Table 1.

Summary of Strengths, Weaknesses, and Required Assumptions of Each Technique.

Technique and overview	Strengths	Weaknesses
Comprehensive structural equations models	Requires clarification of when and how variables are related in a comprehensive model	Assumption: Each mediator is measured equally well so that the variance accounted for by each mediator is not due to one variable’s superior measurement
Include measures of all relevant confounding variables in the statistical analysis of mediation	Includes (possibly latent) measures of competing, alternative, and potentially confounding mediators to help identify the true mediator	Assumption: No unmeasured confounders
	X, M, and Y can be continuous or dichotomous although randomization of X is likely dichotomousa
Instrumental variables	Not all confounders need to have been measured	Assumption: Randomization of X leads to changes in M such that the stronger the relation of X to M, the better the instrument. The ideal instrument has a correlation of 1.00 between X with M, which makes X statistically, but perhaps not conceptually, indistinguishable from M
Assuming no direct effect of X on Y, uses the effect of X on M to predict Y from M	X, M, and Y can be continuous or dichotomous although randomization of X is likely dichotomous to represent a random predictor	Assumption: exclusion restriction, is that M completely mediates the effect of X on Y such that when M is considered there is no relation between X and Y.
Principal stratification	Clearly addresses that there are different types of participants with different responses to an experimental manipulation.	Assumption: Monotonicity assumption is made such that there are no backward-changers because it is unlikely that the mediator would change for participants in the control condition because they are unexposed to the experimental manipulation to influence the mediator
Classification of different possible individual response patterns for how X affects M and M affects Y		Assumption: exclusion restriction that the entire effect of the manipulation on the dependent variable is through the mediator. Assumption: X and M dichotomous
Inverse probability weighting	Does not require that the entire effect of X on Y is through M	Assumption: No unmeasured confounders
Use observed covariates to measure confounder effects and adjust analyses to remove the confounder bias	Confounder effects can be removed from the mediated effect, therefore providing a cleaner estimate of the mediated effect

Note: More detail on the methods can be found in the citations for each method.

^aNeed potential outcome approaches for the most accurate estimation of mediation with combinations of categorical and continuous M and Y (Imai, Keele, & Tingley, 2010; Valeri & VanderWeele, 2013) and see also the more recent version of MPIus (Muthen & Muthen 2012) for causal mediation effect estimation in a structural equation modeling framework.