topic Re: proc logistic - generalized logit model in Statistical Procedures

proc logistic - generalized logit model

SAS49 — Sun, 05 Jun 2022 19:08:31 GMT

Hi, I have a predictor variable Color with 3 categories (0, 1, 2) and an outcome variable Sound_Level with 3 levels (low, medium, and high). I am trying to build a model that also accounts for a 3rd binary variable Size - (0, 1). Before adding that variable, I had been running an ordinal logistic regression with a proportional odds assumption, however, when I add the binary variable to the model the proportional odds assumption doesn't hold. I want to now run a generalized logit model, but I cannot figure out how. Here is the code I have tried, but it isn't working. How can I fix this?

proc logistic data=my_data descending;
	class Color(ref = '0');
	model Sound_level=Color Size link=GLOGIT;
run;

Re: proc logistic - generalized logit model

StatDave — Sun, 05 Jun 2022 20:02:35 GMT

The generalized logit model that treats the response as nominal (unordered) is fit when you add the LINK=GLOGIT option in the MODEL statement. However, you also have the option of fitting a nonproportional odds model that treats the response as ordinal (ordered) but relaxes the proportional odds assumption - either for all predictors or just some. This is described and illustrated in this note: http://support.sas.com/kb/22954

Re: proc logistic - generalized logit model

SAS49 — Sun, 05 Jun 2022 20:29:23 GMT

Great, thank you! Is that the better model to fit then?

Re: proc logistic - generalized logit model

StatDave — Sun, 05 Jun 2022 20:39:56 GMT

It is if you want to make use of the ordered nature of the response. But with the ordinal model, you must be certain that the response levels appear in logical ascending or descending order in the Response Profile table. Also, most people are more used to the 0,1 coding of categorical predictor variables. To do that, specify the PARAM=GLM option in the CLASS statement. You might also want to specify the reference level for your SIZE variable.