topic Re: PROC POWER TWOSAMPLESURVIVAL estimate hazard ratio in Statistical Procedures

PROC POWER TWOSAMPLESURVIVAL estimate hazard ratio

patrick5 — Mon, 29 Jul 2019 14:33:16 GMT

I was asked to help out with planning a study of cancer survival based on transcriptome analysis. As part of that, the PI has requested a power analysis.

We will be able to enroll 300 subjects over three years, with an additional year of follow-up and minimal to no losses to follow-up. However, there is no reasonable way to estimate survival times or hazard ratios for the two groups, as the groups will not even be defined until at least mid-way through the study, so anything I enter for GROUPMEDSURVTIMES or HAZARDRATIO (or similar) would be making numbers up entirely out of thin air. Instead of a "typical" power analysis, i.e. estimating how many subjects needed (or the expected power given the number of subjects), I am wondering if it is possible to instead estimate the hazard ratio given the sample size, 1:1 allocation, power of 0.8, and duration of study. From the documentation it appears that HAZARDRATIO is not amenable to this, but I'm hoping there is a workaround or some other means of generating something useful.

PROC POWER;
   TWOSAMPLESURVIVAL TEST = LOGRANK
      HAZARDRATIO = .
      ACCRUALTIME = 3
      TOTALTIME = 4
      NTOTAL = 300
      POWER = 0.8
   ;
RUN;

Re: PROC POWER TWOSAMPLESURVIVAL estimate hazard ratio

Reeza — Mon, 29 Jul 2019 15:25:34 GMT

There's absolutely no prior research on this topic to give you any estimates of HR to get started?

However, there is no reasonable way to estimate survival times or hazard ratios for the two groups, as the groups will not even be defined until at least mid-way through the study, so anything I enter for GROUPMEDSURVTIMES or HAZARDRATIO (or similar) would be making numbers up entirely out of thin air.

Re: PROC POWER TWOSAMPLESURVIVAL estimate hazard ratio

patrick5 — Mon, 29 Jul 2019 15:28:58 GMT

I have an overall median survival time for the cancer in question, but (for a variety of reasons) absolutely no way to estimate how that survival might differ between groups.