topic Re: Phreg and multiple groups and violated PH ratio in Statistical Procedures

Phreg and multiple groups and violated PH ratio

toyodn01 — Mon, 07 Dec 2020 17:55:50 GMT

Hello, I have a question about proc phreg with multiple groups and violated PH ratio.

I would like to compare the re-intervention on patients who had a procedure. There are three procedures (Group=0,1,2) and the outcome is reintervention (0 ,1).

Comparison of interest is 1) Group 1 vs 0 and 2) Group 1 vs 2.

Proportional hazard ratio is violated, so I would like to include interaction term to evaluate hazard ratio at each time point (5, 10, 15 years after procedure).

I created two dichotomous variables to express Group. Group 0: A=0, B=0, Group 1: A=1, B=0, Group 2: A=0, B=1.

I think contrast statement should be used, but I am not quite sure what interaction term is included.

proc phreg data=X;
model Time_to_reintervention*reintervention(0)=A B interaction/rl ;

interaction=??;
contrast 'time=5' A ? B ? interaction ? /
estimate=exp e;
contrast 'time=10' A ? B ? interaction ? /
estimate=exp e;
contrast 'time=15' A ? B ? interaction ? /
estimate=exp e;

run;

I would greatly appreciate if if you could help...

Thank you very much.

Re: Phreg and multiple groups and violated PH ratio

SteveDenham — Tue, 08 Dec 2020 13:20:45 GMT

I am going to pull a Ksharp trick here and ask someone else who knows quite a bit about PHREG. So calling @FreelanceReinh

SteveDenham

Re: Phreg and multiple groups and violated PH ratio

FreelanceReinh — Tue, 08 Dec 2020 15:25:58 GMT

Thanks, Steve. Sadly, I have never dealt with the case of non-proportional hazards. By a quick web search I found two conference papers on this particular subject:

[1] suggests an "interaction of particular covariate with a function of time variable" (which might be what the OP is trying to implement) or, alternatively, a "stratification model."

[2] uses a Bayesian approach.

Re: Phreg and multiple groups and violated PH ratio

toyodn01 — Thu, 10 Dec 2020 19:42:20 GMT

Thank you both very much for your help! I really appreciate it.