I don't have a software recommendation, but I have some comments
that might be useful. First, I suspect that your intention is misguided
because you have not considered well enough your hypothesis and
what you already know.
What are you trying to learn from your analysis?
Is it really useful, at all, to learn that "Some combination of these 10
variables have a statistically-significant association with the outcome"?
- For that: Look at the size of the effect (X^2, say) that would be
significant for 10 variables; find the corresponding p-level for a 1 d.f.
X^2 (call it p'), and find the sample size, S, for a test of size p'.
This S will be (I'm pretty sure) an underestimate of the necessary
sample size. I would calibrate it by doing similar computations for
simple regression with 1 and 10 variables. Compare the regression
result with the values you can obtain directly for the sample size
needed for regression with 10 variables.
Or, given the marginals of the binary outcome, you can convert the
ORs to their equivalent R-squared values, and try using a power
analysis for multiple regression.
However.
If I had 10 binary predictors that I really wanted to test, I think that
I would have a pretty firm idea of what direction each one of them
predicted. So, if I wanted an efficient test of their combined prediction,
I would construct a total score of how many were scored "good", and
do a test on that. This gives a 1 d.f. test, which you already have
a source for. It gives you an overall test - which you had before - and
it also confirms the direction. So you are already ahead of the game.
- This might actually come in with smaller N that the default minimum
for robust use of logistic regression, since the rule of thumb for LR is
to have 10 or 20 cases per variable for each subject in the *smaller*
of your two outcomes. (With 10 cases per 10 variables, the smaller
group should be at least 100, so your default is already 200 cases.)
If you really want firm tests on 10 variables, separately, what you do
next depends on the narrative that you have. You might be able to
justify using the single-variables tests - either from a combined
regression or for the univariate tests - at the original p-level. Otherwise
you might be stuck with using a Bonferroni correction on the p-level
that you are interested in, and constructing a power analysis based
on that alpha.
Hope this helps.
--
Rich Ulrich
Date: Fri, 28 Oct 2011 17:53:13 -0400
From:
[hidden email]Subject: power cal for logistic with 10 predictor variables
To:
[hidden email]
Dear List. I have logistic regression analysis with 10 binary predictors. I tried to use NCSS PASS to determine the sample size for different effect sizes (OR = 1.5, 2.0, and 2.5), but the program only seems to use one predictor. Does anyone know of any software that would allow one to determine sample size for a logistic analysis with 10 predictors? Thanks,
Locked
