topic Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f in Statistical Procedures

Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model fits

ting1 — Wed, 15 Feb 2023 16:48:25 GMT

I used a multinomial logistic regression to predict whether people have confidence on a certain issue.

The dependent variable has four categories

1	(1)not confident
2	(2)neutral
3	(3)confident
4	(4) unknown

Independent variables include Identities, age, gender, education attainment, employment status, born in a certain place or not, community (urban or rural) and interaction terms.

Following are the outcomes of the model:

Model Fit Statistics
Criterion	Intercept Only	Intercept and Covariates
AIC	11678.775	11421.602
SC	11698.052	12019.195
-2 Log L	11672.775	11235.602

Testing Global Null Hypothesis: BETA=0
Test	Chi-Square	DF	Pr > ChiSq
Likelihood Ratio	437.1734	90	<.0001
Score	456.9761	90	<.0001
Wald	396.1174	90	<.0001

Deviance and Pearson Goodness-of-Fit Statistics
Criterion	Value	DF	Value/DF	Pr > ChiSq
Deviance	4083.7242	4359	0.9368	0.9987
Pearson	4494.1027	4359	1.0310	0.0750

The Deviance and Pearson Goodness-of-Fit Statistics show that P-value for Deviance is high. However, the P-value for Pearson statistics is low, even it greater than 0.05. Can I draw a conclusion that the model fits the data well?

Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f

StatDave — Wed, 15 Feb 2023 17:45:45 GMT

You didn't provide your PROC LOGISTIC statements or indicate if your multinomial response is ordinal or nominal. So, I have to assume that you treated it as nominal and therefore used the LINK=GLOGIT option. I also assume that you used the AGGREGATE option along with the SCALE=NONE option to get these tests, and since their DF are so large, that some of your predictors are continuous. As noted in the Details:Goodness of fit section of the LOGISTIC documentation, the Pearson and deviance statistics require sufficient replication within the populations in order to be valid and that substantial difference between the two are an indication that neither can be used. With one or more continuous predictors there usually is very little, if any, replication within the populations (the populations are defined by the unique settings of the predictors). See note 22630 (https://support.sas.com/kb/22/630.html) which goes more into assessing goodness of fit. As suggested there, you could use the Hosmer-Lemeshow test to assess fit.

Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f

ting1 — Wed, 15 Feb 2023 18:05:06 GMT

Thanks for your answer!!

Yes, I used the LINK=GLOGIT option and AGGREGATE along with the SCALE=NONE option to get these tests. I tried the Hosmer and Lemeshow Goodness-of-fit test with the option lackfit, and the test result did not show up in the output.

Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f

StatDave — Wed, 15 Feb 2023 19:17:13 GMT

As mentioned in the note I referred to, the Hosmer-Lemeshow test is available for the multinomial model beginning with SAS 9.4M3. The current release is SAS 9.4M8.

Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f

ting1 — Wed, 15 Feb 2023 19:43:06 GMT

This is my SAS version:

Copyright (c) 2002-2012 by SAS Institute Inc., Cary, NC, USA.
NOTE: SAS (r) Proprietary Software 9.4 (TS1M2)
Licensed to JUSTICE CANADA, Site 70169614.
NOTE: This session is executing on the X64_8PRO platform.

I guess this version cannot show the "lackfit" test.

Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f

StatDave — Wed, 15 Feb 2023 20:42:03 GMT

Check with the people that administer your SAS license and see if they have a more recent release available to you.

Re: Logistic regression: using Deviance and Person Goodness-of-Fit Statistics to test if the model f

ting1 — Sun, 19 Feb 2023 16:13:47 GMT

Thanks for your help!