PROC NLMIXED random effects model produces estimates that are too low.

RonLev — Fri, 02 May 2014 21:56:11 GMT

I am using PROC NLMIXED to produce a poisson mixed model (I'm using this method instead of GLIMMIX or GENMOD because I need the correct confidence intervals for the Pre/Post difference in means). The first model is the fixed effects model, the second model is the mixed model:

FIXED EFFECTS MODEL (matches results from GLIMMIX, GENMOD)

PROC NLMIXED data=prepost;

PARMS b0=-1.6 b1=0.1;

eta = b0 + b1*prepost2;

lambda=exp(eta);

model countRx=poisson(lambda);

estimate 'pre level ' exp(b0);

estimate 'post level' exp(b0 + b1);

estimate 'pre-post difference' exp(b0+b1)-exp(b0);

run;

RANDOM EFFECTS MODEL (to adjust for correct pre post estimates)

PROC NLMIXED data=prepost;

PARMS b0=-1.6 b1=0.1 s2u=0;

eta = b0 + b1*prepost2 + u;

lambda=exp(eta);

model countRx=poisson(lambda);

random u~normal(0, s2u) subject=enrolid

estimate 'pre level ' exp(b0);

estimate 'post level' exp(b0 + b1);

estimate 'pre-post difference' exp(b0+b1)-exp(b0);

run;

Here's the problem:

The fixed effects model gives these estimates:

parameter estimates

bo -1.5994

b1 0.09337

estimates

period countRx

pre 0.2020

post 0.2218

pre-post 0.01977

But the random effects model gives these estimates:

parameter estimates

bo -6.4813

b1 0.09337

s2u 26.2858

estimates

period countRx

pre 0.001532

post 0.001682

pre-post 0.000150

We would expect some changes because of the switch from independent measures to a repeated measures model, but nothing like this. What is the problem? Is the model specification correct?

Ron Levine

Message was edited by: Ronald Levine:

Since I've been advised the the above code is correct for the Poisson with random effects model, there is another issue at hand: the large random effects variance (s2u=26) may be diminishing the mean estimation. (I tried the same model using PROC GLIMMIX and the estimation was not affected. Why? ) My question to the community is: Should we use a NegBin distribution instead of Poisson in PROC NLMIXED? Or, (since 92% of the count outcomes are zeros), should we use a ZIP or ZINB distribution instead? How would these alternative models be coded in PROC NLMIXED? Ron Levine

Re: PROC NLMIXED random effects model produces estimates that are too low.

SteveDenham — Fri, 09 May 2014 12:45:10 GMT

Hi Ron,

I would definitely consider ZIP or ZINB with that many zeroes. A good place to search for code to do this is in the SAS-L archives. Look for posts by Dale McLerran. There is an especially good one dated 14 April 2003.

Steve Denham

topic Re: PROC NLMIXED random effects model produces estimates that are too low. in Statistical Procedures

PROC NLMIXED random effects model produces estimates that are too low.

Re: PROC NLMIXED random effects model produces estimates that are too low.