Compute pseudo R-square for mixed models

This function computes the pseudo R-square for mixed models by quantifying the reduction on the variance-covariance between a full model and a reduced model.

Usage

r2_pseudo(mf, mr, mnull = mr)

V_table(model)

Arguments

mf: The full model
mr: The reduced model
mnull: The empty model, where some variance of interest is unmodeled (an "unconditional" model)
model: A fitted model object

Value

A tibble with the following columns:

grp: The name of the grouping factor
var: The name of the variance component
r2: The pseudo R-square for that variance component

Details

Variance components are extracted from the 3 models, and where they match (groups and terms must match exactly), we compute:

$$R^2 = \frac{V_{full} - V_{restricted}}{V_{null}}$$

This gives the increase in variance explained by the full model compared to the reduced model. When the restricted model is the empty model, this is the same as:

$$R^2 = 1 - \frac{V_{full}}{V_{restricted}}$$

Supported model classes

{stats}: lm, glm
{lme4}: lmerMod and glmerMod
{rstanarm}: stanreg
{brms}: brmsfit
{glmmTMB}: glmmTMB

Examples


mod0 <- lme4::lmer(Reaction ~ 1 + (1 | Subject), data = lme4::sleepstudy)
mod1 <- lme4::lmer(Reaction ~ Days + (1 | Subject), data = lme4::sleepstudy)
mod2 <- lme4::lmer(Reaction ~ Days + (Days | Subject), data = lme4::sleepstudy)

r2_pseudo(mod1, mod0) # Residual variance explained by the fixed effect
#> # A tibble: 2 × 3
#>   grp      var              r2
#>   <chr>    <chr>         <dbl>
#> 1 Subject  (Intercept) -0.0781
#> 2 Residual NA           0.510 
r2_pseudo(mod2, mod1, mod0) # Residual variance explained by the random effects
#> # A tibble: 2 × 3
#>   grp      var            r2
#>   <chr>    <chr>       <dbl>
#> 1 Subject  (Intercept) 0.599
#> 2 Residual NA          0.156
r2_pseudo(mod2, mod0) # Residual variance explained by the mixed effects
#> # A tibble: 2 × 3
#>   grp      var            r2
#>   <chr>    <chr>       <dbl>
#> 1 Subject  (Intercept) 0.521
#> 2 Residual NA          0.666