Descending option ignored in Proc Genmod

[chrnie]

I am in the early modelling phase and am initially just trying to set up a univariate model (modified Poisson in Proc Genmod) regressing initiation of breastfeeding (0/1) on exposure category (3 categories with varying degree of exposure to a pollutant). The coding is such that initiation=1, but am more interested in modelling initiation=0 (i.e. the default in proc genmod). However, just to make sure the results came out right I reran the model with the descending option. To my surprise the results are identical. I also realize that SAS no longer states what is being modelled as in the good old days.

I would really like to know what I am missing here. Greatful for any input!

First run

Proc GenMod data=final_primi ; 
Class ind expcat/param=ref ref=last; 
Model initiation= expcat / Dist=poisson Link=log type3; 
Repeated subject=ind/ type=exch;

run;

Second run

Proc GenMod data=final_primi desc; 
Class ind expcat/param=ref ref=last; 
Model initiation= expcat / Dist=poisson Link=log type3; 
Repeated subject=ind/ type=exch;

run;

[Rick_SAS]

Perhaps I am misunderstanding your model, but it does not look like a Poisson model to me. If the response variable is binary (0/1), then change DIST=POISSON to DIST=BIN. The DESC option now does what you expect it to do.

FYI, you can also bypass the DESC option on the PROC GENMOD statement and specify the order of the response variable by using the MODEL statement. Either of the following will change the order of the target level of the response variable:

Model initiation(event='1') = ... ;

or
Model initiation(DESC) = ... ;

[MichaelL_SAS · Accepted solution]

As described in the documentation for the DESCENDING option, it applies only to models with a binary or ordinal multinational response. PROC GENMOD determines the response type by the distribution function that is used, so if you are using a modified Poisson approach and have specified DIST=POISSON the response is not treated as a categorical variable with ordered levels and the DESCENDING option is ignored (there also is no Response Profile table in the displayed output). 

I believe for the problem you are interested in you would have to create a new response variable with a 0/1 coding where 1 corresponds to the event you want modeled. 

[chrnie]

I figured the human factor was involved here... Thanks so much for helping me out!

[MichaelL_SAS]

Glad to help, and good catch on your part to notice the difference between what you had and mind and what the procedure was doing.