mixed model analysis_stop when no significant

Discuss statistics related things
Post Reply
Annie
Posts: 1
Joined: Tue Jun 22, 2021 9:52 am

mixed model analysis_stop when no significant

Post by Annie »

Hi all, we are running a mixed model analysis in Jamovi using the Gamlj module. We have several predictors (both covariates and factors) and are clustering the intercept under the participant (1|participant). Further, we have followed the patients for over a year, hence have several time points. Our question concerns whether it is possible to make jamovi stop as long as a polynomial effect is no longer significant,(which if I am not wrong was possible with HLM software). For example, say we have 4 timepoints. The possible polynomial effects are linear, quadratic, cubic and the fourth order effect. In case the linear and the quadratic effects are significant but the cubic and the 4th order effect are not, is it possible for jamovi to stop testing for significance once it reaches the quadratic effect?

Hope this makes sense.
Annie
User avatar
mcfanda@gmail.com
Posts: 457
Joined: Thu Mar 23, 2017 9:24 pm

Re: mixed model analysis_stop when no significant

Post by mcfanda@gmail.com »

Hi
I see your point but I'm afraid this is not going to be implemented. The reason is that when you estimate a set of contrasts, polynomial ones, or any other kind, the whole set should be included in the model no matter each individual contrast significance. If you do not do that, the model results will be biased. In jamovi, like in any other software that estimates full linear models, the number of contrasts should be K-1, where K is the number of levels. Once the model is estimated, testing for significance takes less than .0001 seconds on a normal computer, so I do not really see the advantage of stopping.

There's also another reason not to implement that. Polynomial contrasts are independent, so observing that some order (say linear or quadratic) is not significant does not say anything about the next order contrast significance. Thus, the only way to know whether a contrast is significant is to test it. Once tested, why not present it to the user?

As regards your example, with 4-time points you need 3 contrasts: linear, quadratic, and cubic. They exhaust the main effect of time. Higher-order contrasts will be redundant, but estimating less than 3 will bias the results. So, we estimate 3 contrasts.

Please. let me know if I got your question right
cheers
mc
Post Reply