As far as I know effect size and sample size calculations in G Power are not available for a multinomial logistic regression. Is there another way to calculate effect size for this model?
Dear Nicola,
It is honestly not entirely clear what would be meant by effect size in this context. Although you can extend the pseudo-r2 idea, the idea of multinomial distributions is somewhat at variance with the idea that variance is a good yardstick.
Accordingly on the result side a classification table is considered a good idea. However, the implication of your question is that you are trying to do a power analysis up front. I would suggest that the best method would be simulation to estimate power. The second best would be to consider that the multinomial model will be generally more powerful than the linear association from a chi-squared table. Then the power would be reflecting a single df contrast on a chi-squared distribution. Not entirely trustworthy but then most power analysis is speculative at best. The weakest argument I can think of, but easiest to implement, would be to define the level of successful classification expected across the entire model as the effect size in a 2 by 2 chi-squared.
.
there's a (long) "way" described here
http://www2.sas.com/proceedings/sugi27/p260-27.pdf
but i do not know if it has an implementation available somewhere
.
Dear Nicola,
It is honestly not entirely clear what would be meant by effect size in this context. Although you can extend the pseudo-r2 idea, the idea of multinomial distributions is somewhat at variance with the idea that variance is a good yardstick.
Accordingly on the result side a classification table is considered a good idea. However, the implication of your question is that you are trying to do a power analysis up front. I would suggest that the best method would be simulation to estimate power. The second best would be to consider that the multinomial model will be generally more powerful than the linear association from a chi-squared table. Then the power would be reflecting a single df contrast on a chi-squared distribution. Not entirely trustworthy but then most power analysis is speculative at best. The weakest argument I can think of, but easiest to implement, would be to define the level of successful classification expected across the entire model as the effect size in a 2 by 2 chi-squared.
Hello Nicole,
Hopefully you would have found the answer to power computation for multi-nomial regression. Would it be possible to share your research?
Thanks
Kamalpreet
- Badges
- Science topic