PROC CAUSALMED with poisson distribution and log link

[lindst801]

Hello, I was wondering if someone had experience working with count data and attempted to use PROC CAUSALMED

I am working with count data and I am attempting an mediation analysis with x=continuous m=binary and y=binary. From the Baron and Kenny method, we were able to see complete mediation, but am attempting to gain statistical inference with PROC CAUSALMED and entered the following syntax:

PROC CAUSALMED data=x all;

class _med_bin/ref=last;

model _outc_bin= _exp_cont _med_bin/dist=poisson link=log;

mediator _med_bin = _outc_bin;

run;

However, I only was able to get the effects (TE, CDS, NDE, and NIE) already exponentiated (as if they were odd ratios) as opposed to being raw estimates with the confidence intervals already having 1.00 within it (Ex. total effect being 1.01 with 95% CI of 0.95 - 1.08) but showing a significant z-score (p < 0.001)

Even if I walked backwards and manually calculate from the transformed effects, I am confused why 1) how to obtain the standard error untransformed and 2)  the confidence intervals cross the null (the lower CI's below 1.00) while the p-values were calculated to be significant.

How do I reverse the transformation of the effects within SAS, since it appears that I did not manually calculate from the output correctly? Would there be another way to assess for mediation outside the procedure? I attempted to use the %mediation macro and was unable to get an output that way as well. 

Thank you

[ChrisNZ]

Moved to the procedures forum.

[SteveDenham]

I think this might be something to bring forward to the developers of the PROC as "an undesirable feature" of the program, unless someone could jump in from SAS with a suggestion..

It pains me to suggest this, but what would happen if you changed from the canonical log link to the identity link, but still kept dist=poisson?  How much do estimates change?  What happens to the EVALUATE endpoints?  I only suggest this as it should put all of the output on an untransformed scale.

SteveDenham

[lindst801]

Thank you for getting back to me!

Per your suggestion, I changed the link from log to identity and receive the foillowing error:

ERROR: The specified link function in the MODEL statement is not currently supported for the distribution family specified

If this is a programming issue, how could I forward the problem to the developers of the PROC?

[SteveDenham]

I feared it wouldn't accept it, but it was the only thing I can think of.

So, this should go to SAS Technical Support.  Write up your issue, include code, logs and output, and also indicate your SAS version and operating system. At the end, indicate what you have tried, and what your frustration points are.  Someone associated with Analytics will get back to you to indicate what they think might be the cause of the problem, and what might be done to work around it.  They may point to a hotfix that corrects the issue. And they may say, that is as good as we can provide now, but your issue will be included in the scope of our next upgrade.  They are here https://support.sas.com/en/technical-support/contact-sas.html 

and they are good.

SteveDenham