Ability of ecological deprivation indices to measure social inequalities in a French cohort

The E3N cohort (Epidemiological prospective cohort study among women of the Mutuelle Générale de l’Education Nationale), was initiated in 1990 to study major chronic diseases, among 98,995 women, born between 1925 and 1950, and insured under a health insurance plan covering mostly teachers [33]. Questionnaires are sent, roughly every 2 years, to update information on lifestyle factors and newly diagnosed medical conditions. E3N was approved by the French Commission for Data Protection and Privacy.

The individual-level SEP was evaluated using the women’s educational level in 4 classes (<high school diploma, high school to 2-level university, 3−/4-level and 5-level) collected in 1990.

We calculated 3 DIs (Additional file 1): the Townsend index [5], FEDI [19] and FDep [22] using the 2009 French national census at IRIS level (regrouped statistical information blocks). IRIS is the smallest geographical division in France with 2000 inhabitants in average (towns with more than 5000 inhabitants are divided into several IRIS, while smaller towns form one IRIS each). The homogeneity of each unit is based mainly on habitat type (residential area, public housing, etc.). The Townsend Index is a combination of 4 census-derived variables. In the present study, the proportion of primary residences with more than 1 person per room was used instead of the percentage of overcrowded households (not available in France). FEDI is a combination of 10 weighted census-derived variables associated to average social deprivation in France and identified to best represent individual experience of deprivation [19]. FDep is generated using principal component analysis (PCA) from a set of 4 census-derived variables with both negative and positive socioeconomic dimensions (Additional file 1: Table S1).

We calculated the DIs for 44,709 Metropolitan French IRIS for which the census-derived variables were available (i.e. 89.1%; due to data confidentiality, median income was not available in 5481 IRIS with less than 50 households). We ranked the DIs score (using the population-weighted approach) into five deprivation quintiles from the least (Q1) to the most (Q5) deprived IRIS with approximately 20% of the French population in each ones. The population-weighted approach classify the proportion of the deprived population rather than of the deprived areas [36]. We applied the score and the quintiles of each DI to the women’s residential address in 2005 previously geocoded by a commercial firm, which attributed for each address a level of geocoding accuracy. Geocoding was considered as “precise” if the exact address was found automatically with the highest possible precision (<15 m).

We used the smoking status in two classes (ever smokers vs. never smokers) to evaluate retrospectively smoking initiation [25]. We defined the overweight status using the body mass index (weight(kg)/height(m)²) with a cut-off at 25 kg/m². Both outcomes were reported by participants in the questionnaire sent in 2005.

To take into account the impact of urban-rural settings on the DIs, we classified the addresses using the degree of urbanicity, based on the concept of urban unit, defined at commune-level by INSEE (French National Institute of Statistics and Economic Studies) [22]. An urban unit is a town or a group of towns that includes at least 2000 inhabitants and in which no building is farther than 200 m away from its nearest neighborhood. The degree of urbanicity is defined in 4 classes according to the population size: “Paris-and-suburbs”, “urban” (100,000 to 1,999,999 inhabitants), “quasi-urban” (10,000–99,999), “quasi-rural and rural” (<9999).

Descriptive analyses were performed by t-tests for continuous variables and chi-squared tests for categorical variables. We quantified the degree of agreement between the 3 DIs in classifying the women into the same or a close quintile using weighted Cohen’s Kappa (Kw) statistics (0.81–1.00:almost perfect agreement, 0.61-0.80:substantial, 0.41–0.60:moderate, 0.21–0.40:fair, 0.00–0.20:slight, <0:poor agreement) [37]. We compared the mean of the DIs (as continuous variables) of each educational level. We hypothesized that the better DI, for our specific French elderly women population, would be the one that was associated with well-known socially patterned outcomes.

