Interpretation of multinomial longitudinal regression analysis

[UcheOkoro]

Hello,

Please, I need help interpreting the results of a multinomial longitudinal regression analysis. The outcome Complete_adherence2 variable has 4 levels which include ; complete adherence(1) (reference), hospital only adherence(2), outside hospital only adherence(3) and no adherence (4).   The predictor is ethnicity with levels; Hispanic (1) (reference) and non Hispanic (2). This data was collected weekly over a 21-week period.

The following are the results and SAS code.

Thank you!

 PROC GENMOD DATA=COVERED.Demographics desc;
CLASS  complete_adherence2(ref='4')  record_id ethnicity (ref='1');
MODEL complete_adherence2 = ethnicity 
/ DIST=MULTINOMIAL LINK=cLOGIT;
REPEATED SUBJECT=record_id/ corr=indep;
lsmeans ethnicity/  ILINK diff e cl oddsratio ;
RUN;

[StatDave]

These results should not be used and are not meaningful. When you fit an ordinal model, it is critical to examine the Response Profile table to verify that the response levels are in logical ascending or descending order. That will not be the case because you specified both the DESCENDING option and REF= for the response in the CLASS statement. It is best to never specify the response variable in the CLASS statement. Assuming that you want to model the probabilities of higher levels of the response variable, removing the response variable from the CLASS statement should show the response levels in the order 4, 3, 2, 1 in the Response Profile table. Then in the LSMEANS table, the Mean column gives you the predicted cumulative probabilities (Pr(4), Pr(4 or 3), Pr(4 or 3 or 2)) for each predictor level. The Differences table gives you the odds ratio estimate comparing the two predictor levels.

[UcheOkoro]

Thank you for your response. The response profile is in order and as follow;

Please, how would you interpret the odds ratio for the predictor? If it were a 2 level outcome variable that is pretty straight forward but this has 4 levels.

Thank you,

[StatDave]

Now that the response levels are in logical order, you can use the results. As I stated, the odds ratio estimate is the estimated change in odds comparing the two predictor levels. This is the same as for a binary response ... since this is a proportional odds model the change in odds is the same regardless of the logit.

[UcheOkoro]

Please, could you interpret the odds ratio I posted earlier to make it clearer?

Thank you,