1

Hi I wanted to create a 'mega' linear mixed effects model with nested fixed effects (using the lme4 package) where the interaction effect of type and language was modelled within each level of brain region or Region-of-interest (ROI), type*language nested within ROI. The model did not converge, but there are many ROIs. Hence the number of observations ended up being very large (>4000), and the number of parameters are very large (>100).

m <- lmer(DV ~ ROI/(word*language) + cov1 + cov2 + cov3 + cov4 + cov1:word + cov1:language + cov2:word + cov2:language + cov3:language + cov4:language + (1|subject), data=df)

I was wondering if the number of parameters (>100) would actually be a problem.

There were many interaction effects included between covariates and the factors of interest, hence I used AICc to validate this model. I wanted to also check whether my understanding of AICc was valid. I used AICc because it was recommended for 'small sample sizes' where n/k < 40, n being the number of observations, k being the number of parameters.

Serena
  • 11
  • 1
  • "I was wondering if the number of parameters (>100) would actually be a problem." Yes. Have you checked for multi-collinearity? I would start smaller and first model a single ROI. Possibly `word` could be random effect instead of a fixed effect? – Roland Mar 29 '23 at 06:17
  • I checked for multi-collinearity, using just the main effects of each factor (Pair, word, language, covariates) in the model. It appears that there is no multi-collinearity problem. Would that still posit a problem? I checked for mutli-collinearity in the fuller model using check_collinearity(m) using the performance pacakge but the warning message advised me to just check the main effects. – Serena Mar 29 '23 at 06:33
  • Estimating 100 parameters with default starting values is still a tall order. If you insist on this huge model, you should probably provide custom starting values or fit a Baesian model with pretty strong priors. I'm assuming you have checked the design matrix for balance and structural issues. – Roland Mar 29 '23 at 06:47

0 Answers0