|
|
Parallel Group Designs
Parallel study designs are preferable when there
are strong concerns about carryover effects, or when the natural history of the
disorder makes progression likely during the period required for a crossover study.
Within-patient variability is the major problem posed by parallel group designs,
and several approaches have been suggested to mitigate its impact (Lavori
et al., 1983). For example, baseline pain scores may be subtracted from the
treatment scores to yield pain intensity difference scores, or they may be treated
as a covariate. This often eliminates a large part of the variance, thereby increasing
the power of treatment comparisons.
The investigator should also
make an effort to balance the treatment groups for variables
that predict response, whenever these predictors are known
or suspected. If one wishes to examine response in specific
subgroups, assignments must also be balanced appropriately.
Groups can be balanced using stratification or various techniques
of adaptive randomization (Therneau,
1993; Friedman et al.,
1996). In studies with sample sizes typical of single
center trials, 20-40 patients per group, these methods can
significantly increase the power of a study if the prognostic
variables are well chosen and the statistical methods take
the balancing method into account (Lavori
et al., 1983). With sample size above 50 per group,
the randomization process alone is likely to balance out
most variables (Meinert, 1986).
If stratification is not feasible, post hoc covariate analyses
or other statistical techniques may be an acceptable substitute
if the variables in question are distributed fairly evenly
among the treatment groups.
|