Date: Thu, 16 Jun 2011 14:16:30 -0400
Reply-To: Rich Ulrich <email@example.com>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Rich Ulrich <firstname.lastname@example.org>
Subject: Re: help me
The usual proper use of a power analysis is to describe, even before
data collection, what the chances are that your tests will be able to
detect ("find significant") effect sizes that are (a) interesting, and
(b) at least somewhat likely to be seen, given the range of Ns that
is being considered.
In your instance, it *might* be reasonable that someone could
be asking that you show, for your sample, how small of an effect
could have been detected, say, with even 50% power. If some of
your tests were not significant, it could be interesting to remark about
how large of an effect could have been missed. (I also consider it
possible that the person who asked for your power analysis is an idiot,
who is mindlessly echoing, inappropriately, jargon that he has heard
in the past. It possible to make some descriptive statement, after the
analysis, referring to power, but I agree with the poster who says that
giving confidence limits is often going to be more informative, to
A sample size of 600 is "large" for what Cohen was interested in, for
the usual studies in psychology and social sciences, and their usual
effects. However, there are other topics where 600 is far too small.
The study which showed that aspirin after a heart-attack is a good
idea was the first "mega-study" using multiple sites, and it was
designed from the power analysis which showed that 20,000 patients
was a good number, if one wanted to show that a treatment that
would cut in half the immediate fatalities was worthwhile.
What is an "effect size"? For a dichotomy like a dummy variable,
predicting another dichotomy, the odds-ratio is a good measure of
effect. However, the Odds Ratio is sensitive to the actual sizes of
the proportions, so you often need to anchor the description in the
actual proportions. That is, there is a 20-point difference between
40% and 60%, or between 10% and 30%; but the latter is a more
severe difference than the former.
For your study, your purpose is some sort of illustration -- Obviously,
you had *enough* power for some analyses, since you did have
positive findings. I think I would perform the power analyses one
variable at a time. That makes the presentation easier and clearer,
and it is often the approach to start with, even the power analysis is
done before the experiment.
Date: Wed, 15 Jun 2011 21:00:31 +0000
Subject: Re: help me
Dear Gene and Evan,
Thank very much for your help me.
I got a model (GEE) with 5 independent variables (2 and 3 were categories were continuous covariates), these are all significant. Not consider covariates did not influence my model. As this analysis is an analysis of a secondary database, I have requested that I submit the statistical power. My English is not very good, I hope I could make myself understood.
My understanding is that the sample is large gives a high statistical power, so I do not understand is referred to the effect size.
Thanks in advance.