Problems with assumption checks for LMM
Hey, quick rundown of our muscle experiment: We tracked 2 persons across 5 measurement timepoints each. At every timepoint, we analyzed Type I and Type II muscle fibers (50+ fibers per type per timepoint), creating a hierarchy.
Data structure: Cross-sectional area (µm²) of each individual fiber serves as our dependent variable. Fixed effects include Person, Time, Fiber Type (I vs. II), and their interaction. Each fiber gets its own unique ID as random effects grouping factor (1 | fiberID) .
The LMM assumptions headache: Different stats sources list varying prerequisites—residual normality, homoscedasticity, linearity, no multicollinearity, normal random effects, etc.
Here's the catch: JASP offers zero direct diagnostics for LMM assumptions. But some of the assumptions need to be tested on the LMM like the residals.
What is the right way to test these in JASP? In the linear Regressions modell and ignoring the hierarchy?
Is there a good sources which states clear, that when using bootstrapping all assumptions can be ignored?
Thanks for helping!
Comments
Hi Lena,
Yes, I completely agree. The current Mixed Models module is currently quite limited in this respect and it's long overdue we update it. The good news is that there is a new version of the lme4 package coming to CRAN soon and I am planning to overhaul the module significantly, including incorporating proper assumption checks. With respect to bootstrapping the results - it might alleviate some assumptions but there are still many left that cannot be addressed by bootstrap.
Cheers,
Frantisek