1
$\begingroup$

I was asked this question and I am not sure about my answer that is why I am asking it here. An institute wants to conduct a clinical study. This study has most likely three treatment arms. The design is not yet set up, nor is any information given about the efficacy, endpoints, estimates, methods and so. The (randomized) study basically checks if the weight has decreased. In most simple case this would be a simple mean comparison, but of course also Kaplan-Meier-Estimates and other statistics (ratios and so) can come into play. The study itself is a simple one. So we are not talking about a new drug or whatever. It is from sports life science, basically how can get people to live healthier and it checks if a certain training schedule reduces the weight. The other arm next to placebo is schedule plus addition nutrition consulting.

The problem now is that of course the costs should be estimated, but especially the budget is limited and the question arised how many subjects are needed. Because each subject needs to be assigend to an educated trainer and each subject therefore crucially increases the costs. Of course my first answer was it depends on the design, the estimates and so. But I had to give a specific number. My answer was it should be not less than 30-40 subjects per treatment arm. My thoughts were that if it is more or less some simple mean comparison from my personal point of view it should be sufficient. If I think about weight and comparing weights between groups, what personally is sufficient for me? If I only have 10 subjects to compare this is not enough. 20 already would make T test and so stable, from my point of view. I thought if there are 30-40 subjects per treatment group, this should be sufficient. But later on I thought about Kaplan-Meier plots and estimates (log rank test, Cox-regression, odds ratios to come into play?), especially I do not have any information about the event rate. So I don't know what exactly they will specify - this is also more or less a question to me - and I thought that 30-40 is maybe not sufficient if we talk about comparing weights.

So is there somehow a rule-of-thumb? The problem is really to give a number. With Survival analysis I am not that much used to (those studies I had this never was a problem), so I cannot estimate how many subjects or events are needed to get these plots and estimates working ? So what is the minimum sample size for a clinical study? Also how many events do I need?

$\endgroup$
1
$\begingroup$

If you want to study change in weight, then a Kaplan-Meier curve is irrelevant. That just studies the time till an event occurs. Since you don't have a discrete event, this methodology does not apply.

Before you start investing time and effort into this project make sure you actually can do an experiment on humans. Depending on your jurisdiction the bar can be quite high. See, e.g. https://en.m.wikipedia.org/wiki/Institutional_review_board

To answer your question: you need to perform a power analysis: https://en.m.wikipedia.org/wiki/Power_(statistics) .

$\endgroup$
  • $\begingroup$ Thanks for your answer. Your comment regarding Kaplan-Meier and change in weight is correct, however I was not 100% precisely here. The main point is indeed the weight situation, however this also focusses on dependent factor, like time until a surgery is needed or time until another specific happening occurs. So it could arise that Kaplan-Meier and other survival analysis methodologies and not only a simple weight change check will come up. Yes, I am aware of the ethical point of a clinical study and informed consent and committee. $\endgroup$ – Stat Tistician Dec 10 '19 at 18:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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