Type 3 Tests in Cox Regression

[AJones2020]

Hello. I ran a Cox Regression in which the convergence was satisfied. In the Type 3 Tests portion of the output, two of my independent variables (coded categorically), have dots where the Wald Chi-Square and Pr >ChiSq results should be. I can't figure out why SAS did not calculate results for these two variables. I think it may be due to how the sample is dispersed between the two categories on these variables, but not sure. What factors can prevent the calculation of Wald Chi-Square? 

[Reeza]

Are those your reference levels?

[AJones2020]

Hi, Reeza. The variables are coded as no (i.e., 0) and yes (i.e., 1). Don't know if this answers your question or not.

[SteveDenham]

You identified what I think is the most probable reason for the lack of tests. You could have complete separation. Test some different approaches. You can get score based and likelihood ratio based Type3 tests.  You can also get Type1 tests.  (Well, at least you can try to get those tests, so long as you are careful about using the COVSANDWICH option).  What does a simple PROC FREQ reveal about the pattern of the data?

SteveDenham

[AJones2020]

Hello, Steve. Based on PROC FREQ results: Variable 1- 4% no (n= 28) and 96% yes (n= 672). For Variable 2- 5% no (n= 37) and 95% yes (n= 663).

[SteveDenham]

What about the cross-tab?  Do you get down to 1 or 2 obs in the no-yes or no-no cells?

SteveDenham

[AJones2020]

I have 1 zero cell count in both of the crosstabs. That may be the issue, right?

[SteveDenham]

I think you found it. My recommendation would be to try to run everything with that record excluded/censored to see if it does identify the problem. It may turn out that something entirely different is causing the problem, but this kind of thing has been known to create issues in some sorts of analyses.

SteveDenham

[AJones2020]

Sounds good! Thanks, Steve!

[SteveDenham]

It doesn't sound so good on second reading.  If a cell has a zero in the cross-tabs, there is no record to remove - that's why it is a zero.  So, we are stuck a bit.  Unless we assume some distribution for the times (like a Weibull or a gamma) then there are a couple of approaches (LIFEREG or GENMOD come to mind) that may be more or less useful.

SteveDenham.