The abuse of power: pitfalls
of over-design
Equation (6) is expressingNin terms ofsand.It
is very easy to rewrite this equation and express any
of the three parametersN,sandin terms of the
other two. If we expressin terms ofNands,we
could see that by increasingN, the statistical test
could have high power to detect very small differ-
ences. Thus, it is sometimes tempting to ‘over-
design’ the study; that is, to enroll more subjects
than required so that if the drug is not quite as
efficacious as one hopes, the statistical test would
still be significant at the end; sort of buying an
‘insurance policy’. By enrolling a large number
of subjects, one can assure that the statistical test
is so powerful that it would declare very small and
possibly meaningless differences as statistically
significant. This approach is not only wasteful
but may also lead to false inferences, and is outright
unethical in the drug development arena. Clinical
trials are very expensive enterprises, and it is typi-
cally not feasible to repeat a trial to demonstrate
that the results are reproducible. Furthermore, in
studying therapies for life-threatening diseases, a
trial resulting with a significant outcome often
precludes the possibility of conducting a second
confirmatory trial. The variables studied in clinical
trials are random, thus there will always be differ-
ences between the treatment groups that are due to
chance. An over-powered study could find such
differences statistically significant and lead the
researcher to a false conclusion that a drug is
efficacious when it is not, or that it is harmful
when it is not. In the absence of a second chance,
these finding may never be repudiated.
An underpowered trial is wasteful and unethical
for a different reason. Such a trial may not have
enough power to detect clinically meaningful dif-
ferences resulting in missing clinically important
medical advances. The subjects enrolled into such
a trial, are exposed to the risks involved in all
clinical trials using experimental drugs, without
the anticipated benefit to themselves and to society.
For these reasons, it is important that the size of
the trial is just right: not too small and not too large.
The discussions taking place among the project
research team leading to the appropriate choice
of the sample size are therefore very important,
and although at the end, it is the statistician who
performs the calculations, the input from the other
team members is critical.25.11 Issues in statistical trial
designMulticenter trials
Most phase III clinical trials are multicenter trials;
that is, they are conducted in more than one clinical
center. The number of centers participating in a
clinical trial can vary greatly.
There are a number of good reasons to conduct
phase III trials as multicenter trials. The most
obvious is an administrative and logistical reason.
Spreading theburden of subject recruitment among
many centers will reduce the duration of the subject
enrollment phase of the trial. This is an important
reason considering that often the key to commer-
cial success or failure of a new drug is the timing of
itsintroductionto the market.There are also impor-
tant scientific reasons to conduct the trial as a
multicenter trial.Noise reductionDifferent centers often draw subjects from differ-
ent types of patient populations. Also, different
centers may utilize different procedures and med-
ical practices that are not controlled by the study
protocol. It is, therefore, reasonable to expect that
the within-center variability is smaller than the
overall variability. In a multicenter trial, the center
often serves as a stratification variable, thereby
reducing the variability and increasing the effi-
ciency of the trial design. In order to take advantage
of this aspect of the multicenter trial, the number of
subjects per center cannot be too small so that the
estimate of the intra-center variability is stable. A
rule-of-thumb is that the number of subjects
per treatment group within each center will be at
least 5.25.11 ISSUES IN STATISTICAL TRIAL DESIGN 333