LME-infinite dfs for post-hoc and simple effects

General help and assistance with jamovi. Bug reports can be made at our issues page: https://github.com/jamovi/jamovi/issues . (If you're unsure feel free to discuss it here)

by tutkuoztel » Mon Apr 22, 2019 2:17 pm

hello!

I am running a mixed effects analysis on Jamovi using GAMLj. I have2 continuous covariates and 1 categorical factor (with 4 levels), with all main effects and 2-and 3-way interaction terms as fixed effects (on slope) and subjects as random effects (on intercept) in my model. I have a significant interaction between my categorical factor and one of my continuous covariates on my DV. I want to get to know in which levels of my factor the effect of my covariate on my DV changes, with simple effects. I put my categorical factor as moderator and my covariate as simple effects variable. However, both simple effects and post hoc return infinite num and den dfs, and thus do not compute the t-stats (and F-stats). my number of observations is 6069.

Why do both post-hoc and simple effects return infitine dfs? how can I overcome this problem?

thanks so much in advance,

tutku
tutkuoztel
 
Posts: 3
Joined: Mon Apr 22, 2019 11:57 am

by mcfanda@gmail.com » Tue Apr 23, 2019 1:53 pm

Hi, would you share the .omv file of the analyses that produce the issue ?
mcfanda@gmail.com
 
Posts: 143
Joined: Thu Mar 23, 2017 9:24 pm

by mcfanda@gmail.com » Tue Apr 23, 2019 4:42 pm

HI, after reading more carefully your message I realized that you have 6069 observations. This is the problem. GAMLjMixed uses R package `emmeans`, which has an internal option to turn off the computation of degrees of freedom for models with more than 3000 observations. This is the warning one gets in R.
```
Note: D.f. calculations have been disabled because the number of observations exceeds 3000.
To enable adjustments, set emm_options(pbkrtest.limit = 24000) or larger,
but be warned that this may result in large computation time and memory use.
```
The rationale for this option is that when the observations number is larger than 3000, a t-test is equivalent to a z-test, which does not require computing the DF.
I see now that GAMLj does not return a warning explaining what is going on, which will be very useful for the user. Next version (1.5.0 is coming out in a few days) will sort this problem out. In the meantime, you can easily compute the z-test associated with the simple slopes by taking the ratio estimate/SE in the `Simple effects Parameter estimates`.
mcfanda@gmail.com
 
Posts: 143
Joined: Thu Mar 23, 2017 9:24 pm

by mcfanda@gmail.com » Tue Apr 23, 2019 5:51 pm

for the post-hoc, you simple ignore the df knowing that the tests are z-tests. Adjusting is done anyway
mcfanda@gmail.com
 
Posts: 143
Joined: Thu Mar 23, 2017 9:24 pm


Return to Help