We evaluated the ability to detect well-established associations (external or convergent validity) as previously performed for SEP [34, 35] and occupational exposures [38]. We studied the associations between each SEP indicators and both smoking and overweight status using logistic regressions models adjusted for age (Software package SAS 9.3). The reference category for the SEP indicators was the highest educational level and the least deprived quintile (Q1). The odd ratio estimates were computed with generalized estimating equations (SAS GENMOD procedure) to control for clustering effects from participants within the same IRIS in a context of sparse clustered data (87% of IRIS contained less than 5 participants, Table 1) [39].

More prevalent smoking initiation [25, 26] and less prevalent overweight [28‐30] are well-established associations described in the literature among French elderly women with higher educational level. In addition at area-level, risk factors associated to lifestyle (such as smoking and overweight) are expected to be in the same direction as individual SEP due to normative standards and behavioral characteristics [32]. Therefore, the present analysis relied on the following assumptions: (i) more prevalent smoking and less prevalent overweight among elderly women living in less deprived area; (ii) no association between Townsend index and both outcomes, as this index was not adapted to the French context (iii) The French DIs were not built in the same way (neither for the same purpose) and might classify differently the E3N women regarding their residential deprivation context. We aimed to evaluate, which one, if any, may be more adapted to capture socioeconomic residential conditions of our specific elderly women population (in average, highly educated compared to the general population of the same age group).

We performed several sensitivity analyses. First, to test the robustness of the DIs [3], we stratified the analyses on (i) age (<65 vs. ≥65 years) to check the hypothesis of stronger associations among older women for smoking; (ii) geocoding accuracy to check the hypothesis of stronger associations among those with precise geocoding. Furthermore, to assess the impact of urban-rural settings on the DIs as previously suggested [8], we adjusted on degree of urbanicity. Furthermore, individual SEP was evaluated at baseline and DIs in 2009 (census data not available at baseline) which may induce misclassification bias. Therefore, we performed a sensitivity analysis including only women who did not move between 1991 and 2005 (n = 39,130; 61%). Finally, we performed analyses using the un-weighted approach (i.e. quintile ranked independently of the population size).

Out of 71,411 women who completed the questionnaire in 2005, we excluded 10.6% from the analyses due to missing data on IRIS (n = 462), DIs (n = 424), educational level (n = 2824), smoking status (n = 1115) and body mass index (n = 2698) (Additional file 1: Fig. S1). The excluded women were slightly older, more smokers during life, more educated and overweight (not shown). The final study population included 63,888 women. The women (Table 2) were in average 64 years old, 47% had ever smoked and 30% were overweight. A majority had attended at least the high school diploma (88%) and lived in urban or quasi-urban areas (77%). The older women (≥ 65 years old) had significantly lower educational level, were less often ever smokers and more often overweight compared to the younger. In average, FDep did not vary according to age (p = 0.30) contrary to FEDI and Townsend (p < 0.0001).

Women who lived in urban areas were significantly older, ever smokers and less overweight compared to women living in rural areas (Table 3). The FDep index tended to decrease with the increasing of urbanicity (especially for Paris and suburbs). On the contrary, FEDI and Townsend increased with the increasing of urbanicity, i.e. in urban areas, women were classified as living in more deprived neighborhoods. As expected, when the whole France DIs quintiles were applied to the E3N population, less than 15% of the women were living in the most deprived areas (Q5; Additional file 1: Fig. S2). We observed significant linear associations between educational level and the scores of the DIs (Table 4). FDep and FEDI decreased when educational level increased with a clearer linear trend for FDep whereas an opposite trend was observed for Townsend.

The agreement was substantial between the two French DIs (κw = 0.61), whatever the degree of urbanicity, except for Paris and suburbs (0.28). The agreement between Townsend and FEDI was substantial (0.74) and fair with FDep (0.21).

