Dear allstat,
It would seen to me that significance testing of baseline variables can be
interpreted in two ways.
If one accepts that one's randomization method was in fact random, then
any significant difference in baseline characteristics would seem, by
definition, to be a type 1 error.
If one questions one's randomization method, then significance testing of
baseline characteristics would seem to be a test of the hypothesis: The
randomization was random. For a large number of significance tests, the
hypothesis would be disproved if the proportion rejecting exceeded the
nominal level (e.g. if 7% reject at the .05 level). For a finite number
of significance tests, the Birnbaum statistic can be used to examine the
hypothesis; Feller (1966, p.26) gives the distribution.
Sorry to wade in on an already well aired issue, but on this side of the
Atlantic we often don't get to see the question until after the summary of
replies has been posted.
Andy Dunning
--------------------------------------------------------------------------
Andrew J. Dunning
Department of Biostatistics
University of Washington
--------------------------------------------------------------------------
|