performance 0.11.0

New supported models

• Rudimentary support for models of class serp from package serp.

New functions

• simulate_residuals() and check_residuals(), to simulate and check residuals
  from generalized linear (mixed) models. Simulating residuals is based on the
  DHARMa package, and objects returned by simulate_residuals() inherit from
  the DHARMa class, and thus can be used with any functions from the DHARMa
  package. However, there are also implementations in the performance package,
  such as check_overdispersion(), check_zeroinflation(), check_outliers()
  or check_model().

• Plots for check_model() have been improved. The Q-Q plots are now based
  on simulated residuals from the DHARMa package for non-Gaussian models, thus
  providing more accurate and informative plots. The half-normal QQ plot for
  generalized linear models can still be obtained by setting the new argument
  residual_type = "normal".

• Following functions now support simulated residuals (from simulate_residuals())
  resp. objects returned from DHARMa::simulateResiduals():

  • check_overdispersion()
  • check_zeroinflation()
  • check_outliers()
  • check_model()

General

• Improved error messages for check_model() when QQ-plots cannot be created.

• check_distribution() is more stable for possibly sparse data.

Bug fixes

• Fixed issue in check_normality() for t-tests.

• Fixed issue in check_itemscale() for data frame inputs, when factor_index
  was not a named vector.

performance 0.10.9

Changes

• r2() for models of class glmmTMB without random effects now returns the
  correct r-squared value for non-mixed models.

• check_itemscale() now also accepts data frames as input. In this case,
  factor_index must be specified, which must be a numeric vector of same
  length as number of columns in x, where each element is the index of the
  factor to which the respective column in x.

• check_itemscale() gets a print_html() method.

• Clarification in the documentation of the estimator argument for
  performance_aic().

• Improved plots for overdispersion-checks for negative-binomial models from
  package glmmTMB (affects check_overdispersion() and check_mnodel()).

• Improved detection rates for singularity in check_singularity() for models
  from package glmmTMB.

• For model of class glmmTMB, deviance residuals are now used in the
  check_model() plot.

• Improved (better to understand) error messages for check_model(),
  check_collinearity() and check_outliers() for models with non-numeric
  response variables.

• r2_kullback() now gives an informative error for non-supported models.

Bug fixes

• Fixed issue in binned_residuals() for models with binary outcome, where
  in rare occasions empty bins could occur.

• performance_score() should no longer fail for models where scoring rules
  can't be calculated. Instead, an informative message is returned.

• check_outliers() now properly accept the percentage_central argument when
  using the "mcd" method.

• Fixed edge cases in check_collinearity() and check_outliers() for models
  with response variables of classes Date, POSIXct, POSIXlt or difftime.

• Fixed issue with check_model() for models of package quantreg.

performance 0.10.8

Changes

• Changed behaviour of check_predictions() for models from binomial family,
  to get comparable plots for different ways of outcome specification. Now,
  if the outcome is a proportion, or defined as matrix of trials and successes,
  the produced plots are the same (because the models should be the same, too).

Bug fixes

• Fixed CRAN check errors.

• Fixed issue with binned_residuals() for models with binomial family, where
  the outcome was a proportion.

performance 0.10.7

Breaking changes

• binned_residuals() gains a few new arguments to control the residuals used
  for the test, as well as different options to calculate confidence intervals
  (namely, ci_type, residuals, ci and iterations). The default values
  to compute binned residuals have changed. Default residuals are now "deviance"
  residuals (and no longer "response" residuals). Default confidence intervals
  are now "exact" intervals (and no longer based on Gaussian approximation).
  Use ci_type = "gaussian" and residuals = "response" to get the old defaults.

Changes to functions

• binned_residuals() - like check_model() - gains a show_dots argument to
  show or hide data points that lie inside error bounds. This is particular
  useful for models with many observations, where generating the plot would be
  very slow.

performance 0.10.6

General

• Support for nestedLogit models.

Changes to functions

• check_outliers() for method "ics" now detects number of available cores
  for parallel computing via the "mc.cores" option. This is more robust than
  the previous method, which used parallel::detectCores(). Now you should
  set the number of cores via options(mc.cores = 4).

