Abstrak: Although socioeconomic position is conceptualized by social epidemiologists as a multidimensional construct, most research on socioeconomic disparities in health uses a limited set of observable indicators (e.g., educational attainment, household income, or occupational class) and typically analyzes and reports gradients in relation to one measure at a time. Societal changes in economic structures over time, however, can lead to changes in distributions of and associations between socioeconomic indicators, as has occurred with income returns to education in the United States over the last 50 years. Consequently, temporal comparisons of socioeconomic disparities from repeated cross-sectional surveys can be affected, particularly when salient dimensions of socioeconomic position are unobserved. We discuss this phenomenon within the framework of measurement error and identify sources of variation that can make identification of socioeconomic change difficult. Using simulations, we explore the utility of the quantile, slope index of inequality, and relative distribution approaches to minimizing bias in temporal comparisons and find that these methods yield correct inferences about temporal change only under limited conditions. We contrast these approaches with the use of an imputation model when validation data for the unobserved socioeconomic indicator exist. We discuss implications for analyzing changing socioeconomic health disparities over time.

Abstrak: Chen et al. (Am J Epidemiol. 2013;177(9):870-881) develop a simulation study for comparing various measures of socioeconomic health disparities when bias can arise from temporal changes in the bivariate distribution of education and income. In this commentary, I argue that, in relation to health, the "meaning" of education cannot be reduced to its socioeconomic value; improved health literacy, for instance, can result in important health benefits. Further, I suggest that unless there is a substantial prior understanding of the data-generating mechanism, directed acyclic graph models should be avoided because causal relationships cannot be inferred from regression. An alternative is to resort to conditional independence graphs, which use only undirected edges. Finally, although the slope index of inequality can, in some specific cases, be seen to reduce bias in temporal comparisons of socioeconomic health disparities, it was not designed for causal inference. The slope index of inequality simply describes the average change in the proportion in poor health when the population is ordered by socioeconomic status.