procedure in reverse, although may be conducted
open-label and more rapidly (guided by suitable
PK information) than when therapy is being
introduced.
Sources of bias in this study design arise from
the exposure of patients to lower doses first.
Patients obligatorily must tolerate, and fail to
respond to, lower doses before being exposed to
higher doses. Any degree of treatment familiariza-
tion, tachyphylaxis or patient withdrawal rate
biases dose–response curves to the right (i.e. tend
to overestimate the ED50) in comparison to a
parallel-group study in the same patients with the
same end points.
Crossover studies
Generally, crossover studies are more complicated
thanparallel-groupdesigns. Patients are exposed to
more than one test medication, in sequential treat-
ment periods, perhaps with periods of no therapy
intervening between those of active therapy. Active
therapies may be different drugs, or different doses
of the same drug, or, in complicated studies, both.
The most famous problem is eliminating carry-
over effects (‘washout’). Ideally, end points should
be measured and unambiguously attributable to
one of the test regimens. This requires no residual
effects of the previous regimen(s) (see Laskaet al.,
1983). If this involves intervening placebo treat-
ment periods in between test medications, then
clearly this approach is not possible when placebos
are ethically unjustifiable.
Usually, patients are randomized to a particular
treatment order, and all patients are eventually
exposed to the same variety of treatments. Large
numbers of treatment periods, assigned using a
Latin square, have been reported; however, the
logistics and patient retention in such studies are
usually difficult, and these ideal designs are likely
to be successful only when treatment periods are
short; ideal designs are commonest for normal
volunteer studies (e.g. Aminet al., 1995).
In later phase studies, if there are still numerous
treatments or dose sizes that need to be tested, then
‘partial crossover’ designs can be used. These
expose patients to a random subset of all the
study treatments, again in a random order. ‘Partial
crossover’ designs necessarily require the avail-
ability of large numbers of patients. However,
there can be economies of the amounts of test
drug needed, and the time needed to conduct the
study in comparison to an equivalent, complete,
crossover design. Shorter durations of patient par-
ticipation are also usually associated with less
missing data and fewer patients lost for adminis-
trative reasons. Overall patient recruitment is more
efficient.
Clinical trialists should be wary of using rando-
mized, crossover designs when there are likely to
be appreciable numbers of patients who are with-
drawn before completing the study. This can cause
serious imbalance among treatment groups and
seriously jeopardize the likelihood of achieving a
statistically robust result. Crossover studies with
three or more periods have a substantial advantage
over two-period designs, when the amount of miss-
ing data is likely to be large and statistical salvage
is necessary (Ebbutt, 1984).
9.10 Minimization trials
Less common are trial designs that specifically and
adaptively minimize the number of patients needed
while preserving design integrity for appropriate
statistical analysis. Early ‘Evolutionary’ designs
are now being succeeded by independent treatment
allocation in pursuit of this goal. All minimization
designs involve arduous statistical planning, and
the clinical trialist should seek expert help from the
outset.
Evolutionary designs were devised by Dixon
and Armitage. Although the statistical analysis is
rather different, they have the same objective,
which is to detect a treatment effect at the earliest
moment possible, using the fewest possible
patients, while retaining statistical robustness.
Both types are suited for exploratory clinical
research and diseases which are rare.
The Dixon ‘Up-Down’ technique was first
described in the statistical literature in 1947. It is
designed to estimate an ED50 in clinical trials or
toxicological tests, when a quantal response is
measured (see Figure 9.1). However, it should be
9.10 MINIMIZATION TRIALS 109