topic multinomial logistic regression in Statistical Procedures

multinomial logistic regression

knighsson — Mon, 22 Feb 2021 15:14:57 GMT

Hello,

I want to run a multinomial logistic regression for sample survey data.

1. should I use proc surveylogistic statement for the multinomial logistic regression? what is the difference between binary logistic regression and multinomial logistic regression when it comes to the code.

2. How to do the diagnostics? Should I using Pearson residuals?

Thank you!

Re: multinomial logistic regression

SAS_Rob — Tue, 23 Feb 2021 13:47:00 GMT

Yes, if you have a multinomial response with complex survey data then you should use Proc SURVEYLOGISTIC. If it is an ordinal response then you simply need the usual MODEL statement and it will fit a proportional odds model by default. If you want to fit a generalized logit model then you will also need to place the LINK=GLOGIT on the MODEL statement as well.

Currently, the only available goodness-of-fit tests in PROC SURVEYLOGISTIC are found in the default output under the title "Testing Global Null Hypothesis: BETA=0". This is a general limitation with logit models on complex survey data, in that very few measures to assess fit have been defined. You could compute raw residuals, but for a logit model these are not usually very helpful. For now, all that is available are the ones I mentioned along with the c-statistic.

The Hosmer-Lemeshow test, for example, has been shown to be inappropriate in complex sample designs. This is discussed in "Testing Goodness-Of-Fit for Logistic Regression with Survey Data" by Graubard, Korn and Midthune.

Graubard, B.I., Korn, E.L., Midthune, D., 1997. Testing goodness-of-fit for logistic regression with survey data. In: Proceedings of the Section on Survey Research Methods.American Statistical Association,

Re: multinomial logistic regression

knighsson — Tue, 23 Feb 2021 16:15:25 GMT

Thank you so much!

I found the diagnostic is difficult to code for me. I wonder is ROC curve helpful for diagnostic?

Thank you!

Re: multinomial logistic regression

knighsson — Tue, 23 Feb 2021 16:17:06 GMT

OR does that mean when I run the proc surveylogist statement, I dont have to do diagnostic?