Commentary: Estimating and understanding area health effects

deprived area is associated with poor quality of life in a large population-based sample of older adults living in the UK.1 Their paper adds to a large body of work reporting associations between area socio-economic characteristics or area deprivation and a variety of health outcomes.2 The focus on the elderly population is especially interesting because, as Breeze et al. note, there are reasons to believe that area characteristics may be especially relevant to the health and well-being of elderly people who are likely to spend more time in their local areas and rely on their local areas for services and social interactions. Like other researchers, Breeze et al. analyse data from an observational study to estimate ‘area effects’ after controlling for individual-level social class. The need to control for differences in the socio-economic position of people living in different areas has been a key challenge in the field, especially because the forces shaping residential location generate associations between individual-level social class and area deprivation. The most common approach in the literature is to use regression methods to adjust for individual-level characteristics in the estimation of area effects. An important assumption in the use of regression methods is that the overlap in the distribution of measured individual-level confounders across categories of area deprivation is sufficient for the regression adjustment to yield valid estimates of the ‘independent’ effect of area. In the absence of sufficient overlap, adjusted estimates necessarily imply extrapolations beyond the information available in the data, the validity of which cannot be tested.3 There is no unique answer as to how much overlap is sufficient to allow meaningful estimates. A certain amount of extrapolation is inherent in all scientific inquiry. In fact, the process of adjustment always involves varying amounts of extrapolation from the data at hand to what would have been observed if certain features of the data were different (e.g. if the age distributions of two groups being compared were not as different as they are). Nevertheless, it is important to be explicit about how far we are straying from the data in our extrapolated conclusions, so that readers can judge for themselves whether the assumptions implicit in the extrapolation are likely to be valid. Breeze et al. address the issue of overlap between area deprivation and social class in their sample by reporting the distribution of individual-level confounders by categories of area characteristics in Table 2 of their paper. They show that although, as expected, social class and area deprivation are associated, there is still substantial variability in deprivation of the area of residence within social class categories. A similar pattern has been reported in other contexts,4 suggesting that lack of overlap is unlikely to be as important a problem in estimates of area effects from observational studies as is sometimes implied. In addition, the analytical approach followed by Breeze et al. is slightly different from the usual approach in the literature in that they estimate the combined (as opposed to the ‘independent’) effects of social class and area deprivation. This is accomplished very simply by estimating prevalence ratios for cross-classified cells of social class and area deprivation. This approach has the advantage that estimates are based on the people in each cross-classified cell, and is closer to reality than adjusted estimates which smooth (and extrapolate) over cells and artificially separate out effects which are inextricably linked in reality. A disadvantage is of course that the cells can get very small, as they sometimes do in the analyses reported by Breeze et al., although they are somewhat protected from this problem by relatively large overall sample size. Another disadvantage is that presenting results for each cell can become very cumbersome. Breeze et al. avoid this by reporting results only for the most extreme social class and area deprivation categories, but this means that we do not see all the data. An important assumption in the use of adjustment (including stratification) strategies to estimate causal area effects from observational data is that the adjusted comparison approximates the counterfactual contrast of interest. In the results reported by Breeze et al., the prevalence ratio of 1.49 reported in Table 3 for home management in people of social class I/II living in the most deprived Carstairs quartile compared with people of similar social class living in the least deprived Carstairs quartile can be interpreted as meaning that if people living in the least deprived area quartile lived instead in the most deprived quartile, their probability of having a mobility impairment would be 49% higher. The assumption is that the people living in the least deprived areas are a good proxy for what the people living in the most deprived areas would be like if they did not live in the most deprived areas. Usually in epidemiological jargon, this implies no residual confounding by measured variables and no unmeasured confounders. The inability to undeniably confirm the validity of this assumption is the crucial limitation of observational analyses, like those reported by Breeze et al. However, there are ways to test the sensitivity of results due to this assumption, for example, by examining how much results change when covariates are modelled in different ways or by estimating how strongly an omitted confounder would have to be associated with area deprivation and with the outcome to create the associations observed. These types of sensitivity analyses have only recently Published by Oxford University Press on behalf of the International Epidemiological Association International Journal of Epidemiology © The Author 2005; all rights reserved. doi:10.1093/ije/dyi020