesting a protective effect of informality on health (all else being equal).

For males, no individual level or neighbourhood level SES variables were significantly associated with SRH after having adjusted for all other variables in the models. However, men who reported that they suffered at home from outdoor noise ‘a lot’ compared with ‘somewhat’ or ‘not at all’ were more likely to have poor or normal SRH versus excellent (see Table 4).

Firstly, this is a secondary analysis of cross-sectional data and hence it was not possible to consider changes over time in the explanatory or outcome variables. It therefore potentially suffers from the problem of self-selection, whereby people are selected into residential areas based on unmeasured individual characteristics that are relevant to health [30, 31]. Findings from previous studies in the West show that people’s ratings of health depends on current health status as well as on changes over time in relation to socioeconomic status, chronic disease, functional disability and mental health [32, 33]. Although the AHS included information on functional disability and chronic disease, these variables were not included in our analysis since these are measures which are highly correlated with self-rated health. Additionally, cross-sectional surveys do not include longitudinal and historical data on neighbourhoods which interact with, and are heavily influenced by, a multitude of macro-level factors such as social, cultural, economic and policy contexts of states [34, 35]. They also change over time as a result of societal processes such as economic cycles or demographic shifts and migration [36]. This is a particular issue for informal areas, which may change on a day to day basis. The study also lacked data on neighbourhood SES which are not available in Syria. In the absence of such information the importance of contextual SES has been assessed by aggregating individual level indicators from the survey data to the neighbourhood level. For some neighbourhoods, aggregate SES measures were therefore based on the responses of only a few participants (as noted in the Results section, the lower quartile for achieved neighbourhood size was 9 for females and 10 for males). Such averaging of data from very few respondents might have introduced measurement error into the estimation of our neighbourhood level estimates. To explore the impact of these ‘small’ neighbourhoods on the estimation of the fixed effects at the neighbourhood level, the analyses were repeated omitting neighbourhoods with <10 respondents. The resultant final model ORs for neighbourhood level average item ownership were the same to one decimal place. Other authors have encountered similar numerical issues [37]. Those authors also observed a stronger relationship between an individual outcome variable and an aggregated contextual level variable than between the outcome variable and the corresponding individual level variable (as we do here for self-reported health and neighbourhood average item ownership among females). We conclude, like them, that this may represent a genuine contextual effect. However, further research via simulation studies, is needed to investigate the utility of aggregated neighbourhood level variables as measures of contextual effects. A further limitation of this study is the low statistical power to estimate between neighbourhood variation, mainly as a result of the small number of individuals available within a high proportion of the neighbourhoods [31, 38, 39]. This may explain why as a proportion of the total variation, variation between neighbourhoods in poor/normal self-rated health for females and males was estimated from the model to be small and non-significant.

Despite these limitations, this study is unique in that it considers neighbourhood informality status which is a relevant neighbourhood characteristic for the health of urban residents in developing countries. Therefore, by using an integral measure of neighbourhood (defined as a feature of an area only measurable at an ecological level of neighbourhood) this study has avoided relying solely on a derived neighbourhood measure such as SES which is based on the aggregate characteristics of individuals [36].

Additional strengths of this study include the way in which neighbourhoods were operationalised. Whist formal and informal areas correspond to administrative units that are defined by the local government, the way the neighbourhoods were measured corresponds to smaller, more homogenous clusters or units that are more likely to match people’s perceptions of their local neighbourhoods and hence could be more related to the way they rate their health [2, 35]. Another strength of this study was the inclusion in the analysis of correlates of SRH, such as social support, physical activity and smoking status, that have been found to vary by gender in Aleppo [3].