So the answer is: don't do it.
Thanks everyone - and for the ammunition in the form of references.
Responses below.
Best wishes
Sue
From: [log in to unmask]
You could look at Stephen Senn's Statistical Issues in Drug Development
(John Wiley & Sons, 1997) Chapter 7.2, page 98. This explains clearly and
succinctly why baseline hyp. testing is a misuse of significance testing.
I doubt you would find anyone defending it these days.
Regards,
Sean McGuigan
From: Ian Bradbury <[log in to unmask]>
Stephen Senn's book, Isue in Pharmaceutical Statistics, has some
references and comments on this. For what it's worth, I'd argue that the
decision on whether to include a covariate should be based on its
influence, not the imbalance (this is particularly clearly the cadse, of
course, if you look at, say, a Cox model rather than a linear model).
Dr Ian Bradbury
From: "Manktelow, B." <[log in to unmask]>
While applying a statistical test to the baseline data may indicate which
variables to adjust for in the final analysis there are problems with
this. A good (and very short) discussion of this is:
Roberts C, Torgerson DJ. Baseline imbalance in randomised controlled
trials. BMJ 1999; 319: 185.
The main idea is that common sense is better than a statistical test.
Best wishes
Brad
From: "Robert Newcombe" <[log in to unmask]>
I can't give you chapter and verse. But what I understand to be the
perceived wisdom is as follows. When you measure a characteristic X at
times 0 (pre-randomisation baseline) and 1 (after a period on treatment A
or B as determined by the randomisation) for each subject in groups A
and
B, and X is well-behaved i.e. Gaussian, with a linear relationship of X1
to X0 within each treatment group - then the method of choice is to model
X1 on Rx (binary) and X0 (as a continuous covariate). That is, analysis
of covariance using the baseline of the variable in question as your
covariate. Full stop. It doesn't depend on whether X0 differed
significantly or importantly between groups A and B, nor on whether the
correlation of X1 with X0 within groups was significant or important. Nor
do you consider adjusting X1 for some other baseline characteristic Y0 -
unless that was a decision made in advance and set out in the protocol.
Personally, my only reservation about this is that X0 is measured
with error, so that b, the slope of the line used to adjust for the
baseline difference in X0 will be under-estimated. Often b comes out
around +0.5. This is reasonable as adjusting using a slope +1 is the same
as analysing increments, which often over-adjusts for baseline
differences, whereas using b=0 disregards the baseline. But I suspect
that using the estimate of b that comes straight from the data also
under-adjusts, and that a calculated b=0.5 should often really be 0.7 or
thereabouts. You can't estimate the degree of attenuation of the slope
from the data that is usually collected, though if you replicate the
baseline measurements, this should give you the information that is
needed
to tune the adjustment.
There's also possibilities such as non-linear behaviour, or treatment-
baseline interaction, but these are probably not all that frequent.
Hope this helps.
Robert Newcombe.
From: "Andy Vail" <[log in to unmask]>
A good starting point would be Assmann's paper in Lancet last
March, Vol 355, p1064.
The only time I think that hypothesis testing would make sense is
if you were testing to see whether the randomisation procedure had
been fiddled in some way. Otherwise you 'know' the population
difference regardless of your sample estimate.
Apart from this, logically it doesn't follow that differences between
samples will necessarily affect the treatment estimate in any case.
For that you'd need to consider interaction terms, for which you
almost certainly don't have realistic power (unless your name's
Peto).
Best wishes,
Andy
From: "CDSC - Swan, Tony" <[log in to unmask]>
After randomisation imbalances involving the y variable may occur
giving
rise to type I errors. That is what randomisation is for. Imbalances in
explanatory or X variables are the mechanism for that and all the
randomisation guarantees is that these will not take the form of
systematic bias in the X's and as a consequence in the y's. Significance
testing of X differences is inappropriate because it is a) testing nothing
more than the randomisation method itself; and b) it is not important
whether differences in X variables are 'significant' or not - what
matters is whether these differences inflate the underlying error and thus
diminish the power of the study. In most RCTs comparing two treatments
the
primary hypothesis that 'the two treatments do not differ' should be
tested with a simple two group comparison supported by an auxiliary
analysis adjusting for X variables thought to affect the y variable (and
listed in the protocol). It is unwise to make the analysis adjusting for
the Xs the primary analysis because you then get into the problems of
which Xs should be used and how representative these patients are of all
such patients in the X space. Tony Swan
From: <[log in to unmask]>
I don't know if you have already looked at any of the following
references, but if not you should find them useful!
Roberts C, Torgerson J. (1999) Baseline imbalance in randomised
controlled
trials. BMJ; 319: 185
Altman DG. (1996) Better reporting of randomised controlled trials: the
CONSORT statement. BMJ; 313: 570-571
Altman DG. (1985) Comparabillity of randomised groups. Statistician; 34:
125-136
Kennedy A, Grant A. (1997) Subversion of allocation in a randomised
controlled trial. Control Clin Trial; 18 (suppl 3): 77-78S
Senn SJ. (1989) Covariate imbalance and random allocation in clinical
trials. Stat Med; 8: 467-475
Senn S. (1994) Testing for baseline balance in clinical trials Stat Med;
13: 1715-1726.
Hope this is of some help,
JOY
From: Irene Stratton <[log in to unmask]>
Stephen Senn, Statistical Issues in Drug Development is about the best
reference I know of. He is respected, though not universally.......
Irene
___________________________________________
Dr S M Bogle
e-mail: [log in to unmask]
_________________________________________
|