Mediation Analysis with Multiple Mediators

Abstract

Recent advances in the causal inference literature on mediation have extended traditional approaches to direct and indirect effects to settings that allow for interactions and non-linearities. In this paper, these approaches from causal inference are further extended to settings in which multiple mediators may be of interest. Two analytic approaches, one based on regression and one based on weighting are proposed to estimate the effect mediated through multiple mediators and the effects through other pathways. The approaches proposed here accommodate exposure-mediator interactions and, to a certain extent, mediator-mediator interactions as well. The methods handle binary or continuous mediators and binary, continuous or count outcomes. When the mediators affect one another, the strategy of trying to assess direct and indirect effects one mediator at a time will in general fail; the approach given in this paper can still be used. A characterization is moreover given as to when the sum of the mediated effects for multiple mediators considered separately will be equal to the mediated effect of all of the mediators considered jointly. The approach proposed in this paper is robust to unmeasured common causes of two or more mediators.

medflex: An R Package for Flexible Mediation Analysis using Natural Effect Models

“…Specifically, if the aforementioned ignorability conditions A1-A4 hold with respect to each mediator separately 11 , natural indirect effects, as defined as causal pathways through single mediators, are identified since these conditions imply that the given mediators are independent given exposure and baseline covariates VanderWeele and Vansteelandt 2013). Recently, Lange et al (2014) demonstrated how independent intermediate pathways can be assessed in a single natural effect model using the weightingbased approach.…”

“…Alternatively, the total causal effect can be decomposed into the effect transmitted through L and M simultaneously and the effect not mediated by any of the given mediators ( VanderWeele and Vansteelandt 2013;VanderWeele, Vansteelandt, and Robins 2014). Although such a joint mediation approach might not target the initial mediation hypothesis, it may still shed some light on the underlying causal mechanisms if there are reasons (either theoretical or empirical) to question the validity of condition A4 (with respect to a single mediator) 12 , since this decomposition relies on a weaker set of ignorability assumptions.…”

“…Specifically, if, as under the NPSEM-IE depicted in Figure 5, we assume that a set of baseline covariates C satisfies 'no omitted variables' assumptions A1-A3 with respect to L and M jointly (rather than separately) and that no measured or unmeasured confounders of the (L, M ) − Y relation are affected by the exposure 13 , the joint mediated effect and corresponding direct effect are identified. The appeal of this joint mediation approach is that by defining a natural indirect effect with respect to a set or vector of mediators (rather than a single mediator) assumption A4 can be made more plausible by simply including mediator-outcome confounders that are deemed likely to be affected by the exposure in the joint set of mediators (VanderWeele and Vansteelandt 2013).…”

“…VanderWeele and Vansteelandt (2013) and VanderWeele et al (2014) described how additional insight into the causal mechanisms can be gained when the ordering is (assumed to be) known. These authors advocated a sequential approach which enables further effect decomposition of the total causal effect into multiple path-specific effects (Avin et al 2005; also see Huber 2013 for an inverse-probability weighting approach and Albert and Nelson 2011 and Daniel, De Stavola, Cousens, and Vansteelandt 2015 for a parametric g-computation approach for estimating some of these path-specific effects).…”

Mediation analysis to estimate direct and indirect milk losses associated with bacterial load in bovine subclinical mammary infections

“…at the same time SCC and milk were collected) was unavailable and was replaced by information on CFU observed at three consecutive dates. If we had obtained information on test-day CFU, the mediation model would have included two 'causally ordered' mediators (CFU and SCC), with test-day CFU affecting test-day SCC (VanderWeele and Vansteelandt, 2014). With such model, it is possible to compute eight estimates of milk loss not mediated by any changes in SCC and CFU, eight estimates of milk loss mediated only by changes in CFU, eight estimates of milk loss mediated by changes in SCC alone and eight estimates of milk loss mediated by changes in both SCC and CFU (Albert and Nelson, 2011).…”