Support Functions for Model Extensions"

In effectsize: Indices of Effect Size

library(knitr)
knitr::opts_chunk$set(
  comment = ">",
  warning = FALSE,
  message = FALSE
)
options(digits = 2)
options(knitr.kable.NA = "")

set.seed(333)

library(effectsize)

Supporting ANOVA Effect Sizes

To add support for you model, create a new .anova_es() method function. This functions should generally do 3 things:

1. Build a data frame with all the required information.
2. Pass the data frame to one of the 3 functions.
3. Set some attributes to the output.

Simple ANOVA tables

The input data frame must have these columns: - Parameter (char) - The name of the parameter or, more often, the term. - Sum_Squares (num) - The sum of squares. - df (num) - The degrees of freedom associated with the Sum_Squares. - Mean_Square_residuals (num; optional) - if not present, is calculated as Sum_Squares / df. (Any other column is ignored.)

And exactly 1 row Where Parameter is Residual.

Optionally, one of the rows can have a (Intercept) value for Parameter.

An example of a minimally valid data frame:

min_aov <- data.frame(
  Parameter = c("(Intercept)", "A", "B", "Residuals"),
  Sum_Squares = c(30, 40, 10, 100),
  df = c(1, 1, 2, 50)
)

Pass the data frame to .es_aov_simple():

.es_aov_simple(
  min_aov,
  type = "eta", partial = TRUE, generalized = FALSE,
  include_intercept = FALSE,
  ci = 0.95, alternative = "greater",
  verbose = TRUE
)

The output is a data frame with the columns: Parameter, the effect size, and (optionally) CI + CI_low + CI_high,

And with the following attributes: generalized, ci, alternative, anova_type (NA or NULL), approximate.

You can then set the anova_type attribute to {1, 2, 3, or NA} and return the output.

ANOVA Tables with Multiple Error Strata

(e.g., aovlist models.)

The input data frame must have these columns:

• Group (char) - The strata
• Parameter (char)
• Sum_Squares (num)
• df (num)
• Mean_Square_residuals (num; optional)

And exactly 1 row per Group Where Parameter is Residual.

Optionally, one of the rows can have a (Intercept) value for Parameter.

An example of a minimally valid data frame:

min_aovlist <- data.frame(
  Group = c("S", "S", "S:A", "S:A"),
  Parameter = c("(Intercept)", "Residuals", "A", "Residuals"),
  Sum_Squares = c(34, 21, 34, 400),
  df = c(1, 12, 4, 30)
)

Pass the data frame to .es_aov_strata(), along with a list of predictors (including the stratifying variables) to the DV_names argument:

.es_aov_strata(
  min_aovlist,
  DV_names = c("S", "A"),
  type = "omega", partial = TRUE, generalized = FALSE,
  ci = 0.95, alternative = "greater",
  verbose = TRUE,
  include_intercept = TRUE
)

The output is a data frame with the columns: Group, Parameter, the effect size, and (optionally) CI + CI_low + CI_high,

And with the following attributes: generalized, ci, alternative, approximate.

You can then set the anova_type attribute to {1, 2, 3, or NA} and return the output.

Approximate Effect sizes

When sums of squares cannot be extracted, we can still get approximate effect sizes based on the F_to_eta2() family of functions.

The input data frame must have these columns:

• Parameter (char)
• F (num) - The F test statistic.
• df (num) - effect degrees of freedom.
• (Can also have a t col instead, in which case df is set to 1, and F is t^2).
• df_error (num) - error degrees of freedom.

Optionally, one of the rows can have (Intercept) as the Parameter.

An example of a minimally valid data frame:

min_anova <- data.frame(
  Parameter = c("(Intercept)", "A", "B"),
  F = c(4, 7, 0.7),
  df = c(1, 1, 2),
  df_error = 34
)

Pass the table to .es_aov_table():

.es_aov_table(
  min_anova,
  type = "eta", partial = TRUE, generalized = FALSE,
  include_intercept = FALSE,
  ci = 0.95, alternative = "greater",
  verbose = TRUE
)

The output is a data frame with the columns: Parameter, the effect size, and (optionally) CI + CI_low + CI_high,

And with the following attributes: generalized, ci, alternative, approximate.

You can then set the anova_type attribute to {1, 2, 3, or NA} and return the output, and optionally the approximate attribute, and return the output.

Example

Let's fit a simple linear model and change its class:

mod <- lm(mpg ~ factor(cyl) + am, mtcars)

class(mod) <- "superMODEL"

We now need a new .anova_es.superMODEL function:

.anova_es.superMODEL <- function(model, ...) {
  # Get ANOVA table
  anov <- suppressWarnings(stats:::anova.lm(model))
  anov <- as.data.frame(anov)

  # Clean up
  anov[["Parameter"]] <- rownames(anov)
  colnames(anov)[2:1] <- c("Sum_Squares", "df")

  # Pass
  out <- .es_aov_simple(anov, ...)

  # Set attribute
  attr(out, "anova_type") <- 1

  out
}

# This is for: https://github.com/easystats/easystats/issues/348
.anova_es.superMODEL <<- .anova_es.superMODEL

And... that's it! Our new superMODEL class of models is fully supported!

eta_squared(mod)

eta_squared(mod, partial = FALSE)

omega_squared(mod)

# Etc...

References

effectsize documentation built on Sept. 14, 2023, 5:07 p.m.