Dear Steve, Thanks for your time and response. I will give more detail to the problem (from a theoric point of view and also from the code) In this plot, I have been comparing five parameters in different models, mainly Random intercept Mixed models with distributions Normal-Normal, Binomial-Normal, and Bayesian Binomial-Normal and Binomial_mixture of Normal. The 3 first rows are covariance parameters, the last two rows my response variable a logit of a probability in a multivariate hierarchical model. The blue lines are the estimates, the curves are the profile likelihood (PL). So far I have been using the book of Russell B. Millar - Maximum Likelihood Estimation and Inference_ With Examples in R, SAS, and ADMB as a guide in the coding process. Also the manuals for MIXED, GLIMMIX and NLMIXED. In Proc Mixed, and Glimmix is possible to do a PL of the covariates parameters with this approach (in orange): ods output Parameters=all21_covsesp; proc nlmixed data=ma_all cov; PARMS covsesp=-2 to 2 by .004; bounds s2usens>=0;bounds s2uspec>=0; logitp = (msens + usens)*sens + (mspec + uspec)*spec; p = exp(logitp)/(1+exp(logitp)); model true ~ binomial(n,p); random usens uspec ~ normal([0 , 0],[s2usens,covsesp,s2uspec]) subject=study_id;run; data tdata2;do i = 1 to 501;output;end;run;
data tdata2;set tdata2; if i=1 then do; covp2=-2; end; else covp2 + 0.008; drop i;run;
ods output Parameters=gl.all21_covsesp;
PROC GLIMMIX data=ma_all;
class study_id status;
model true/n = status / noint s cl corrb covb ddfm=bw;
random status / subject=study_id S type=chol G;
covtest tdata=tdata2 / parms;
ods output covtests=ct;
run; proc mixed data=bi_meta NOPROFILE;
class study_id ;
model logit = dis non_dis / noint cl df=1000, 1000, 1000, 1000, 1000, 1000;
random dis non_dis / subject=study_id V VCorr;
repeated / group=rec;
parms (0.9828) (-0.7498) (0.01 to 25.00 by 0.01); ods output ParmSearch=pl_variance2;
run; So far all is easy to follow in manuals and the referred book. The issue now is when I want to explore the probability in the average of my responses variable (logit of the probabilities: msens and mspec in the case of NLMIXED). NLMIXED allows me to provide starting values for my fixed estimations. Like this: ods output Parameters=all21_msens;
proc nlmixed data=ma_all cov ecov QPOINTS=20 TECH=TRUREG;
PARMS msens=-1 to 10 by 0.011;
etc... So far with MIXED or GLIMMIX I only see ways to provide starting values for the covariance parameter, not for the fixed estimation in my hierarchical model. Of course, I can pass the Normal-Normal made in MIXED to NLMIXED, but I was trying to avoid this approach. Thanks in advance
... View more