topic Re: SURVEYLOGISTIC Wald chi-square to f-test justification in Statistical Procedures

SURVEYLOGISTIC Wald chi-square to f-test justification

pcl — Fri, 24 Apr 2020 17:04:28 GMT

I recently went back to an older analysis to rerun the code for documentation purposes.

Currently using SAS 9.4 and the code uses PROC SURVEYLOGISTIC, originally run in SAS 9.3 back in 2015. Looking at the saved output from the old run, I noticed that the hypothesis tests used Wald chi-squares, but rerunning the same code in SAS 9.4 now uses the F-stat by default. This leads to different findings and different conclusions now.

I know I can manually set the DF=NONE/INFINITY to produce the chi-square tests again with the same results as the old run, but I wanted to know what the justification was in switching from using a chi-square test to the F-test. I haven't found the documentation for this switch yet.

Re: SURVEYLOGISTIC Wald chi-square to f-test justification

SAS_Rob — Fri, 24 Apr 2020 17:13:05 GMT

This change was in SAS/STAT 14.1 (SAS 9.4TS1M3) and was meant to reflect the fact that the test statistic more accurately follows an F distribution for smaller numbers of DF. As the DF approaches infinity then it can be approximated using a Chi-Square distribution.

There is a mention of this change in the 14.1 documentation

https://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_whatsnew_sect036.htm

And the theoretical explanation is there as well.

https://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_surveylogistic_details37.htm#statug.surveylogistic.splogitdfdetails

Re: SURVEYLOGISTIC Wald chi-square to f-test justification

ballardw — Fri, 24 Apr 2020 17:13:59 GMT

SAS in always working to improve. Changes in output or defaults may reflect requests, check the SASWare Ballot section of this forum for example or perhaps for better results with some other changed default.

If you post your original code it may help to see what might generate closer to the same results. Of course, if you previously ran the code on a 32bit system you may get minor differences than running with a 64bit system and/or operating system.