Brent Beasley writes:
>In reading NEJM's article by Poldermans et al (The Effect of
>Bisoprolol on Perioperative Mortality and Myocardial Infarction
>in High-Risk Patients Undergoing Vascular Surgery, December 9,
>1999 -- Vol. 341, No. 240), I noticed something that has struck
>me before in randomized trials.
>
>There were 50-60 patients in each arm of the placebo controlled
>trial. This was enough patients to show a statistically
>significant difference in their endpoint (cardiac death and
>nonfatal MI). BUT, in the characteristics of patients who began
>the study, more patients in the standard-care group had "limited
>exercise capacity" (43% vs 27%). Although to me this difference
>appears "clinically" significant, it did not reach statistical
>significance because of the relatively small number in each group.
>
>It would seem that forethought should be done in sample size
>calculations to avoid having a "clinically" important difference
>between groups.
I haven't had a chance to look at the Poldermans et al article yet, but I
have to share a story that seems to be related to your question.
One of my first studies at Children's Mercy Hospital in Kansas City was a
randomized trial for breastfeeding (bf) in pre-term infants. The infants
were randomized into an intervention group or into a control group. One of
the outcome measures was duration of bf.
There are lots of covariates that can influence duration of bf: birth
weight, gestational age, marital status of the mother, type of delivery
(vaginal vs C-section), length of stay in the birth hospital, and so forth.
It's unreasonable to expect that randomization will result in a near perfect
balance across each and every one of these covariates. In this study, the
mothers in the treatment group had an average age of 29 years, compared to
25 years for the control mothers. And a four year gap in average age is
clinically important. Older moms are generally more successful at bf.
So we had to worry about whether the large difference in duration of bf that
we saw between the treatment and control groups was caused by the imbalance
in average mother's age. The solution was to adjust for mother's age in all
of the statistical models. It turns out that the gap in the duration of bf
decreased a bit after adjusting for mother's age, but it was still large and
achieved both clinical and statistical significance.
So the bottom line is that even with randomization, you can't expect all
your covariates to divide themselves up nicely. If there is one covariate
that is really critical, such as smoking status in a cancer study, then you
should try to match on this variable if possible. But if there are a dozen
or so covariates, don't have unrealistic expectations. A large sample size
will not necessarily prevent problems with covariate imbalance.
I talk a bit more about this study on my web pages, though the discussion is
more on the mechanics of data analysis. But if you are curious, look at the
"Steps for developing a statistical model" section of my web pages.
Steve Simon, [log in to unmask], Standard Disclaimer.
STATS - Steve's Attempt to Teach Statistics: http://www.cmh.edu/stats
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|