In easystats/performance: Assessment of Regression Models Performance

library(knitr)
knitr::opts_chunk$set(
  dpi = 300,
  fig.width = 7,
  fig.height = 5,
  out.width = "100%",
  collapse = TRUE,
  comment = "#>",
  warning = FALSE,
  message = FALSE
)
options(knitr.kable.NA = "")
options(digits = 2)

pkgs <- c("DHARMa", "glmmTMB", "see")
successfully_loaded <- vapply(pkgs, requireNamespace, FUN.VALUE = logical(1L), quietly = TRUE)
can_evaluate <- all(successfully_loaded)

if (can_evaluate) {
  knitr::opts_chunk$set(eval = TRUE)
  vapply(pkgs, require, FUN.VALUE = logical(1L), quietly = TRUE, character.only = TRUE)
} else {
  knitr::opts_chunk$set(eval = FALSE)
}

The basic workflow for simulated residual checks using simulate_residuals() is as follows.

First, fit a model:

model <- glmmTMB::glmmTMB(
  count ~ mined + spp + (1 | site),
  family = poisson,
  data = glmmTMB::Salamanders
)

Next, simulate residuals from the model:

library(performance)
simulated_residuals <- simulate_residuals(model)

simulated_residuals

The raw residuals can be extracted using residuals():

head(residuals(simulated_residuals))

Note that since this inherits the DHARMa class, all the methods implemented in DHARMa just work, including all the tests:

DHARMa::testUniformity(simulated_residuals, plot = FALSE)

Finally, run specific checks on the simulated residuals:

check_residuals(simulated_residuals)

Further implemented checks are tests for overdispersion, outliers and zero-inflation.

check_overdispersion(simulated_residuals)

check_zeroinflation(simulated_residuals)

check_outliers(simulated_residuals)

The above three functions internally call simulate_residuals() for more complex models automatically, so you don't need to call simulate_residuals() yourself. Simulated residuals are usually more reliable than the standard residuals, especially for complex models.

Finally, you can even perform a visual check for the entire model, either by passing the model object directly, or the object returned from simulate_residuals().

check_model(simulated_residuals, size_dot = 1.5)

The check_model() function is the main reason we don't want to prematurely extract the residuals in simulate_residuals(), because if we do then the simulated residual won't contain the model fit (fittedModel in the output below), so we won't be able to do all of the checks we would want to do using the model (e.g., posterior predictive checks).