Using multilevel models applied to the entire sample including each neighborhood variable separately (500-m radius buffers), living in neighborhoods with lower income, lower real estate prices, and particularly lower education level were associated in a dose-response pattern with higher BMIs and waist circumferences, after adjustment (Table 3, full models in eAppendix 2, http://links.lww.com/EDE/A494). We selected neighborhood education for the next steps of the analyses, based on the following 3 reasons: (i) neighborhood education was more strongly associated with the outcomes than neighborhood income or real estate prices; (ii) neighborhood education led to a better fit to the data than the other neighborhood variables; and (iii) only neighborhood education remained negatively associated with the outcomes when neighborhood variables were introduced 2-by-2 into the models.
In models estimated with an interaction term between sex and neighborhood education, as well as models stratified by sex, the associations between neighborhood education and BMI and waist circumference were stronger among women.
Regarding the buffer spatial scales, sensitivity analyses indicated that the point estimates for the relationships between neighborhood education and BMI and waist circumference slightly increased with the radius of the area from 100 to 500 m, and then decreased from 500 to 10,000 m, particularly among women (Fig. 2). The lowest Akaike information criterion was observed when using 500-m radius buffers (with more important differences among women), supporting the selection of this neighborhood spatial scale for the further analyses.
Logistic models estimated for the odds of living in low educated neighborhoods (to construct propensity score) suggest, for both men and women, that having a low individual education and parents with a low education, being born in a low development country, having a low income, and having a low social class occupation were associated with increased odds of living in low-educated neighborhoods. Furthermore, among men, reporting no vacations and no cultural entertainments, as well as giving a high priority to health, were also associated with increased odds of living in low-educated neighborhoods. In the opposite direction, perceived precariousness was associated with decreased odds of living in low-educated neighborhoods (full models in eAppendix 3, http://links.lww.com/EDE/A494).
In Figure 3, the probability (propensity score, according to individual characteristics) of living in a neighborhood with a lower education level is plotted for the RECORD participants who lived in high and in lower educated neighborhoods using the BMI sample, similar results were obtained for waist circumference. As expected, based on their individual characteristics, residents from neighborhoods with a lower education had higher probabilities of living in lower educated neighborhoods, and residents from high educated neighborhoods had lower probabilities of living in neighborhoods with lower education levels. Comparing the top, middle, and bottom parts of Figure 3 (in which residents from the fourth quartile of neighborhood education were successively represented with residents from the first, second, and third quartiles of education), we observe that the overlap between the curves increases when comparing neighborhood education categories that are closer to each other (see Fig. 3 legend).
The propensity score matched sample for the first and fourth neighborhood education quartiles was, for example, 57% smaller (n = 1026) than the original sample (n = 2368) in the analysis for BMI among men, and 59% smaller (n = 516) than the original sample (n = 1246) among women (Table 4). The decrease in sample size was less important when we compared closer neighborhood education categories, ie, the second and fourth quartiles (sample size was reduced by around 33% for BMI and waist circumference), and the third and fourth quartiles (around 17% smaller sample sizes were observed).
Estimating models in the propensity score matched samples, we obtained point estimates that were largely comparable to those from models based on the entire sample: residents from low-educated neighborhoods had an increased BMI and waist circumference (Table 5, full models in eAppendix 4, http://links.lww.com/EDE/A494). As shown in the Table, a notable difference between the 2 approaches was that the 95% confidence intervals (CIs) were wider when based on the propensity-score matched samples, as a result of their smaller size (descriptive information on matched and not-matched samples in eAppendix 5, http://links.lww.com/EDE/A494).
We observed that living in low socioeconomic status neighborhoods was associated with an increased BMI and waist circumference even after adjustment for individual and maternal characteristics. Although a considerable number of studies investigated relationships between area socioeconomic characteristics and BMI and waist circumference,6 none of these studies implemented any of the following methodological improvements: (i) the measurement of neighborhood socioeconomic status in person-centered areas based on building-level data and the investigation of the optimal spatial scale of measurement of neighborhood variables; and (ii) the comparison between multilevel analyses performed with the entire sample and with restricted propensity score matched samples to assess the exchangeability of exposed and unexposed participants. These 2 aspects are largely complementary because they are related to the measurement of neighborhood socioeconomic status and the modeling of its associations with weight.
Strengths and Limitations
Regarding study strengths, very few other studies9,11 have elaborated neighborhood socioeconomic variables in person-centered areas defined on various spatial scales using administrative sources geocoded at the building level, and none of them in relation to weight. This approach in our study probably contributed to reducing exposure misclassification biases. Regarding health data, as noted in our recent review,6 few neighborhood studies considered indicators of central adiposity as we did.
