topic Re: Goodness of Fit for Multinominal regression in Statistical Procedures

Goodness of Fit for Multinominal regression

jjjunyi — Thu, 24 Aug 2017 17:44:15 GMT

Hey all,

so if i am using a multinominal procedure (proc logistic with Class), how shall i test the goodness of fit? i suppose they are not the same as the binary logistic regression (i.e., AUC, BIC,Breier Score).

so any ideas on that?

Re: Goodness of Fit for Multinominal regression

PGStats — Thu, 24 Aug 2017 18:22:06 GMT

Make sure you understand what the model is, i.e. how to interpret the estimates of logistic multinomial regression (when the response is not binomial). The model fitting information is described in the Model fitting information section of the Logistic Procedure documentation. It applies to binomial and multinomial models.

Re: Goodness of Fit for Multinominal regression

jjjunyi — Thu, 24 Aug 2017 19:30:57 GMT

could you give me some example for that?

Re: Goodness of Fit for Multinominal regression

PGStats — Thu, 24 Aug 2017 20:46:36 GMT

Read the documentation! You will find example

http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_logistic_examples04.htm

Re: Goodness of Fit for Multinominal regression

StatDave — Thu, 31 Aug 2017 19:28:35 GMT

Note that the presence of the CLASS statement does not make the model a multinomial model. In PROC LOGISTIC, the model is a multinomial model if the response variable has more than two distinct values. Further, a multinomial model can be a ordinal or nominal multinomial model depending on whether the multiple levels of the response have a natural ordering or not. PROC LOGISTIC fits an ordinal model by default (using cumulative logits) when there is a multilevel response. If the response level are not ordered, then the LINK=GLOGIT option in the MODEL statement can be specified to fit a nominal multinomial model.

Now to your question. As with the binary logistic model, you can use the SCALE=NONE and AGGREGATE options in the MODEL statement to obtain Pearson and deviance statistics which are valid goodness of fit tests only if there is sufficient replication within the subpopulations defined by the distinct covariate patterns in the data. For purposes of comparing competing models you can use the R-square statistic available with the RSQUARE option in the MODEL statement. Also for comparing models, you could use the AIC and SC statistics which are provided by default. Of course, you can always save predicted probabilities from the model using the PREDPROBS=INDIVIDUAL option in the OUTPUT statement. Similarly, you can score a new set of data and save predicted probabilities using the SCORE statement. See the example titled "Scoring Data Sets" in the LOGISTIC documentation.