topic Change the mediator distribution in Proc Causalmed in Statistical Procedures

Change the mediator distribution in Proc Causalmed

Giampaolo — Mon, 14 Sep 2020 11:59:14 GMT

Dear SAS users,

By default Proc Causalmed assigns a normal distribution to the mediator variable and an identity link. Is it possible to customize these parameters? The actual distribution of my mediator variable is closer to a gamma distribution.

Re: Change the mediator distribution in Proc Causalmed

SAS_Rob — Mon, 14 Sep 2020 13:17:42 GMT

PROC CAUSALMED assumes a normal distribution and the identity link function for a continuous mediator, and a binary distribution and the logit link function for a binary mediator. Unfortunately there is no way currently to change this default behavior.

Re: Change the mediator distribution in Proc Causalmed

Giampaolo — Mon, 14 Sep 2020 13:20:06 GMT

Thank you.

Re: Change the mediator distribution in Proc Causalmed

SteveDenham — Mon, 14 Sep 2020 17:44:34 GMT

If your mediator looks like it may have a gamma distribution, consider a close approximation such as a lognormal. To do this, take the natural log of the mediator variable - this will have a normal distribution of residuals in the log space and has a default identity link in the other PROCs that address alternatives to the normal distribution. It then becomes even more important that you consider how you report the mediating effect.

SteveDenham

Re: Change the mediator distribution in Proc Causalmed

Giampaolo — Mon, 14 Sep 2020 18:07:54 GMT

Thank you for your suggestion!

Then how should I interpret the output given that the outcome is binary?

In summary I would have:

a binomial distribution and logit link for the outcome.

a normal distribution and identity link for a log transformed mediator.

The output would be given in Odds Ratios or excess relative risk, for the direct, indirect and total effects.

Can I still express mediation per unit change of the mediator, rather than its log transformed value? If yes how?