Weighted analysis in Proc NLMixed or other Procedure?

[deleted_user]

Dear fellow sas user,

Are you aware of any ways for sas to run weighted non-linear mixed effect analysis? Proc mixed has a weight statement to minimize weighted residual sum of squares for linear mixed effect analysis. Unfortunately that statement is not available in proc nlmixed. I tried replicate statement but it seems not working for the purpose because degrees of freedoom changes .

Does anyone know the solution?Is there any other procedure that could do the same thing for non-linear mixed effect analysis? I need to do mixed-effect grouped-time survival analysis while using Inverse Probablity Weighting method.

Thanks,

[Dale]

It sounds as if you are able to code the log-likelihood model yourself so that a user-specified likelihood can be maximized in NLMIXED. If that is the case, then you can perform a weighted analysis by multiplying the log-likelihood by your weight function. Thus, you would employ a construction like the following:

proc nlmixed data=mydata;
  parms ...;
 
 
  model y ~ general(weightfunc*LL);
run;

I would note that the REPLICATE statement in NLMIXED is very easy to abuse. If you code

REPLICATE M;

then you are indicating that there are M SUBJECTS who have the same identical values of the predictors and the response. As the SAS documentation states, this sort of condition might apply when you have a binomial response modeled as:

P(Nsuccess{i} | Ntot{i}, p=f(x1{i}, x2{i}, ..., x{i}, gamma{i}) )

where Nsuccess{i} and Ntot{i} are the number of successes and total number of trials, x{i} are values for a set of predictor variables, and gamma{i} is a random effect. All of these are specific to the i-th subject. If you have M subjects that share the same values for the number of successes, total number of trials, and predictor variables, then all of those subjects will be assigned the same random effect. You can drop all but one of the subjects from the analysis in order to improve computational efficiency.

This is a much different model than one would employ for fitting a weighted analysis. The weighted analysis gives different importance to certain contributions to the total log-likelihood. I believe it is the weighted analysis employing user-specified log-likelihood which you wish to employ.