Great disparity in binomial regression odds ratio between proc genmod and proc logistic

[UcheOkoro]

Hello,

Please, I need help figuring out why there is a great disparity in the binomial regression odds ratio between proc genmod and proc logistic. I would also like to know the more reliable result.

The data was obtained after matching. The following are my codes and results.

Thank you!

proc genmod data=COVID.COVID_TRAUMA_MATCHED2 descending;
class  subclass COVID_19_Positivity(DESC) ; 
model  In_hosp_mortality(event='2')= COVID_19_Positivity
/ dist=bin ;
 repeated sub=subclass /type=cs corrw covb printmle;
 lsmeans COVID_19_Positivity/ diff exp cl or;
run;


proc logistic data=COVID.COVID_TRAUMA_MATCHED2; 
CLASS COVID_19_Positivity(DESC); 
strata  subclass;
   model In_hosp_mortality(event='2')= COVID_19_Positivity;
  run;

The result from proc genmod

The results from proc logistic

[StatDave · Accepted solution]

It is because those are entirely different models - one a Generalized Estimating Equations logistic model using GEE (not maximum likelihood) estimation, and the other is a conditional logistic model that maximizes a conditional likelihood. Differences should be expected when fitting such different models using differing estimation algorithms. Importantly, the GEE model is a population-averaged model while the conditional logistic model is a subject-specific model. Yet another possible approach is the random-effects model as can be fit in PROC GLIMMIX which is also a subject-specific model. If you want to make inferences about population differences, then the population-averaged GEE model is more appropriate. If you want to make inferences about, say, change over time (if the repeated measures are over time), then a subject-specific approach can be used. This distinction in model types and estimation methods is discussed in many texts and papers. One paper available online is Hu et. al. (1998), "Comparison of Population-Averaged and Subject-Specific Approaches for Analyzing Repeated Binary Outcomes."

For these and other kinds of logistic models, see this note.

[UcheOkoro]

Thank you so much!