Simplify random effects

If a maximal model will not converge, even after you change the optimiser (see Troubleshoot), then you will need to simplify the model’s random effect structure.

Let’s assume a maximal model with the formula y ~ x * z + (1 + x * z | g). This model is predicting y as a function of x, z, and their interaction, and it includes by-g adjustments to the intercept, to the slopes of x and z, and to the interaction. It also estimates the correlations between all the by-g adjustments.

There is no single correct way to simplify the random effect structure. But here’s a guide to a sequence of simplifications you might make. This sequence starts with the parameters that are generally least relevant for your research question (but if you really need a given parameter to address your research question, then definitely leave it in!).

Try to fit the model again after every step, and stop simplifying once it successfully converges without singularities.

• Remove the by-group adjustment to the interaction term: y ~ x * z + (1 + x + z | g).
• Remove the correlations between by-group adjustments: y ~ x * z + (1 + x + z || g).
• Remove one of the by-group slope adjustments (if one is a covariate, remove that one first): y ~ x * z + (1 + x || g).
• Remove both of the by-group slope adjustments: y ~ x * z + (1 | g).
• (Usually not a good idea, only do this if you have very low group-level variance) Remove the random effect entirely: y ~ x * z.