development team to detect drug toxicity early, but
not efficacy. Similarly, depending upon the hypoth-
eses under test, control charts can be rhomboidal,
parallelogram or of many other shapes. Whitehead
(1999) is the best entry to the literature on this
specialized topic.
Contemporaneous independent
treatment allocation
Taves (1974) has described a study design that
requires an independent coordinator who allocates
each patient, as he or she is recruited to one or other
treatment group. The independent coordinator
allocates each patient so as to minimize the differ-
ence between the two treatment groups according
to prospectively defined patient characteristics, for
example, age, sex, genotype, disease state or stage,
or concomitant therapy. This allocation is therefore
also based upon the cumulating characteristics of
the treatment groups as has developed during the
study to date. Patients are therefore not allocated to
a treatment group by the chance of a randomization
schedule.
Bias in minimization trials can be avoided when
three conditions are met. Firstly, those performing
the clinical trial itself, that is administering test
medications and measuring end points, should be
double-blind and unaware of which treatment the
patient has received. Secondly, the independent
coordinator need only allocate patients to anon-
ymous groups A or B, and the study pharmacist
need be the only person who knows which treat-
ments these codes represent. Thirdly, the criteria
for which the treatment groups should be balanced
must be prospectively identified and rigidly
adhered to, using a recorded, quantitative system
of scoring the factors.
In its simplest form, this class of minimization
designs usually results in treatment groups of
nearly equal size. By equitably assigning patients
to three or more treatment groups, and yet having
identical treatments for two or more of these,
unbalanced sample sizes can be created. This is
of use when, for example, it may be desirable to
expose fewer patients to placebo than to active
therapy, especially when conducting a trial of com-
pounds whose properties are fairly well known or
may be predicted with some confidence.
Note that minimization trials can only alter
power calculations when assumptions of the size
of worthwhile differences in effect are also pro-
spectively defined. For example, from a clinical
point of view, a small-sized improvement in out-
come (perhaps a few percent of patients more than
that observed for placebo treatment) may be
viewed as very worthwhile in an extremely hetero-
geneous patient population when subjected to mul-
tivariate analysis (this is common in large, simple
studies; see below). On the other hand, when
designing a minimization study, the assumption
is that the treatment groups will be devoid of rele-
vant differences in baseline characteristics and,
therefore, clinical significance might only be
assumed to follow from a large-sized difference
in patient response. The size of the difference that
is assumed to be of interest, as it increases, may
compensate for the reduction in variability
amongst study group samples, and thus have less
than expected impact on the sample sizes needed to
conduct the clinical trial.
Minimization designs are probably under-used
by the pharmaceutical industry. This approach is
not well designed for pivotal clinical trials nor for
diseases with large numbers of prognostic factors,
where, in any case, large numbers of patients are
especially needed for a tolerability database. If the
controlled clinical trial is a gold standard, then it
would be wrong to assert that the independent
treatment allocation design is the ‘platinum stan-
dard’ (paceTreasure and MacRae, 1998). The
interested reader is referred to a good published
example (Kalliset al., 1994), and to more detailed
statistical treatments (Pocock and Simon, 1975;
Freedman and White, 1976).
9.11 The ‘large simple study’
and stratification designs
These similar classes of study require large num-
bers of patients. The choice between them lies in
being able to ‘hedge one’s bets’ with a partial
indication approval, versus ‘all or nothing’ with
huge logistical costs and potentially huge rewards.
9.11 THE ‘LARGE SIMPLE STUDY’ AND STRATIFICATION DESIGNS 111