Jeromy Anglim's Blog: Psychology and Statistics

Wednesday, November 12, 2008

Causality and Mediation Analysis

This post discusses issues with causal inference in mediation analysis. It proposes a set of steps that researchers can use when analysing and reporting mediation analyses.
I have previously critiqued the use of mediation analysis in psychological research. Many researchers write things like "the study showed that the effect of the Independent Variable on the Dependent Variable  was mediated by the Mediator Variable", where you can replace IV, MV, and DV with your variables of interest. A few seconds thought can usually yield an alternative causal explanation.

For example, in self-efficacy research some researchers have argued that self-efficacy mediates the effect of past performance on future performance. While this may or may not be true, an alternative explanation is that past performance causes self-efficacy and future performance with self-efficacy being largely epiphenomenal. The resulting pattern of correlations between the three variables would indicate support for the partial mediation model (past performance leads to self-efficacy leads to future performance). However, this may be a completely incorrect understanding of the causal mechanisms involved.

My tentative procedure for justifying causal inference in mediation
Simplistic conclusions of mediation that ignore alternative causal explanations are a problem in psychology. They are likely to be related to not properly considering the nature of causal mechanisms. Thus, I am now working out in my own mind, what would be an appropriate set of thoughts about causal mechanism.
Here is a tentative procedure:

  1. Express your best theory on HOW the IV causes the MV. 
  2. Express your best theory on HOW the IV causes the DV
  3. Think about all the ways that the IV could cause the DV and whether the MV is likely to be one of those intermediary variables. If there are multiple plausible mediators are they all on the same sequence of causation? 
  4. Think about alternative causal explanations, which are essentially rival hypotheses. In particular this includes: a) causation is in the opposite direction to what you think (DV causes IV; MV causes IV; DV causes MV; etc.); b) causation is reciprocal (IV causes DV and DV causes IV); b) Variables not included in your model cause some of your variables (e.g., IV, MV and DV) and this produces a pattern of correlations which could be confused with mediation.
  5. If the statistical tests support complete or partial mediation, authors should write something like "the study showed a pattern of correlations consistent with the mediational hypothesis that the effect of IV on DV is mediated by MV". Past research will usually provide substantial information about the plausibility of various causal mechanisms. IV=Independent Variable; MV=Mediator Variable; DV=Dependent Variable

Additional Resources

1 comment:

  1. What if I have a model, theoreticaly justified, where my IV can cause my MV and my MV can cause my IV (I say theeoreticaly justified because it is an open debate in the theory on these specific variables). And, say, I found support for both paths by following the Baron and Kenny steps, but with smaller size effects for one than the other. Should the mediation hypothesis remain and what can be said about it?