Bug fixes

• Fixed issues is check_model() for models that used data sets with
  variables of class "haven_labelled".

performance 0.10.5

Changes to functions

• More informative message for test_*() functions that "nesting" only refers
  to fixed effects parameters and currently ignores random effects when detecting
  nested models.

• check_outliers() for "ICS" method is now more stable and less likely to
  fail.

• check_convergence() now works for parsnip _glm models.

Bug fixes

• check_collinearity() did not work for hurdle- or zero-inflated models of
  package pscl when model had no explicitly defined formula for the
  zero-inflation model.

performance 0.10.4

• icc() and r2_nakagawa() gain a ci_method argument, to either calculate
  confidence intervals using boot::boot() (instead of lmer::bootMer()) when
  ci_method = "boot" or analytical confidence intervals
  (ci_method = "analytical"). Use ci_method = "boot" when the default method
  fails to compute confidence intervals and use ci_method = "analytical" if
  bootstrapped intervals cannot be calculated at all. Note that the default
  computation method is preferred.

• check_predictions() accepts a bandwidth argument (smoothing bandwidth),
  which is passed down to the plot() methods density-estimation.

• check_predictions() gains a type argument, which is passed down to the
  plot() method to change plot-type (density or discrete dots/intervals).
  By default, type is set to "default" for models without discrete outcomes,
  and else type = "discrete_interval".

• performance_accuracy() now includes confidence intervals, and reports those
  by default (the standard error is no longer reported, but still included).

Bug fixes

• Fixed issue in check_collinearity() for fixest models that used i()
  to create interactions in formulas.

performance 0.10.3

New functions

• item_discrimination(), to calculate the discrimination of a scale's items.

Support for new models

• model_performance(), check_overdispersion(), check_outliers() and r2()
  now work with objects of class fixest_multi (@etiennebacher, #554).

• model_performance() can now return the "Weak instruments" statistic and
  p-value for models of class ivreg with metrics = "weak_instruments"
  (@etiennebacher, #560).

• Support for mclogit models.

Changes to functions

• test_*() functions now automatically fit a null-model when only one model
  objects was provided for testing multiple models.

• Warnings in model_performance() for unsupported objects of class
  BFBayesFactor can now be suppressed with verbose = FALSE.

• check_predictions() no longer fails with issues when re_formula = NULL
  for mixed models, but instead gives a warning and tries to compute posterior
  predictive checks with re_formuka = NA.

• check_outliers() now also works for meta-analysis models from packages
  metafor and meta.

• plot() for performance::check_model() no longer produces a normal QQ plot
  for GLMs. Instead, it now shows a half-normal QQ plot of the absolute value
  of the standardized deviance residuals.

Bug fixes

• Fixed issue in print() method for check_collinearity(), which could mix
  up the correct order of parameters.

performance 0.10.2

General

• Revised usage of insight::get_data() to meet forthcoming changes in the
  insight package.

Changes to functions

• check_collinearity() now accepts NULL for the ci argument.

Bug fixes

• Fixed issue in item_difficulty() with detecting the maximum values of an
  item set. Furthermore, item_difficulty() gets a maximum_value argument
  in case no item contains the maximum value due to missings.

performance 0.10.1

General

• Minor improvements to the documentation.

Changes to functions

• icc() and r2_nakagawa() get ci and iterations arguments, to compute
  confidence intervals for the ICC resp. R2, based on bootstrapped sampling.

• r2() gets ci, to compute (analytical) confidence intervals for the R2.

• check_predictions() accepts a bw argument (smoothing bandwidth), which is
  passed down to the plot() methods density-estimation. The default for the
  smoothing bandwidth bw has changed from "nrd0" to "nrd", which seems
  to produce better fitting plots for non-gaussian models.

• The model underlying check_distribution() was now also trained to detect
  cauchy, half-cauchy and inverse-gamma distributions.

• model_performance() now allows to include the ICC for Bayesian models.

Bug fixes

• verbose didn't work for r2_bayes() with BFBayesFactor objects.

• Fixed issues in check_model() for models with convergence issues that lead
  to NA values in residuals.

• Fixed bug in check_outliers whereby passing multiple elements to the
  threshold list generated an error (#496).

• test_wald() now warns the user about inappropriate F test and calls
  test_likelihoodratio() for binomial models.