7

I want to assess if a single variable is mediating the effects of a set of IVs on a single DV. All variables in the model (IV, DV and the mediator variable) are dichotomous (0, 1) and observed. What is the appropriate statistical procedure to test this model?

I have run separate binary logistic regressions, and have found that:

  1. IVs are associated with the mediator variable, and
  2. the mediator is associated with the DV. However,
  3. IVs are not associated with DV when removing the mediator variable from the model.
chl
  • 50,972
  • 18
  • 205
  • 364
Marvin
  • 71
  • 1
  • 3
  • 1
    Informally, if you observe those three things, then there appears to be mediation (although, as always, you can't, statistically, rule out the possibility of some sort of confounding creating this pathway spuriously). But, it appears this question is asking for a formal statistical testing framework - is that the case? – Macro Dec 28 '12 at 01:18
  • Your third finding makes this sound more like a case of suppression than mediation. You may want to check out [this article](http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2819361/) on the topic. – Nick Stauner Dec 31 '13 at 22:13

2 Answers2

1

You can easily model this in structural modeling software such as Mplus. You need a model of

X --> Z --> Y

where Z is the mediator and inspect fit and/or residual correlations. If the model fit is poor, then Z may be an imperfect mediator, and residual correlations between the three variables should be inspected to see where residual effects persist, e.g. between X and Y if you have imperfect mediation.

I am suggesting to use Mplus because it can very easily deal with categorical variables in its general linear modeling framework.

tomka
  • 5,874
  • 3
  • 30
  • 71
1

Nathaniel Herr discussed the problem (and solution) of using Logit regression with dichotomous mediator, predictor and outcome in his blog http://www.nrhpsych.com/mediation/logmed.html

STAToSphere
  • 91
  • 4