Proc nlmixed GEE for ZINB

[rdemass]

Hello,

I am looking to fit a GEE ZINB model using proc nlmixed. I have turned to proc nlmixed as neither proc genmod or proc GEE can handle zero-inflated distributions for GEE analysis.

Following SAS sources, such as note 2 here , I believe I can set up the nlmixed statement correctly as needed. However, I am not sure how to specify clustering. In general, my code looks like:

proc nlmixed data = data;
	/* linear predictor for zero inflation */
	eta_zero = a0;
	prob_zero = 1/(1 + exp(-eta_zero));
	/* linear predictor for negative binomial */;
	eta_mean = b0 + b1*dTLN + b2*dNLLN + b3*dT;
	mean = exp(eta_mean + myoff); 
	IF y= 0 THEN lglk = LOG(prob_zero + (1-prob_zero)*(1+(phi*mean))**(1/phi));
			 ELSE lglk = LOG(1-prob_zero) + y*LOG(phi*mean) - (y+(1/phi))*LOG(1+(phi*mean))
			 + lgamma(y+(1/phi)) - lgamma(1/phi) - lgamma(y+1);
	model y ~ general(lglk);
        random ??? ;
run;

I understand proc nlmixed can utilize a 'random' statement which can include a subject identifier, but I do not want to include any random effects (I am trying to avoid a linear mixed effects model). I would just like to account for clustering/correlation as a subject can provide more than one response. Note, the data is not a longitudinal study setup.

Is there a way to manipulate the random statement to allow for some correlation structure but not to estimate any random effects, setting up for interpretation as a GEE? Or is there another statement/line in nlmixed I can do this?

Thanks!

[StatDave · Accepted solution]

A marginal, GEE type model uses R-side random effects and is not supported in NLMIXED which only allows for G-side effects. That is, the equivalent to the _RESIDUAL_ keyword in the RANDOM statement is not available in NLMIXED like it is in GLIMMIX. The best you can do would be to add random effects for your subjects as you suggest.

[rdemass]

@StatDave  Thanks for the insight. Are you saying there is a way to implement the model I would like in GLIMMIX?

[StatDave]

No, I'm just noting, syntactically and statistically, what NLMIXED lacks and would need to fit that model