Where the variation lives: a variance partition for a mixed model

variance_partition() turns a fitted mixed model's variance components into a one-row-per-source table: each random-effect variance plus, when it is defined, the residual (within-group) variance. The pct column is each source's share of the total, answering a biologist's first question of a mixed model – where does the variation live?

The residual row (and therefore pct) appears only when the residual variance is defined on a single scale: a homoscedastic Gaussian fit (one residual SD) or a binomial fit (the known latent-scale residual, \(\pi^2/3\) for logit, 1 for probit). For other families, or a location-scale Gaussian whose residual SD varies across observations, the random-effect variances are still shown but pct is NA with a reason – shares of a total are not meaningful without a residual.

Point estimates only; no confidence intervals (the package's uncertainty contract).

Usage

variance_partition(x, ...)

Arguments

x

A symbolized_model whose underlying fit is retained on x$metadata$fit.

...

Reserved for future use.

Value

A tibble (S3 class symbolizer_variance_partition) with one row per variance source. Columns: component, variance, sd, pct. Carries a reason attribute (character or NA) explaining an all-NA pct.

See also

icc() for the single-number intraclass correlation.

Examples

# variance_partition(symbolize(lme4::lmer(Reaction ~ Days + (1 | Subject),
#                                         data = lme4::sleepstudy)))