topic Re: Conditional Logistic Regression, Approximation Method in Statistical Procedures

Conditional Logistic Regression, Approximation Method

longitudinal — Tue, 10 Jan 2017 02:29:36 GMT

Dear Community,

I am wondering if SAS has similar approximation procedure when fits the conditional logistic regression as R does?

I.e. In R, there is an option to specify likelihood approximation method as "efron" or "breslow".

I don't think PROC LOGISTIC has this option.

Thank you!

Re: Conditional Logistic Regression, Approximation Method

Reeza — Tue, 10 Jan 2017 04:58:02 GMT

There are several proc that perform logistic regression, PROC GLM and PHREG are others.

This may be what you're looking for:
http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_phreg_examples05.htm

Re: Conditional Logistic Regression, Approximation Method

JacobSimonsen — Tue, 10 Jan 2017 11:34:22 GMT

It is true that proc logistic can not make the Breslow or Efron approximation. But, construction af variable "dummytime" with values 1 for cases and 0 for controls, and analyze this as survival analysis is equivalent to conditional logistic regression when using the option ties=discrete in PROC PHREG. If, instead, the option ties=breslow (default) or ties=efron is used, then you have the approximation.

Re: Conditional Logistic Regression, Approximation Method

JacobSimonsen — Tue, 17 Jan 2017 11:45:01 GMT

Sorry, I made a typing error. dummytime should be 1 for cases and 2 for controls (not 0). And the dummytime should be censored, so no riskset after 1:
model dummytime*dummytime(2)/ties=discrete

Re: Conditional Logistic Regression, Approximation Method

longitudinal — Wed, 18 Jan 2017 00:47:17 GMT

I assume it should be the same as

model dummytime*Y(0)/ties=discrete

given that Y is binary with value 1 and 0?

Re: Conditional Logistic Regression, Approximation Method

JacobSimonsen — Wed, 18 Jan 2017 07:40:15 GMT

Yes, that wil work also. If those with dummytime=2 is the same as those with Y=0, then

model dummytime*Y(0)/ties=discrete

is equivalent to

model dummytime*dummytime(2)/ties=discrete

given of course that the rest has dummytime=1.