Regarding study limitations, first, even if our sample is not strictly representative of the Paris Metropolitan area,20 we selected a priori a panel of municipalities from the region to ensure the presence in the sample of people from all socioeconomic backgrounds. Second, the present study considered only participants' current residential neighborhood, disregarding socioeconomic status of previous residential neighborhoods. Third, regarding the analyses, we did not assess whether increases in BMI and waist circumference with decreasing neighborhood education were of different magnitudes at different quantiles of these continuous anthropometric variables. Using quantile regression, eAppendix 6 (http://links.lww.com/EDE/A494) shows that these associations were larger at higher levels of BMI/waist circumference.
New Methodological Insight
The influence of residential neighborhood delimitations on the associations between neighborhood characteristics and health is an important issue that is often neglected. Most researchers do not conduct sensitivity analyses for the spatial scale of measurement of neighborhood characteristics, and when a sensitivity analysis is performed, it is generally not reported. For example, Auchincloss and colleagues35 used kernel-density techniques to smooth census-block group (rather than building level) data to determine variables within circular areas, but they did not perform any sensitivity analysis of the spatial scale of measurement. Mason and colleagues36 conducted but did not report a sensitivity analysis on the spatial scale of neighborhood factors, but their neighborhood measures were based on census tract data rather than on building-level census data as reported in this article.
The choice of the buffer scale may be based on 2 criteria: (1) plausible social and biologic hypotheses of environmental health influences (if there are a priori hypotheses on the spatial scale to retain) and (2) explicit exploratory comparison of the different spatial scales. The choice of a 500-m buffer scale in our analyses was a consequence of these 2 aspects. We assumed a priori that environmental conditions may be associated with health when factors are measured in “walkable” areas from the residence. Areas larger than 500-m radius buffers may not be easily “walkable,” particularly for many of the aged participants of our sample. At the other extreme, smaller geographic areas (eg, with a 100-m radius) may be too small to represent “walkable” areas, therefore leading to weaker associations with weight. This hypothesis of exposure within “walkable” areas may explain why the empirical associations with weight or with abdominal fat were weaker when neighborhood education was measured within areas that were either too small or too large (in areas with a radius below or above 500 m in our case). However, the choice of spatial scale should be made carefully, rerunning sensitivity analyses in each specific study area, for each specific contextual factor and each outcome.
As a second methodological innovation, we used propensity score matching to perform analyses among participants who are exchangeable between neighborhood exposure groups on the basis of a number of individual sociodemographic characteristics that influence the likelihood of exposure to a low socioeconomic status environment. Propensity-score matching was not used in itself as an alternative to adjustment. In the literature, propensity-score matching is typically employed to reduce model dependence, and to estimate associations in a more empirical way than what would be necessary without matching.33 In line with this practice, propensity-score matching was employed as a diagnostic tool to identify potential situations of structural confounding, and as a validation tool to verify that the adjusted neighborhood effects of interest can be estimated without excessive model extrapolations, ie, with a reasonable amount of data in the various cells of the cross-tabulation between explanatory variables (“on-support” inferences).12–13
There was considerable socio-spatial segregation in our sample, because sample size for estimating the effect of the first versus the fourth quartile of neighborhood education on BMI and waist circumference was halved when we applied propensity-score matching. However, matching exposed and unexposed study participants by their propensity score to live in a low-educated neighborhood showed that, in our French sample, there was some overlap in propensity score between them (which might not be the case in other countries, eg, in certain US territories).12 Therefore, to measure this overlap, quantify the socio-spatial segregation level, and compare it across studies, we recommend explicitly presenting the percentage of reduction in sample size (as in Table 4) when undertaking propensity-score matching. This may help to reach the more balanced conclusion that structural confounding is perhaps not a systematic threat to studies of neighborhood and health, as recently claimed.12 Furthermore, it is difficult to provide a definite cutoff for the percentage reduction in sample size beyond which it would seem unreasonable to estimate the adjusted neighborhood effect; the percentage reduction in sample size with matching is somewhat dependent on the caliper size selected for matching.
