Help using Base SAS procedures

PROC LOGISTIC case-control STRATA can't do ctables/ROC?

Reply
New Contributor
Posts: 3

PROC LOGISTIC case-control STRATA can't do ctables/ROC?

I don't have a lot of experience with matched case-control studies, so I was surprised to see that PROC LOGISTIC doesn't support ROC analysis and ctable options when using the STRATA command. Is there a statistical or mathematical issue with ROC and specificity and sensitivity determinations, or is it simply that they have not been implemented yet (as with profile-likelihood CIs, which aren't available with STRATA, but which tablcan be obtained using the alternative transformed-data method)?
TIA,
Elliot

Contributor
Posts: 62

Re: PROC LOGISTIC case-control STRATA can't do ctables/ROC?

Posted in reply to ElliotSprecher

Hi,

Do you want the ROC for each strata or for overall prediction ?

Contributor
Posts: 62

Re: PROC LOGISTIC case-control STRATA can't do ctables/ROC?

Posted in reply to MohammadFayaz

For both of them, you can use  %rocplot by specifying the right input arguments based on your analysis.

For further information check the results tab in this page :

25018 - Plot ROC curve with labelled points for a binary-response model

New Contributor
Posts: 3

Re: PROC LOGISTIC case-control STRATA can't do ctables/ROC?

Posted in reply to MohammadFayaz
Thanks for the replies. I did try the macro, as you suggested, with a conditional logistic regression with strata. (Overall prediction is the goal.) Unfortunately, SAS also doesn't support OUTROC = when STRATA is used, so the macro cannot be run.
If it is simply a matter of SAS not currently supporting the necessary calculations in the current version, I suppose either other software might, or some rather tedious hand calculation could be employed (to at least produce a ctable from inverse fit of the logistic regression parameters). Of course, it may be that SAS doesn't implement ROC for STRATA for some statistical/mathematical reason which makes the calculation invalid – not that any come to mind, but as I wrote, I don't have a lot of experience with matched case-control studies.
Ask a Question
Discussion stats
  • 3 replies
  • 332 views
  • 0 likes
  • 2 in conversation