As expected in this French elderly women population, educational level was associated with smoking with a significant trend (Fig. 1). Women with lower educational level were less prone to be ever smokers, compared to those higher educated (Low vs. High; OR [95% CI] = 0.43 [0.40; 0.46], (p for trend <0.0001). At area-level, only FDep showed the same pattern (i.e., women living in more deprived areas were less prone to be ever smokers (most deprived (Q5) vs. least deprived quintile (Q1); 0.77 [0.73; 0.82], p for trend <0.0001). We observed opposite significant associations using FEDI (1.13 [1.07; 1.20]) and Townsend (1.51 [1.43; 1.59]). Regarding overweight status (Fig. 2), as expected, women with lower educational level were significantly more often overweight (1.89 [1.77; 2.01]), p for trend <0.0001) compared to women higher educated. We observed a similar pattern with both FDep (1.52 [1.44; 1.62] and FEDI (1.20 [1.13; 1.28]), and an opposite association with Townsend (0.93 [0.88; 0.99]).

We observed stronger associations among older women (≥ 65 years old) for smoking (Additional file 1: Fig. S3). When we restricted the analysis to women with the better geocoding accuracy (50% of the population), we observed stronger associations for both smoking and overweight status (not shown). Analyses adjusted on degree of urbanicity showed similar results with FDep for both smoking and overweight whereas associations became non-significant for smoking and stronger for overweight when using FEDI (Table 5). Similar results were also observed when the analysis was restricted to non-movers between 1991 and 2005 (n = 39,130; 61%) (Additional file 1: Table S2). Finally, weighted quintiles ranking method gave slightly stronger associations with both smoking and overweight status, compared to unweighted ones, especially for FDep (Additional file 1: Fig. S4-S5).

The ability of area-based indicators to predict known socially patterned outcomes have been previously studied mostly in Anglo-Saxon countries [40]. Our study, the first one performed in a very large French epidemiological cohort, showed the reliability of FDep to capture socioeconomic residential conditions among E3N elderly women. Results for FDep were consistent with those of the literature for both smoking and overweight status with a clear gradient between the quintiles [27, 31]. Moreover, we observed a stronger association among older women between FDep and smoking, similarly to what was observed with educational level. It has been suggested that commonly used DIs suited poorly to study inequalities in older people especially because they included variables related to the active population or male-centered (social class) [14]. The use of a DI in our population could therefore be a limitation. However, interestingly, FDep did not varied according to age, contrary to FEDI and Townsend for which older women were classified as more deprived in average. Stronger associations were also observed with FDep when we restricted the analyses to women with precise geocoding linkage, which was expected in the case of non-differential geocoding errors regarding SEP. [41] In French ecological studies, stronger associations were observed with FDep compared to Townsend, studying DIs and all-cause mortality at commune-level [22], consistently to our results. In addition, FDep was found less sensitive to urban-rural differences than Townsend, studying associations with colorectal cancer screening attendance [8] consistently to our results.

As expected, Townsend was not adapted to evaluate residential deprivation in a French context [20, 22]. Although FDep and FEDI showed a substantial agreement in classifying the women across the range of deprivation, we observed discrepancies in predicting smoking and overweight social patterns in E3N. This discrepancy could be explained by their different mode of construction. FDep was constructed to maximize the heterogeneity of the components using a PCA [22]. FEDI was composed of weighted variables identified to best represent individual experience of deprivation and based on average social deprivation [19] and thus might be less adapted to capture the variety of socio-spatial situations that composed the French territory. Moreover, DIs we applied here have not been created in the same context. FEDI was constructed to proxy individual SEP whereas FDep was setup in the context of ecological approaches.

