Skip to contents

Check the assumptions for an ANOVA with a variable number of factors: Normality and Homogeneity of variance assumption.

Source: R/utils.R, R/zzz-aliases.R
checkAssumptionsForAnova.Rd

Checks the normality and variance assumptions of a factorial ANOVA and recommends the parametric or the non-parametric analysis. Every factor is treated as categorical – numeric condition codes (1, 2, 3) included, which stats::lm() would otherwise fit as a single linear covariate – and the message names the tests that were run and their results.

Usage

checkAssumptionsForAnova(data, y, factors, subject = NULL)

check_assumptions_anova(data, y, factors, subject = NULL)

Arguments

data

the data frame

y

The dependent variable for which assumptions should be checked

factors

A character vector of factor names

subject

the participant-ID column for a design with within-subject factors, as a string; NULL (default) for a between-subjects design.

Value

The guidance text (also emitted as a message), invisibly, with attributes:

parametric

TRUE if every check passed.

design

"between", "within" or "mixed".

method

named character vector naming the tests run (residuals, groupwise, homogeneity).

residuals

the Shapiro-Wilk result on the model residuals (columns as in check_normality_by_group()'s tests).

groupwise

between subjects: the per-cell Shapiro-Wilk results with Holm-adjusted p-values; NULL otherwise.

homogeneity

the Brown-Forsythe result(s) (df1, df2, statistic, p, p_adjusted), or NULL when no between-subject factor exists.

Details

  • Between subjects (subject = NULL, the default): a Shapiro-Wilk test on the residuals of the linear model y ~ A * B * ...; a Shapiro-Wilk test within every cell of the design, Holm-corrected across the cells (as in check_normality_by_group()); and the Brown-Forsythe test (median-centred Levene's test) across the cells. This assumes one row per participant: a within-subject factor analysed this way is tested on residuals that still contain each participant's overall level, so pass subject for repeated measures.

  • Within subjects or mixed (subject given): a Shapiro-Wilk test on the residuals of y ~ A * B * ... + subject, with the participant as a fixed factor – the residuals after removing each participant's overall level, which is what a repeated-measures ANOVA assumes to be normal. The raw scores per cell are not tested, as they also carry the between-participant spread. Variance homogeneity is checked only for between-subject factors (those constant within a participant), with the Brown-Forsythe test across their groups at each combination of the within-subject factors (Holm-corrected). For within-subject factors the corresponding assumption is sphericity, which rstatix::anova_test() tests (Mauchly) and corrects (Greenhouse-Geisser) itself. More than one row per participant and cell is an error: aggregate repeated trials first.

A residual set or cell that cannot be tested – fewer than three values, or no variance – counts as not normal, so the advice errs towards the non-parametric analysis. Rows with a missing value in y, factors or subject are left out.

Naming

check_assumptions_anova() is the spelling used throughout the documentation and the one to prefer in new code: the report_* / plot_* / check_* prefixes make the API discoverable through autocomplete.

checkAssumptionsForAnova() [Superseded] is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

# \donttest{
if (requireNamespace("rstatix", quietly = TRUE)) {
  set.seed(123)

  main_df <- data.frame(
    tlx_mental      = rnorm(40),
    Video           = factor(rep(c("A", "B"), each = 20)),
    DriverPosition  = factor(rep(c("Left", "Right"), times = 20))
  )

  checkAssumptionsForAnova(
    data    = main_df,
    y       = "tlx_mental",
    factors = c("Video", "DriverPosition")
  )

  # The same two factors measured within each of 10 participants
  within_df <- expand.grid(
    id             = 1:10,
    Video          = c("A", "B"),
    DriverPosition = c("Left", "Right")
  )
  within_df$tlx_mental <- rnorm(10)[within_df$id] + rnorm(40)
  checkAssumptionsForAnova(
    data    = within_df,
    y       = "tlx_mental",
    factors = c("Video", "DriverPosition"),
    subject = "id"
  )
}
#> You may take parametric ANOVA (function anova_test). Checks: Shapiro-Wilk on the residuals of tlx_mental ~ Video * DriverPosition: W = 0.988, p = 0.949; Shapiro-Wilk test per cell (Holm-adjusted): no cell deviates; Brown-Forsythe test (median-centred Levene's test): F(3, 36) = 0.54, p = 0.661. See https://www.datanovia.com/learn/biostatistics/anova/anova-in-r#check-assumptions-1 for more information.
#> You may take parametric ANOVA (function anova_test). Checks: Shapiro-Wilk on the residuals of tlx_mental ~ Video * DriverPosition + id: W = 0.988, p = 0.941; sphericity of the within-subject factor(s) Video, DriverPosition is tested (Mauchly) and corrected (Greenhouse-Geisser) by rstatix::anova_test(). See https://www.datanovia.com/learn/biostatistics/anova/anova-in-r#check-assumptions-1 for more information.
# }