What's New in Causal Mediation Analysis?

Mediation analysis, which started in the mid-1980s, is used extensively by applied researchers. Indirect and direct effects are the parts of a treatment effect that are mediated by a covariate (indirect effect) and the part that is not (direct effect). Subsequent work on natural and pure indirect and direct effects provides a formal causal interpretation, based on cross-worlds counterfactuals: outcomes under treatment with the mediator set to its value without treatment. Organic indirect and direct effects avoid cross-worlds counterfactuals, using so-called organic interventions on the mediator while keeping the initial treatment fixed. We argue that pure and organic indirect effects are very relevant for drug development: 1] They are often the effect of a treatment through its intended pathway, and 2] they can be estimated without on-treatment outcome data. We illustrate our approach by estimating the pure/organic indirect effect of αDEspR in COVID-19 patients in the ICU. αDEspR targets elimination of circulating DEspR+ neutrophils extruding neutrophil-extracellular traps (NET+Ns) to attenuate or prevent multi-organ failure in critical COVID-19 in the ICU. Thus, we estimate the pure/organic indirect effect of αDEspR mediated by DEspR+[NET+Ns]. Using the sequential organ failure assessment (SOFA)-score as a measure of multi-organ failure, we estimated the pure/organic indirect effect of αDEspR using outcome data from patients with COVID-19 in the ICU not treated with αDEspR. Our analysis illustrates the pre-clinical promise of αDEspR, to be used as an argument to fund an early-stage randomized clinical trial to collect on-treatment outcomes and estimate αDEspR's overall effect – thus giving insight into clinical trial design. Causal mediation analysis can also be used as a way to evaluate DEspR+[NET+Ns] as a clinical trial stratifying and monitoring biomarker in critical COVID-19. This illustrates how causal mediation analysis can be used as a much-needed translational bridge from animal model testing towards clinical trial testing. 

Public health interventions may spill over to others in the network of those directly receiving the intervention. In the past, this phenomena has been termed 'contamination' or 'interference' and often considered undesirable. When it occurs, standard analysis methods need further refinement, but from a public health perspective, the phenomena is highly desirable. We present methods for the estimation of spillover effects in cluster randomized trials in the presence of non-adherence. In addition, we present methods for the evaluation of mediation of overall intervention effects and of individual and spillover effects, that are appropriate for cluster randomized trials. Finally, we present methods for evaluation of the total effect of interventions, including within-treated-cluster spillover as well as spillover between treated and untreated clusters, as may occur in practice. Methods are illustrated in the Botswana Community Prevention Program, a cluster-randomized trial of a HIV prevention package conducted in 16 villages between 2013-2020, focusing on the evaluation of the spillover and mediation effects of voluntary medical male circumcision, one of the package's components.

Donna Spiegelman, Melody Owen, Laura Forastiere, Fan Li
Center on Methods for Implementation and Prevention Science
Department of Biostatistics
Yale School of Public Health
 

Causal mediation has traditionally been framed as the effect of an exposure on an outcome through some intermediate variable, where each variable is measured at three sequential time points. However, definitions of mediated effects and their corresponding identification assumptions generally ignore the fact that the mediator of interest is, in many if not most circumstances, a stochastic process indexed by time from baseline to follow-up. I demonstrate that the failure to account for the mediator process has important implications for defining the relevant causal estimand of interest as well as its identification and estimation. Additionally, I discuss versions of direct and indirect effect definitions that account for the entire mediator process and how they relate to corresponding versions that do not.