topic Re: alprocsurvey logistic (multinomial regression) in Statistical Procedures

alprocsurvey logistic (multinomial regression)

amng — Sat, 28 Nov 2020 21:54:44 GMT

Hi all,

For a class project I am modeling the impact of a binary health literacy variable (getadvice2) on the likelihood of completing up to 5 self-management behaviors for diabetic patients (manage5). I originally tried to run an ordinal logistic regression, but following the violation of the proportional odds assumption I instead modeled a multinomial regression. However, the odds ratios under "point estimate" are not showing up properly. I wanted to ask whether this is some kind of SAS Error? There is nothing out of the ordinary in the log.

PROC LOGISTIC DATA=diabetes;
class getadvice2(ref=first) manage5(ref="0") / param=ref;
MODEL manage5(Desc) = getadvice2 / link=glogit;
title "Model 1";
RUN;

Re: alprocsurvey logistic (multinomial regression)

PaigeMiller — Sat, 28 Nov 2020 22:27:46 GMT

When a SAS user says that there is an error in SAS, I generally believe SAS, and that the error is the user's fault, not SAS's fault.

But unless you can share your data with us, there's no way we can resolve this.

Does this help? https://support.sas.com/kb/22/954.html

I also think you should not use your response variable MANAGE5 in the CLASS statement, but I don't know if that makes any difference here.

Re: alprocsurvey logistic (multinomial regression)

amng — Sun, 29 Nov 2020 16:28:24 GMT

Hi thanks. I wasn't trying to suggest there was an error in SAS itself - just trying to understand why it was giving me these results. The data is attached here.

Re: alprocsurvey logistic (multinomial regression)

amng — Sun, 29 Nov 2020 16:28:50 GMT

Sorry - the dataset is too large to be posted. What is the best way to upload it in that case?

Re: alprocsurvey logistic (multinomial regression)

amng — Sun, 29 Nov 2020 16:30:25 GMT

Re: alprocsurvey logistic (multinomial regression)

PaigeMiller — Sun, 29 Nov 2020 16:54:26 GMT

We need you to show us a portion of the data (not the whole thing) by providing the data in this format and not any other format.

Re: alprocsurvey logistic (multinomial regression)

FreelanceReinh — Sun, 29 Nov 2020 18:34:31 GMT

@amng wrote:

PROC LOGISTIC DATA=diabetes;
class getadvice2(ref=first) manage5(ref="0") / param=ref;
MODEL manage5(Desc) = getadvice2 / link=glogit;
title "Model 1";
RUN;

Hi @amng,

Look at a cross tabulation of getadvice2 and manage5 (PROC FREQ). If the cell getadvice2=0 & manage5=0 is empty, the estimated odds of manage5=5 (and likewise 4, 3, 2, 1) vs. manage5=0, which you chose as the reference category, will have a zero in the denominator for getadvice2=0, leading to the "zero" odds ratio estimates in your table. In this case, pick a different reference category for manage5, i.e., one with non-empty cells for both getadvice2 categories, so that non-trivial odds ratio estimates (for manage5 ne 0) have a chance to occur.