It's not clear what your colleague means "to show that there is a change in
the treated group but not the untreated" group. Does your colleague mean
that he wishes to determine whether the difference from baseline is greater
for the experimental treatment group than for the control treatment group?
Or does he actually want to determine whether the difference from baseline
is statistically significantly different for the experimental treatment
*and* whether the pretreatment and posttreatment scores are equivalent
(within some specified delta) for the control treatment? Hypothesis tests
of equivalence are set up differently from those of a difference.
The following assumes that your colleague is interested in determining
whether there is a difference between the experimental and control
treatments, whether by ANCOVA or change-from-baseline. As to the Bonferroni
adjustment: if it's any guidance for your colleague's situation, according
to the rule that I have heard is applied to multiple-endpoint clinical
studies for medical product approval, the Bonferroni or other adjustment
must be made if your colleague will be willing to declare a difference
between treatments if even only one of the four indicators displays a
statistically signficant difference. But no adjustment for multiple
comparisons is needed if your colleague requires all four indicators to
individually display a statistically significant difference before he is
willing to declare a difference between the two treatments.
There are a couple of options for an omnibus test as an alternative, for
example, multivariate techniques (MANOVA) or construction of a so-called
composite endpoint from the four indicators. For these it probably would be
better to use Monte Carlo simulation for power analysis. (See A. H.
Feiveson, Power by simulation. _Stata Journal_ 2:107-124, 2002, for
pointers.) Does your colleague have assumptions for correlations?
Joseph Coveney
Chris Wallace wrote:
----------------------------------------------------------------------------
I have been asked by a colleague for help to perform a sample size
calculation. As this is not an area I'm familiar with, I was hoping
someone in this group might give me feedback to check I'm doing this
correctly.
My colleague plans to sample 2n people, randomized to treated and
untreated groups (each of size n). A single baseline measurement of
four indicators will be taken from all individuals, and he has provided
expected means and standard deviations for these. The treated group
will then be treated, and the measures repeated (again, he has provided
expected means and SDs for these).
He wants to be powered to show there is a change in the treated group,
but not the untreated.
He is not sure how best to analyse these data, but looking at the
documentation for sampsi, I note there are 3 common methods listed -
post, change, and ancova. I assume all would be appropriate, though the
latter two more powerful.
So I need to do
sampsi mean_post_treatment mean_pre_treatment, sd1(sd_post_treatment)
sd2(sd_pre_treatment) pre(1) post(1) meth(all)
r01(corr_pre_post_treatment)
But then I also need to adjust for the multiple testing of four
measures. As a conservative approach, I was planning to just append
alpha(0.0125)
to force a Bonferroni-type correction and maintain a overall 0.05
significance level.
I would appreciate any advice, particularly if I'm making some big
error!
Thanks, Chris.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
