topic Re: Proc Causalmed: Mediation Analysis in Statistical Procedures

Proc Causalmed: Mediation Analysis

zaldarsa — Fri, 22 Jul 2022 05:21:29 GMT

Hi!

I am trying to do a mediation analysis in which brain volumes are the mediator (continuous), meditteranean diet the exposure (continuous), and cognitive test scores (continuous) the outcome. However, whenver I run the analysis I get the following message in the log and my code:

WARNING: The covariance matrix for the outcome model is not positive definite. The convergence is
         questionable.
WARNING: The covariance matrix for the mediator model is not positive definite. The convergence is
         questionable.
WARNING: Standard errors are not computed for the effect estimates.


proc causalmed data=cog;
   model  icogstroop  = Meddietavg per_TBV25 ;
   mediator per_TBV25 =Meddietavg ;
   covar age0 sex race center education avgpact smoke avgbmi caloriesavg htn anycvd diab2hhba1ccum25;
run;

Thank you in advance for your help!

Re: Proc Causalmed: Mediation Analysis

pink_poodle — Mon, 25 Jul 2022 22:33:49 GMT

Would the warning go away if you comment out the covar statement?

Re: Proc Causalmed: Mediation Analysis

SAS_Rob — Tue, 26 Jul 2022 14:39:06 GMT

The general meaning of the message is that the data does not support either of the causal models. The likely culprit is one or more of the covariates causing a singularity in the model. Unfortunately it is sometimes difficult to know which one(s) so in practice the easiest thing to do is a process of elimination by starting with just a single covariate and then adding from there until the problem occurs.

Re: Proc Causalmed: Mediation Analysis

zaldarsa — Wed, 03 Aug 2022 03:57:20 GMT

Yes, the error message did go away, so I decided to just do the mediation analysis without the covariates!

Re: Proc Causalmed: Mediation Analysis

pink_poodle — Wed, 03 Aug 2022 16:25:49 GMT

That is because of overfitting. It is usually one or two of them that have a lot of categories. Like @SAS_Rob said, you can try adding them back one-by-one and leave out the trouble-makers.