topic Re: Fisher's Exact in Proc Surveyfreq in SAS Programming

Fisher's Exact in Proc Surveyfreq

eabc0351 — Tue, 19 Nov 2019 18:11:09 GMT

Hello. I would like to run a fisher's exact test. However, my data is weighted so I am using proc surveyfreq to obtain p-values for my contingency tables. The exact option will not work with the proc surveyfreq command. Any suggestions?

I should also add that one of my cells is 0. The nosparse option can suppress the cell but I still cannot get a p-value with the chisq option.

Thank you.

Re: Fisher's Exact in Proc Surveyfreq

PaigeMiller — Tue, 19 Nov 2019 21:16:03 GMT

I don't think that there is such a thing as Fisher's Exact test for survey data, as the existence of the survey weights means you aren't dealing with integers in the analysis, which is what Exact tests require (but shhhh don't tell Sir Ronald)

Re: Fisher's Exact in Proc Surveyfreq

eabc0351 — Thu, 21 Nov 2019 18:50:43 GMT

Yes, that is the error message that I am getting from SAS. @PaigeMiller do you know any solution for getting a p-value with survey data when one of the cells is 0?

Thank you.

Re: Fisher's Exact in Proc Surveyfreq

PaigeMiller — Thu, 21 Nov 2019 20:09:06 GMT

PROC SURVEYFREQ has other ways to obtain confidence intervals. There are many choices within the CL option in PROC SURVEYFREQ.

https://documentation.sas.com/?docsetId=statug&docsetTarget=statug_surveyfreq_syntax06.htm&docsetVersion=15.1&locale=en#statug.surveyfreq.sfreqclwt

Re: Fisher's Exact in Proc Surveyfreq

Watts — Thu, 21 Nov 2019 20:22:42 GMT

When there's an empty (zero-frequency) table cell, the Rao-Scott chi-square can't be computed because some quantities are undefined/missing. Here's one possible work-around, if you want to use zero as the standard error for the missing cell(s).

1) Construct a data set that contains one observation for each empty table cell, and assign a small weight (relative to the real survey weights) to each of these observations -- e.g., 1e-8.
2) Merge that data set with the analysis dataset. Because of the very small weights, the additional observations should not affect the estimates of proportions and totals. But all table cells will have positive frequency and SURVEYFREQ will compute the Rao-Scott chi-square test.

Re: Fisher's Exact in Proc Surveyfreq

eabc0351 — Thu, 21 Nov 2019 22:54:06 GMT

That worked perfectly without changing the proportions, as you said. Thanks!