In terms of point estimates, we found a striking similarity between those obtained using the entire sample and those derived from the restricted propensity-score matched sample, consistent with the intermediate level of social segregation observed in France. Regarding 95% CIs, our propensity-score matching strategy is useful to diagnose situations in which measures of uncertainty are spuriously narrow in the analyses based on the entire sample. If an association is documented in the entire sample, but the 95% CI of the association becomes too large in the reduced propensity-score matched sample to document any association, we would conclude that the association documented in the initial sample had an excessively narrow CI (because it included too many unexchangeable participants who are not usable in the determination of the adjusted neighborhood effect). On the other hand, 95% CIs are perhaps excessively large in the propensity-score matching analyses (conservative analyses), in that an excessive number of participants are excluded from the sample (regression models require data in all cells of the cross-tabulation, not a strict balance in the number of participants between exposure groups at each level of the propensity score).
Overall, the estimations obtained from the propensity- score matched samples are not intended to replace those derived from the entire sample, but rather are meant to provide information allowing assessment of the quality of the associations obtained from the full sample. We do not necessarily view the estimates and related measures of uncertainty from propensity-score matched samples as better than those derived from the entire sample, or as a better trade-off between validity and precision. We recommend that future studies provide the estimates of neighborhood effects obtained for the entire sample (more generalizable), and comparatively those from propensity-score matched samples, to validate that the adjusted neighborhood effect can be estimated without excessive model extrapolations.
Although propensity-score matching is informative when estimating contextual effects, this method does not solve in itself issues of residual confounding related to the selective migration of participants toward specific neighborhoods.16 The critical question is the selection of variables to include in the propensity-score calculation.18,37 Although we attempted to ensure that persons are exchangeable on the basis of several characteristics, a major study limitation is that important individual variables related both to the choice of living in a particular type of neighborhood and to weight were not available in the database (eg, general attitudes toward own body weight and residential strategies), resulting in a misspecification of the propensity score.
In summary, studies that measure and model neighborhood socioeconomic effects should include sensitivity analyses of the buffer-area size to identify the spatial scale on which the environment–health associations operate. Propensity score matching can be implemented to verify that modeling results are not based on excessive extrapolations.
We particularly thank Alfred Spira, head of the French Institute for Public Health Research, for his advice and support. We are grateful to INPES (and Pierre Arwidson) for its continued support since the beginning of the study. We also thank Danièle Mischlich from the Ile-de-France Regional Health Agency and to Nathalie Catajar and Muriel Hirt from the Ile-de-France Youth and Sports Regional Direction for their support in our project. We are grateful to Insee, the French National Institute of Statistics and Economic Studies, which provided support for the geocoding of the RECORD participants and allowed us to access to relevant geographical data (with special thanks to Pascale Breuil and Jean-Luc Lipatz). We thank Geoconcept for allowing us to access to the Universal Geocoder software and Alain Weill from the Caisse Nationale d'Assurance Maladie des Travailleurs Salariés (CNAM-TS) for his support in merging the healthcare consumption data from the SNIIR-AM to the RECORD database. Regarding the data used in the present analysis, we are grateful to Paris-Notaires and to the National Geographic Institute. We also thank CNAM-TS and the Caisse Primaire d'Assurance Maladie de Paris (CPAM-P, France) for helping make this study possible.
1.Chaix B, Chauvin P. Tobacco and alcohol consumption, sedentary lifestyle and overweightness in France: a multilevel analysis of individual and area-level determinants. Eur J Epidemiol
2.Coogan PF, Cozier YC, Krishnan S, et al. Neighborhood socioeconomic status in relation to 10-year weight gain in the Black Women's Health Study. Obesity (Silver Spring)
3.Grow HM, Cook AJ, Arterburn DE, et al. Child obesity associated with social disadvantage of children's neighborhoods. Soc Sci Med
4.Oliver LN, Hayes MV. Effects of neighbourhood income on reported body mass index: an eight year longitudinal study of Canadian children. BMC Public Health
5.Stafford M, Brunner EJ, Head J, et al. Deprivation and the development of obesity a multilevel, longitudinal study in England. Am J Prev Med
