R/nice_lm.R

In rempsyc: Convenience Functions for Psychology

#' @title Nice formatting of lm models
#'
#' @description Formats output of [lm()] model object for a
#' publication-ready format.
#'
#' @details The effect size, sr2 (semi-partial correlation squared, also
#' known as delta R2), is computed through [effectsize::r2_semipartial].
#' Please read the documentation for that function, especially regarding
#' the interpretation of the confidence interval. In `rempsyc`, instead
#' of using the default one-sided alternative ("greater"), we use the
#' two-sided alternative.
#'
#' To interpret the sr2, use [effectsize::interpret_r2_semipartial()].
#'
#' For the *easystats* equivalent, use [report::report()] on the [lm()]
#' model object.
#'
#' @param model The model to be formatted.
#' @param b.label What to rename the default "b" column (e.g.,
#' to capital B if using standardized data for it to be converted
#' to the Greek beta symbol in the [nice_table] function). Now
#' attempts to automatically detect whether the variables were
#' standardized, and if so, sets `b.label = "B"` automatically.
#' Factor variables or dummy variables (only two numeric values)
#' are ignored when checking for standardization.
#' *This argument is now deprecated, please use argument
#' `standardize` directly instead.*
#' @param standardize Logical, whether to standardize the
#' data before refitting the model. If `TRUE`, automatically sets
#' `b.label = "B"`. Defaults to `FALSE`. Note that if you have factor
#' variables, these will be pseudo-betas, so these coefficients could
#' be interpreted more like Cohen's *d*.
#' @param mod.id Logical. Whether to display the model number,
#' when there is more than one model.
#' @param ci.alternative Alternative for the confidence interval
#' of the sr2. It can be either "two.sided (the default in this
#' package), "greater", or "less".
#' @param ... Further arguments to be passed to the
#' [effectsize::r2_semipartial] function for the effect size.
#'
#' @keywords moderation interaction regression
#' @return A formatted dataframe of the specified lm model, with DV, IV, degrees
#' of freedom, regression coefficient, t-value, p-value, and the effect
#' size, the semi-partial correlation squared, and its confidence interval.
#' @export
#' @examplesIf requireNamespace("effectsize", quietly = TRUE)
#' # Make and format model
#' model <- lm(mpg ~ cyl + wt * hp, mtcars)
#' nice_lm(model)
#'
#' # Make and format multiple models
#' model2 <- lm(qsec ~ disp + drat * carb, mtcars)
#' my.models <- list(model, model2)
#' x <- nice_lm(my.models)
#' x
#' @examplesIf requireNamespace("effectsize", quietly = TRUE) && packageVersion("effectsize") >= "0.8.3.5"
#' # Get interpretations
#' cbind(x, Interpretation = effectsize::interpret_r2_semipartial(x$sr2))
#'
#' @seealso
#' Checking simple slopes after testing for moderation:
#' \code{\link{nice_lm_slopes}}, \code{\link{nice_mod}},
#' \code{\link{nice_slopes}}. Tutorial:
#' \url{https://rempsyc.remi-theriault.com/articles/moderation}
#'

nice_lm <- function(model,
                    b.label = "b",
                    standardize = FALSE,
                    mod.id = TRUE,
                    ci.alternative = "two.sided",
                    ...) {
  rlang::check_installed("effectsize",
    version = get_dep_version("effectsize"),
    reason = "for this function."
  )
  if (inherits(model, "list") && all(unlist(lapply(model, inherits, "lm")))) {
    models.list <- model
  } else if (inherits(model, "lm")) {
    models.list <- list(model)
  } else {
    stop("Model must be of class 'lm' or be a 'list()' of lm models (using 'c()' won't work).")
  }

  if (!missing(b.label)) {
    message(paste(
      "The argument 'b.label' is deprecated.",
      "If your data is standardized, capital B will be used automatically.",
      "Else, please use argument 'standardize' directly instead."
    ))
  }

  if (model_is_standardized(models.list)) {
    b.label <- "B"
  } else if (isTRUE(standardize)) {
    data.list <- lapply(models.list, function(x) {
      scale(x$model)
    })
    models.list <- lapply(seq_along(models.list), function(i) {
      data <- as.data.frame(data.list[i])
      stats::update(models.list[[i]], data = data)
    })
    b.label <- "B"
  }

  sums.list <- lapply(models.list, function(x) {
    summary(x)$coefficients[-1, -2]
  })
  df.list <- lapply(models.list, function(x) x["df.residual"])
  ES.list <- lapply(models.list, function(x) {
    z <- effectsize::r2_semipartial(x, alternative = ci.alternative, ...)
    dplyr::select(z, "r2_semipartial", "CI_low", "CI_high")
  })
  
  if (as.data.frame(ES.list) %>% nrow() == 1) {
    sums.list <- lapply(sums.list, function(x) {t(as.matrix(x))})
  }
  
  stats.list <- mapply(cbind, df.list, sums.list, ES.list, SIMPLIFY = FALSE)
  stats.list <- lapply(stats.list, function(x) {
    x <- as.data.frame(x)
    IV <- row.names(x)
    x <- cbind(IV, x)
  })
  table.stats <- do.call(rbind.data.frame, stats.list)
  response.names <- unlist(lapply(models.list, function(x) {
    rep(as.character(x$terms[[2]]), each = length(x$assign) - 1)
  }))
  row.names(table.stats) <- NULL
  table.stats <- cbind(response.names, table.stats)
  good.names <- c(
    "Dependent Variable", "Predictor",
    "df", "b", "t", "p", "sr2", "CI_lower", "CI_upper"
  )
  if (length(models.list) > 1 & mod.id == TRUE) {
    model.number <- rep(seq_along(models.list), times = lapply(sums.list, nrow))
    table.stats <- cbind(model.number, table.stats)
    names(table.stats) <- c("Model Number", good.names)
  } else {
    names(table.stats) <- good.names
  }
  names(table.stats)[names(table.stats) == "b"] <- b.label

  table.stats
}