Jim Wang writes:
>I have a experiment design problem here and would be
>grateful if anyone can give me some suggestions. The study
>design is an intervention RCT with two arms. The treatment
>groups are not homogenous by strict criteria, but a recent
>change means the introduction of a new factor. I would like
>to stratify according to this new factor, but that would
>increase the sample size to a level too high to reach. On
>the other hand, leaving the factor out and hope the
>randomisation will even the effect of the factor between the
>two arms may be ok, but it may mean I cannot analysis it
>retrospectively (can I? How about the power of the
>analysis?). What would you think?
You may wish to get some advice from a local statistician. Stratification
will usually reduce the sample size.
Also, a randomized trial is never retrospective because you have to
randomize before you intervene.
I suspect that there is a language problem here. What you are describing as
stratification may actually be a subgroup analysis. What you are describing
as retrospective may actually be a post hoc covariate adjustment. These are
just guesses, though.
As a general piece of advice for randomized trials, only stratify on those
things that are very, very important. Things like smoking status in a cancer
trial, or gestational age in a neonatal study. The difficult logistics of
stratification can often outweigh the benefits that it provides. Also, as
you already know, randomization prevents any serious imbalance in your
stratifying variable. So save stratification for the important stuff.
Stratification is more important, perhaps, in an observational study. In a
randomized trial, it doesn't prevent bias much better than a randomized
study, but serves mostly to improve precision. In an observational study,
stratification can help with both better precision and less bias.
Steve Simon, [log in to unmask], Standard Disclaimer.
STATS: STeve's Attempt to Teach Statistics. http://www.cmh.edu/stats
|