What is the best tool to calculate the sample size for a study in which I want to use ordinal logistic regression? My dependent variable is ordinal with three levels. I cannot use G*Power for this purpose, as it does not support this a priori method. Thank you!
Hello Fani Minti. I suspect you would need to use simulation for that. I found the following R code, for example:
https://gist.github.com/BioSciEconomist/3d1b61b202b1e164574211a20ec15951
But another idea occurs to me too. I wonder if you could estimate the sample size for two separate binary logistic regression models, and pick the larger of the two sample size estimates. For the first model, let your binary DV = 1 vs [2,3]; and for the second model, let your binary DV = [1,2] vs 3 (where 1, 2, and 3 are the levels of your original 3-level DV).
FYI, this is just an off-the-top-of-my-head (OTTOMH), untested idea. I have no idea whether anyone else has tried this previously. But I am curious to know how it would compare to simulation.
Good luck!
Bruce Weaver thank you for your answer!
I found several libraries in R, but none specifically for ordinal logistic regression. It seems that Monte Carlo simulation might be the solution, although I'm not sure what it means to generate simulated data.
I finally had the same idea!! I used G*Power with the approach you have already suggested, by dividing the outcome into two groups: [1,2] versus 3.
I believe a post hoc analysis, once I have the data, will help me determine whether I have achieved the necessary statistical power.
Hello Fani Minti. Two things:
1) Notice that my idea was to get sample size estimates for two binary logit models, one with 1 v [2,3], and another for [1,2] v 3.
2) If the post hoc analysis you are describing is conventional post-hoc power analysis, beware! See this note by Russ Lenth, for example: https://homepage.stat.uiowa.edu/~rlenth/Power/2badHabits.pdf
Cheers,
Bruce
- Badges
- Science topic
- Similar topics
- Publishing