Measuring the impact of differences in risk factor distributions on cross-population differences in disease...

In cross-population comparisons of disease occurrence (prevalence, incidence), a common public health question is the extent to which variations in the distribution of risk factors for the disease explain observed differences. Limited work has been done on formalizing this problem, which is conceptually tantamount to quantifying the degree of confounding for the 'population effect' induced by different factors. A common approach is to compare 'unadjusted' and 'adjusted' regression-based estimates of that parameter, but the interpretation of the resulting 'contribution' measures may be hindered by other confounding sources and non-collapsibility issues. Interactions also raise interpretational challenges.

We formalized this problem using directed acyclic graphs and the potential outcomes framework, on the basis of which we defined a series of estimands that address specific questions and are identifiable under certain causal assumptions. We subsequently determined possible estimators. A study of regional differences in egg allergy prevalence in 1-year-olds was used for illustration.

The main estimands defined were: the change in the prevalence or incidence difference induced by compositional variations in measured risk factors, all at once and individually, relative to a reference population; and the proportion of the crude difference that remains unexplained by measured factors. Standardization (g-computation), inverse probability weighted (IPW) and doubly robust IPW estimators of these estimands were considered.

This work provides a causal theoretical basis for studying disease occurrence differences between populations. The proposed measures can be used to answer the questions that arise in this context under a set of clearly stated assumptions.