6.Leal C, Chaix B. The influence of geographic life environments on cardiometabolic risk factors: a systematic review, a methodological assessment and a research agenda. Obes Rev
8.Chaix B, Merlo J, Evans D, et al. Neighbourhoods in eco-epidemiologic research: delimiting personal exposure areas. A response to Riva, Gauvin, Apparicio and Brodeur. Soc Sci Med
9.Chaix B, Merlo J, Subramanian SV, et al. Comparison of a spatial perspective with the multilevel analytical approach in neighborhood studies: the case of mental and behavioral disorders due to psychoactive substance use in Malmo, Sweden, 2001. Am J Epidemiol
10.Chaix B, Leyland AH, Sabel CE, et al. Spatial clustering of mental disorders and associated characteristics of the neighbourhood context in Malmo, Sweden, in 2001. J Epidemiol Community Health
11.Propper C, Jones K, Bolster A, et al. Local neighbourhood and mental health: evidence from the UK. Soc Sci Med
12.Messer LC, Oakes JM, Mason S. Effects of socioeconomic and racial residential segregation on preterm birth: a cautionary tale of structural confounding. Am J Epidemiol
13.Oakes JM. The (mis)estimation of neighborhood effects: causal inference for a practicable social epidemiology. Soc Sci Med
14.Oakes JM. Commentary: advancing neighbourhood-effects research–selection, inferential support, and structural confounding. Int J Epidemiol
15.Merlo J, Chaix B. Neighbourhood effects and the real world beyond randomized community trials: a reply to Michael J Oakes. Int J Epidemiol
16.Chaix B. geographic life environments and coronary heart disease: a literature review, theoretical contributions, methodological updates, and a research agenda. Annu Rev Public Health
17.Diez Roux AV, Borrell LN, Haan M, et al. Neighbourhood environments and mortality in an elderly cohort: results from the cardiovascular health study. J Epidemiol Community Health
18.Oakes JM, Johnson PJ. Propensity score matching methods for social epidemiology. In: Oakes JM, Kaufman JS, eds. Methods in Social Epidemiology
. San Francisco, CA: Jossey-Bass; 2006:370–392.
19.Chaix B, Bean K, Leal C, et al. Individual/neighborhood social factors and blood pressure in the RECORD Cohort Study: which risk factors explain the associations? Hypertension
20.Chaix B, Billaudeau N, Thomas F, et al. Neighborhood effects on health: correcting bias from neighborhood effects on participation. Epidemiology
21.Havard S, Reich B, Bean K, et al. Social inequalities in residential exposure to road traffic noise: an environmental justice analysis based on the RECORD Cohort Study. Occup Environ Med
22.Beckman A, Hakansson A, Rastam L, et al. The role country of birth plays in receiving disability pensions in relation to patterns of health care utilisation and socioeconomic differences: a multilevel analysis of Malmo, Sweden. BMC Public Health
23.United Nations Development Programme. Human Development Report 2007/2008
. New York: United Nations Development Programme; 2008. Available at: http://hdr.undp.org/
. Accessed December 1, 2010.
24.Thomas F, Bean K, Pannier B, et al. Cardiovascular mortality in overweight subjects: the key role of associated risk factors. Hypertension
25.WHO. Obesity: preventing and managing the global epidemic. Report of a WHO Consultation. Geneva: World Health Organization; 2000. WHO Technical Report Series 894.
26.Benetos A, Thomas F, Pannier B, et al. All-cause and cardiovascular mortality using the different definitions of metabolic syndrome. Am J Cardiol
27.Berke EM, Koepsell TD, Moudon AV, et al. Association of the built environment with physical activity and obesity in older persons. Am J Public Health
28.Rose D, Hutchinson PL, Bodor JN, et al. Neighborhood food environments and body mass index: the importance of in-store contents. Am J Prev Med
29.Rundle A, Neckerman KM, Freeman L, et al. Neighborhood food environment and walkability predict obesity in New York City. Environ Health Perspect
30.Tilt JH, Unfried TM, Roca B. Using objective and subjective measures of neighborhood greenness and accessible destinations for understanding walking trips and BMI in Seattle, Washington. Am J Health Promot
31.Akaike H. A new look at the statistical model identification. IEEE Trans Automat Contr
32.Brookhart MA, Schneeweiss S, Rothman KJ, et al. Variable selection for propensity score models. Am J Epidemiol
33.Ho DE, Imai K, King G, et al. Matching as nonparametric preprocessing for reducting model dependence in parametric causal inference. Polit Anal
34.Westreich D, Cole SR, Funk MJ, et al. The role of the c-statistic in variable selection for propensity score models. Pharmacoepidemiol Drug Saf
35.Auchincloss AH, Diez Roux AV, Brown DG, et al. Association of insulin resistance with distance to wealthy areas the multi-ethnic study of atherosclerosis. Am J Epidemiol
36.Mason SM, Kaufman JS, Emch ME, et al. Ethnic density and preterm birth in African-, Caribbean-, and US-born non-Hispanic black populations in New York City. Am J Epidemiol
© 2011 Lippincott Williams & Wilkins, Inc.
37.Sturmer T, Schneeweiss S, Rothman KJ, et al. Performance of propensity score calibration–a simulation study. Am J Epidemiol