| Summary: | In recent years, there have been increased calls for epidemiology to provide evidence that is relevant to policymakers. To meet these calls, a prominent approach uses the potential outcomes framework of causation and focuses on estimation of intervention effects in future target populations (future intervention effects) using results from epidemiologic studies (realized effects). This approach entails a number of assumptions that merit further investigation in the literature, including most fundamentally whether future intervention effect estimates are considered by policymakers to be the only epidemiologic evidence of direct policy relevance. Additionally, several assumptions are required for even internally valid realized effects to be unbiased estimates of future intervention effects, but the mechanisms by which they may be violated and the potential impact of violations remain under development in the literature. To advance understanding of what it means to use epidemiologic evidence to inform policy, and improve the utility and relevance of such data for policymakers, the overarching goal of this dissertation was to investigate several assumptions related to the methodological problem of future intervention effect estimation. To demonstrate real-world relevance and utility of the work for applied research, a case study focused on estimation of the future effect of depression treatment on antiretroviral adherence.
First, a structured review of antiretroviral treatment guidelines and their methodological references tested the assumption that intervention effect estimates represent the totality of policy-relevant epidemiologic evidence; the review revealed a strong emphasis on estimation of intervention effects in target populations, but countered the assumption that they were the only types of evidence that should be considered “policy-relevant.” Subsequently, two simulation studies examined the impact of violations of particular assumptions needed for realized effects (effects from epidemiologic studies) to be unbiased estimates of future intervention effects. The first study showed that even when using the results of an intervention study (e.g. a randomized controlled trial), non-exchangeability between the study and target populations can develop over time, resulting in large under- or over-estimates of the future intervention effects over long time intervals. The second study examined the implications of using effects of harmful exposures to estimate effects of interventions to remove the exposures (e.g. attributable risks), and showed that such estimates may be substantially biased due to violations of the treatment variation irrelevance assumption, when real interventions differ from hypothetical ones due to unremovable consequences of exposures or unintended consequences of intervention. Overall, this dissertation contributes to the literature by clarifying the larger conceptual approaches to generalizing or transporting evidence to future target populations, and by showing the potential impact of violations of certain assumptions required to interpret results from epidemiologic studies as future intervention effects.
|
|---|
