Power analysis for poisson regression with one interaction effect

[Jennifer Carr]

Can anyone help point me in the right direction here? I need to conduct a power analysis for poisson regression models that include an interaction term between two binary predictors. Ideally, I'd also like to be able to estimate this with an added fixed effect (as the sample will be drawn from four different sites), however, at this point I'd settle for just being able to get power estimates for the interaction models without an added fixed effect. Is there a relatively easy way to do this (for a non-statistician)? Maybe an existing package or even another program outside of Stata? I've been reading through the PowerSim tutorial and some other materials I've found related to this package but I'm afraid it's a bit above my paygrade- I'm not confident that I'm entering correct parameters. I do have pilot data that I can pull various estimates from, as well as access to PASS power analysis software. Thanks in advance!

[Rich Goldstein]

since the two variables in your interaction are binary, you will find the power analysis easier if you view these as a 4-category variable to be estimated with 3 indicator (dummy) variables; I note that this will not answer the same question but that should not impact the power analysis too seriously (but note that if the interaction is your main term of interest, you should not do this)

I am not familar with PASS (it does not run on Mac's) but you can download the free "G*Power" program which has, under "z tests" a routine for Poisson regression

[Bruce Weaver]

Hello Jennifer Carr. At the risk of suggesting something else that may be "above your paygrade", the Stata Blog includes a series of posts on power analysis (or sample size estimation) via simulation. I don't know if there is one on Poisson regression specifically, but perhaps the 3rd in the series (on linear and logistic regression) will help you understand the basic approach. You can find it here:

• https://blog.stata.com/2019/08/13/ca...ic-regression/

I hope this helps.


PS- Before clicking Post, I added "Poisson regression" to my search terms and found this FAQ item:

• https://www.stata.com/support/faqs/s...by-simulation/

Maybe Chuck Huber (StataCorp) can point to other resources. (Hi Chuck.)

[Leonardo Guizzetti]

In times like this, I would opt to "roll my own". You might find the following paper as a useful starting point for conducting power analyses in Stata by hand. The trick with simulating data from regression models is selecting suitable and realistic choices for the regression coefficients.

Feiveson, A. H. (2002). Power by Simulation. The Stata Journal: Promoting Communications on Statistics and Stata, 2(2), 107-124. https://doi.org/10.1177/1536867X0200200201

[Jennifer Carr]

Thank you all so much! At least I'm confident that I'm not missing some super easy way to do this! I'll dig through the resources that you all sent along and will hopefully figure it out. I appreciate the help!