Thanks for taking the time to answer my query. Your e-mail was helpful
and I'll be seeking out the papers you mention re: the D'Amico and SEM
I was under the impression that DFA was preferable to logistic
regression as it has more power (or so I have read). Thus, when data
meets the correct assumptions, DFA is to be used.
Jeremy, would you mind elaborating on what you mean exactly by "run
I had a feeling point biserial r would be equivalent to normal pearsons
r power analyses. Regarding partial correlation though, you said to use
the calculations for regression, but would it not be hierarchical
regression, as the partial correlation procedure is analogous to the HMR
procedure ( i.e., looking for an effect after controlling for the
influence of another predictor or set of predictors)?
>>> "Jeremy Miles" <[log in to unmask]> 23/02/2007 17:59:09 >>>
On 23/02/07, Kathryn Jane Gardner <[log in to unmask]> wrote:
> Dear list,
> Does anybody know of any resources for power anaysis for two-group
> discriminant function analysis, or know how of the formula to do the
> power analysis? Either hand calculations/formula, internet based
> calculators or sample size tables etc in journals will do. I can't
> to find anything, which is a shame as i've used power analysis for
> my other analyses.
Unless you really, really have a good reason, then don't use
discriminant functions. Use logistic regression instead. DFA is more
complex, and much more sensitive to distributional assumptions.
It's possible to do power analysis for logistic regression as well.
If you're really desperate, and need power for DFA, you could run some
simulations. The problem is that the model parameters in DFA are a
bit fiddly and weird (at least, I always find them fiddly and weird).
> I also need to find out the power calculations for: point-biserial
> correlation, partial correlation & partial point-biserial
Point biserial correlation is a correlation, so you can use standard
correlation power analysis. Alternatively PB is just another way of
thinking about a t-test, so use t-test power instead.
For partial correlation, the p-value is the same as the p-value you
get for a regression (try it) so you can use standard regression
If you're ever really desperate for power and you are doing something
weird, then there are three solutions.
1) Simulate it. This can be a bit fiddly (especially in SPSS, it's
much easier in Stata or R). The book Data Analysis Using Regression
and Multilevel/Hierarchical Models by Gelman and Hill describes how to
do this in R. (It's not desperately straightforward, unless you're
already familiar with R.)
2) Use MANOVA. Most analyses can be converted into a MANOVA design.
In SPSS MANOVA you can input summary statistics (means, correlations,
regressions) and analyses the data, to get a power for any sample
size. There's a paper that describes this: D'Amico, E. J.,
T. B., & Zambarano, R. (2001). Power analysis for multivariate and
repeated measures designs: a flexible approach using the SPSS MANOVA
procedure. Behavior Research, methods, instruments and computers, 33,
479-484. The book Statistical Power Analysis, by Murphy and Myers
expands on this. Basically, every test is turned into an F test, and
they are all powered the same way.
3. Use a structural equation model. Almost every analysis can be
thought of as a structural equation model, and power in structural
equation models is always estimated in the same way - by reducing
everything to a chi-square, and chi-squares are (relatively) easy to
work out the power for. If you're already familiar with SEM, this is
a bit fiddly, if you're not, then it's very fiddly. I wrote a paper
on this method, which can be found here:
Although I do a lot of power analysis, and I've only used it once
since I wrote the paper. (That's how fiddly it is). I use the
D'Amico method more. (Coincidentally, Liz D'Amico works just down the
corridor from me now - I didn't know about her paper when I wrote
mine, or I might not have bothered.)
Learning statistics blog: www.jeremymiles.co.uk/learningstats