topic Power Analysis for Logistic Regression in Statistical Procedures

Power Analysis for Logistic Regression

SasTinker — Tue, 28 Mar 2023 13:50:40 GMT

I conducted an analysis using complex weighted survey data, running both trivariate (stratified bivariate) and stratified logistic regression models. Stratification variable was race.

I want to do a post-hoc power analysis to estimate the number of respondents needed to see a hypothetical effect size. I basically want to make sure the population for each race is sufficiently large for analysis.

I've never done a power analysis, much less use SAS for it. I've been trying to browse online for details and examples, but I'm still lost.

Do I need to use Proc POWER or Proc GLMPower? What exactly is the "dataset" I need to be referring to--the dataset I used for the original analysis, or do I need to create a new dataset--if so, what does that dataset need to include?

Any kind of guidelines or suggestions would be greatly appreciated.

Re: Power Analysis for Logistic Regression

PaigeMiller — Tue, 28 Mar 2023 14:20:44 GMT

How about the LOGISTIC statement in PROC POWER?

Details and example: https://documentation.sas.com/doc/en/statcdc/14.2/statug/statug_power_syntax17.htm

Re: Power Analysis for Logistic Regression

SteveDenham — Wed, 29 Mar 2023 13:44:05 GMT

I like @PaigeMiller 's answer (LOGISTIC option in PROC POWER) but want to point out a particular pet peeve of mine. A post hoc calculation of power is pretty useless, except as a jumping off point for a planned future survey/experiment. Here is the reasoning that I was taught: You already know the p value of interest. Depending on your decision rule, then you either detected a difference (power=1) or you did not (power=0). It's why effect size is a major consideration in your interpretation of results.

So, as a prospective tool, you can use the effect size observed, a desired power or desired sample size, and an alpha level, you can calculate the remaining free variable. But that is for a collection of new samples from a population with characteristics similar to those in hand, not for the current sample from such a population. As I mentioned above, you already know the power for your current sample.

And you have to consider that the error estimates and point estimates from your weighted survey analysis have a built-in finite adjustment, which PROC POWER doesn't have. In my opinion, that means simulation is probably the best way to calculate prospective power/sample size.

SteveDenham