performance Files

Breaking Changes

• binned_residuals() now defaults to using response residuals, in line with the suggestion from Gelman and Hill (2007).

Changes

• New functions performance_mor(), performance_poor() and performance_ior() to calculate median odds ratios, proportions of opposed odds ratios, and interval odds ratios for logistic multilevel regression.

• check_group_variation() now returns a numeric effect size of the grouping variable's predictive association strength.

• Updated test-files to fix deprecated names.

• r2() now also covers remaining families from package glmmTMB and defaults to returning McFadden's R2. Documentation of r2() was updated accordingly.

Bug fixes

• compare_performance(rank = TRUE) now ranks Log_loss in the right direction. It was normalized along with the other indices but never flipped, so the model with the largest log-loss got the highest performance score and the best-fitting model was listed last (#917).

• r2_mcfadden() now uses the number of estimated parameters in the penalty of the adjusted R2. Previously the component list returned by insight::find_parameters() was counted instead of the parameters inside it, so the penalty was almost always 1 and the adjusted value barely differed from the unadjusted one (#933).

• check_collinearity() now properly warns when the vcov matrix is rank deficient (#922).

• r2() (and hence model_performance()) no longer errors for glmmTMB negative-binomial models (nbinom1/nbinom2) without random effects, and now returns McFadden's R2 for them.

• check_model() now uses Pearson residuals for the homogeneity-of-variance plot of glmmTMB and MixMod models. Previously these residuals were divided by a single scalar, which is only correct when the variance function does not depend on the mean; for non-mixed binomial and Poisson models that scalar was 1, so the plot could suggest heteroscedasticity for correctly specified models (#926).

• Fixed issue in compare_performance() for objects from package lavaan.