Hi all,
When you perform a 2-outcome logistic regression with say, two categorical variables (with categories A/B/C and 1/2, say), what you get is the significance of the effects highlighted in the picture below (with reference levels B and 1 for convenience):
If I'm interested in the significance of the OR comparing the odds in category C1 with the odds in category C2 (the rightmost arrow) I can change the reference levels but this would not account for multiple testing.
Same problem if I also want to know the significance of the (undepicted) interaction arrow between the leftmost and rightmost vertical arrows.
Getting the significance of all arrows at once would probably have to be called post-hoc tests. Is there a way to get them in jamovi that I'm not aware of?
UPDATE:
I thought I was going to answer my own question, as I realized that Gamlj has a post-hoc option, so I ran a generalized linear model, and I cannot figure out why the results differ from the output of a logistic regression. Only the interaction terms agree, which usually indicates that there are covariates that aren't scaled the same in both models, yet here all I have is factors and simple contrasts, so this shouldn't be a possible issue.
Thanks for any ideas!
Arnaud
Post-hoc in logistic regression
- mcfanda@gmail.com
- Posts: 460
- Joined: Thu Mar 23, 2017 9:24 pm
Re: Post-hoc in logistic regression
Hi
Logistic regression uses dummies variables (0 vs 1), whereas GAMLj default uses "simple" which are dummies centered to zero. With no interaction, results are identical, with the interaction in the model, they are different. If you set GAMLj factor coding to "dummy", you get identical results.
Keep in mind that when you have an interaction, the "usual" main effects are obtained using centered conding
Logistic regression uses dummies variables (0 vs 1), whereas GAMLj default uses "simple" which are dummies centered to zero. With no interaction, results are identical, with the interaction in the model, they are different. If you set GAMLj factor coding to "dummy", you get identical results.
Keep in mind that when you have an interaction, the "usual" main effects are obtained using centered conding
- Arnaud Mortier
- Posts: 32
- Joined: Tue Apr 27, 2021 11:39 am
Re: Post-hoc in logistic regression
Oh I see! I thought that simple coding was being used in both cases.
Thanks a lot!
Thanks a lot!