Digital version of the poster "Causal inference in health disparities research and the No-Multiple-Versions-of-Treatment assumption" presented at the EAM Conference 2023
Author: Lizbeth Burgos Ochoa, Tilburg University.
Abstract: Health disparities research is often interested in identifying the drivers of observed health differences across population groups. Such research questions are of causal nature as they involve the estimation of the effect of a certain exposure (e.g., socioeconomic status and neighbourhood poverty) on a health outcome of interest. Causal inference in health disparities research is challenging as in most scenarios researchers can only make use of observational data for this purpose. The Neyman-Rubin potential outcomes framework has provided a conceptual framework and supported methodological approaches to estimate causal effects from observational data. To identify a causal effect, the framework relies on the following assumptions: exchangeability, positivity, and no-multiple-versions-of-treatment. While a large part of the literature has focused on the exchangeability assumption (and to a lesser extent to positivity), the last assumption has received little attention and it has often been taken for granted. This work focuses on the implications of the no-multiple-versions-of-treatment assumption for bias in health disparities research from a conceptual and model specification perspective. Various scenarios in which this assumption may be violated are discussed, including those related to the application of particular analytical strategies. Moreover, guidance is provided on the situations under which, even when there is a violation of the assumption, the estimates could still be interpreted as causal.