Hello Chen,

I am not sure if mediation models could be tested in logistic regression. Did you run a separate mediation model?

Regards

Hi, Orlik,

Thanks for the quick reply.

Yes, I ran separate regression models to test for mediation according to the four step approach proposed by Barson and Kenny (1986).

In step 1: A (the main exposure; categorical variable) was significantly associated with the outcome  (binary variable).

In step 2: A was significantly associated with M (the hypothesized mediator; categorical variable).

In step 3: M  was significantly associated with the outcome.

However, in the final step, neither A nor M was significantly associated with the outcome when I put them together into the logistic model. 

So I wonder whether M is a potential mediator (or just correlates with A), and it may be inappropriate to adjust M in the exposure-outcome model. 

Best,

Chen 

Hey Orlik,

Check  variable M for interaction and confounding in the multivariable model. if it is not either of it then it would be better to remove it from the final model.

Regards,

Dr Nida

Hi,

I also suggest for checking if there is confouding in your model firstly: to check if M is a confounder of the relationship between your outcome and exposure A.

If M and outcome are conditionally independent given A

OR

M and A are conditionally independent given the outcome,

THEN M is not a confounder.

Otherwise, M is a confounder.

If it is a confounder, it is necessary to adjust for it in the logistic model where A is a covariate.

There are many possible explanations, including parameter redundancy and multicollinearity.

I would probably start by looking at muticollinearity. Are the predictors correlated?

You may want to read about the 3 Ms :)

mediators

moderators

multicolinearity