005 - Causally-Interpretable Random-Effects Meta-Analysis

Differences between the sample of individuals participating in biomedical research and larger populations relevant to health policy questions constitute an important bottleneck in the pipeline from therapeutic discovery to clinical application. Recent work-rooted in the tools of causal inference-has made important advances in tackling this problem in the context of meta-analysis. That is, given a target population of interest and a set of results from possibly several clinical trials, how can we transport the results from those trials to the target population in a way that adjusts for heterogeneity in baseline covariates? Roughly, the current state-of-the art in addressing this question involves pooling individual patient data from the set of clinical trials and subsequently applying one of several techniques from causal inference in order estimate the effect of interest as observed in the pooled trial population had the treatment been applied in the target population.

While such covariate-induced heterogeneity constitutes an important way in which the target population may differ from the set of trial participants in a meta-analysis, further sources of heterogeneity remain. In particular, many of the same concerns implicated in the development of random-effects meta-analysis apply equally in this case. Numerous differences in trial conduct between the set of trials included in the meta-analysis may persist even after accounting for covariate differences. For instance, if clinicians at two trial sites applied distinct versions of a given treatment, we might expect systematic differences in trial outcomes even if the same individual participated in both trials. Under a circumstance where one such site has a much larger sample size than the other, pooling their data and adjusting for covariate differences with the target population may inadvertently overrepresent the treatment version applied in the larger trial. Across the numerous, varied settings of real-world clinical practice, we might more reasonably assume that each such version would be applied with equal probability.

Our work builds on existing developments in causally-interpretable meta-analysis by accounting for this residual between-trial heterogeneity, which persists even after adjusting for differences in baseline covariates. We develop causal estimands and corresponding estimators that explicitly reference residual heterogeneity and, moreover, help to clarify the conditions under which we might still make meaningful, policy-relevant conclusions within a target population. In general, transporting results from a set of clinical trials to a target population is highly relevant to health policy, as policies which incorporate information on the costs and benefits of treatments are often designed to apply to diverse populations that may differ significantly from participants in clinical research. Equally diverse are the clinical settings in which such treatments are ultimately put into practice. Our work aims to extend recent methodological advances addressing the former type of diversity by also accounting for the latter. By doing so, we hope to expand the scope of clinical research to make conclusions that are more actionable, policy-relevant, and precise.

Meta-Analysis

Causal Inference 

Mid-Level

Knowledge

International Conference on Health Policy Statistics 2023