1
$\begingroup$

I have a Clinical Trial where I'll be comparing a drug vs placebo.

Then I intend to run a Log-rank test on it.

Now for the sample size calculations, I would be requiring the hazard ratio and the median survival time for the treatement and placebo.

Now my confusion is that how would I get the median survival time without the dataset, beforehand ?

Do I need to consult previous studies to find the median survival for treatement and placebo or is there any other method.

Please help me with my confusion.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
2
$\begingroup$

Don't use a log-rank test. This is just a special case of the Cox proportional hazards model and doesn't allow for covariate adjustment. You need to adjust for pre-specified predictors that explain a clinically important amount of outcome variation to get the proportional hazards assumption to be better satisfied.

The power for a comparison using the Cox model (and hence also using the log-rank test) is not a function of the median survival time. It is a function of the expected number of events in the first group and the expected number of events in the second group. To get the expected number of events you need to make various assumptions, usually based on the cumulative incidence at a fixed time point. The R Hmisc package cpower and especially the spower functions help. spower does a clinical trial simulation where various complexities can be handled (non-constant hazard ratio, non-constant hazard ratio, complex accrual rate variation, etc.).

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you for explaining and the suggestion. @Frank Harrell $\endgroup$
    – OJS
    Commented Mar 2, 2022 at 16:58
  • $\begingroup$ What software are you using? I'm used to seeing these power calculations done from an event driven analysis where, say, you conduct primary analysis after 350 PFS events - obviously you may plan to enrol 500 and wind up with 300 PFS events after LPLV. But the approach gives different N when you plan to conduct an analysis, say, 5 years after first site active. $\endgroup$
    – AdamO
    Commented Mar 9, 2022 at 0:01
2
$\begingroup$

You would indeed want to consult previous (ideally recent with similar background therapies, as medical care may change over time) studies regarding what you would expect on placebo. If there are no previous similar trials, then non-trial data (e.g. dedicated observational databases/registries, electronic medical records databases, medical claims data etc.) could be alternatives. If hardly any data exists or its relevance is unclear, one can also try to elicit the judgments of a set of experts.

For the treatment group this would rather be an assumption and/or a number for which you want your study to have good power (e.g. a minimum important clinical difference, or perhaps something larger that constitutes a "great drug").

Also note, that given assumption about the placebo group survival time distribution (of course, just a median survival time does not define a whole distribution, a hazard ratio for drug vs. placebo and the assumption that the hazard ratio is constant over time, you already have the whole distribution for the treatment group.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thank you for the ellaborate explaination @Bjorn $\endgroup$
    – OJS
    Commented Mar 2, 2022 at 16:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.