The different items included in the DIs may also explain the disagreement. For example, FEDI and Townsend included the "proportion of households not owner occupied", "primary residence with more than 1 person per room" and “without a car”. These items are known to vary according to the degree of urbanicity and specifically in rural vs. urban areas [13, 42]. For example, in rural areas, not possessing cars could be an obstacle for mobility and though be a proxy of deprivation, whereas in urban areas, especially in large cities, it is common to have no car as public transport is particularly developed in France. Likewise, overcrowding and home-ownership are not comparable between urban and rural settings and could be a marker of deprivation in rural areas but not always in urban ones. We observed that the FDep index tended to decrease (i.e. less deprivation) with the increasing of urbanicity, especially for Paris and suburbs. On the contrary, FEDI and Townsend increase (i.e. more deprivation) with the increasing of urbanicity. This opposite trend which appeared clearly on the maps of the distribution of the IRIS (Additional file 1: Fig. S6), could ensue from these items. For example, the map with the Townsend index showed light shades because the majority of the French territory is rural. On the contrary, the map with the FDep index is darker because rural areas are classified as more deprived than urban ones. Furthermore, individual characteristics of the participants varied according to the degree of urbanicity of their place of residence, with higher prevalence of ever smokers and less prevalence of overweight in urban areas that might also explain the unexpected associations observed for Townsend and FEDI with smoking.

Our study presented several strengths. We used a very large population sample homogeneously distributed across the French territory. At area-level, composite indicators were more effective to take into account the multidimensionality of the SEP than a single one [4]. Our results confirmed that FDep, initially developed at commune-level [22], was also able to capture the inter- and intra-urban socio-spatial divisions existing in France at IRIS level. In addition, we used the population-weighted approach to construct the DIs quintiles, which allowed a better classification of population and gave stronger associations between FDep and the outcomes. To the best of authors’ knowledge the present epidemiological study is the first one to compare weighted and un-weighted methods. We tested two different outcomes with established social pattern, smoking and overweight, to assess the robustness of the DIs [3]. We used GEE models to control for clustering effects from participants within the same IRIS in a context of sparsely clustering data, as recommended [39]. We used the finest spatial unit with socio-demographic data available in France to minimize misclassification and potential ecological bias as recommended [43]. Associations observed between FDep and known SEP related outcomes fit within 2 a priori criteria described as external validity and robustness by Krieger et al. [3].

The study nevertheless had some limitations. Models including both individual and area-level variables were not performed in the present study because the hypotheses were based on studies using either SEP indicators at individual- or at area-level separately. Therefore, we were not able to distinguish compositional from contextual effects. Women’s residential history were not taken into account. However, less than 30% of movers were identified between 1991 and 2005 in a sub-E3N population [44]. E3N women were 45 years at baseline, thus we hypothesized that their social trajectory was already settled and did not change much during this period. In addition, census data were not available to calculate the DIs at baseline. However, French studies have shown that spatial distribution of deprivation did not change substantially since 1991 [45, 46]. Nonetheless, we performed a sensitivity analysis including only women who did not move and the conclusion was similar.

Historically, area-based SEP has been used as a surrogate of individual-SEP in medical records [47], but this strategy have been questioned particularly in Anglo-Saxon countries [48]. Some methodological studies have compared the agreement between individual and area-based SEP and their ability in predicting health outcomes [35, 49] with conflicting results. Poor agreement has been reported between self-reported individual income and area-based income [49]. While, in others studies, area-based SEP was considered as a good proxy of individual-level SEP [35] allowing the prediction of socially patterned outcomes. Finally, it has been underlined that area-based SEP indicators fairly classify socially homogenous areas (most and least deprived neighborhoods) but failed sometimes to classify the in-between situations that are more heterogeneous [50]. In the relatively highly educated E3N population, we observed a clear gradient across the quintiles of FDep for both outcomes, whatever the strategy of analysis. The E3N population is not representative of the French elderly women. They have in average higher educational level than French elderly women and probably healthier conditions. However, even in this specific population, we found that social disparities in smoking and overweight do not affect only extreme social situations but rather the socioeconomic gradient [43]. Our objective was to determine which area-based SEP could meaningfully be used to further study social disparities in health in an elderly women population. It has been underlined that DIs might not be suitable in specific populations, such as elderly [15] or women [15, 17]. FDep appeared to be a good indicator to capture inter- and intra-urban socio-spatial divisions existing in France and seemed reliable to capture socioeconomic residential conditions of the E3N elderly women population, mostly teachers living in urban areas.