In a standard randomised clinical trial you will have a primary hypothesis
for which the study is powered. Randomisation should give you balanced
groups, but is not guaranteed. In practical terms the analysis to be
performed must be defined a priori. Previous comments have suggested that
it is influence, rather than imbalance, of a covariate on this primary
hypothesis that matters. But surely a balanced covariate will have little
impact, regardless of its influence. This suggests a two-stage process:
Identification of whether variables are imbalanced, followed by determination
of their prognostic relevance.
In a pre-defined analysis plan, use of hypothesis testing as a diagnostic
tool, followed by a clinical review of prognostic significance, provides
a clearly defined process. I welcome any suggestions how an approach
without hypothesis testing can be clearly defined a priori without leaving
open accusations of data dredging?
Many thanks
----------------------------------
Andrew Charlesworth
Head of Statistics
Nottingham Clinical Research Group
[log in to unmask]
Tel: 44 115 9567711
Fax: 44 115 9220960
----------------------------------
|