summarize_balances: Summarize group balances

In groupdata2: Creating Groups from Data

Summarize group balances

Description

\Sexpr[results=rd, stage=render]{lifecycle::badge("experimental")}

Summarize the balances of numeric, categorical, and ID columns in and between groups in one or more group columns.

This tool allows you to quickly and thoroughly assess the balance of different columns between groups. This is for instance useful after creating groups with fold(), partition(), or collapse_groups() to check how well they did and to compare multiple groupings.

The output contains:

1. `Groups`: a summary per group (per grouping column).

2. `Summary`: statistical descriptors of the group summaries.

3. `Normalized Summary`: statistical descriptors of a set of "normalized" group summaries. (Disabled by default)

When comparing how balanced the grouping columns are, we can use the standard deviations of the group summary columns. The lower a standard deviation is, the more similar the groups are in that column. To quickly extract these standard deviations, ordered by an aggregated rank, use ranked_balances() on the "Summary" data.frame in the output.

Usage

summarize_balances(
  data,
  group_cols,
  cat_cols = NULL,
  num_cols = NULL,
  id_cols = NULL,
  summarize_size = TRUE,
  include_normalized = FALSE,
  rank_weights = NULL,
  cat_levels_rank_weights = NULL,
  num_normalize_fn = function(x) {
     rearrr::min_max_scale(x, old_min = quantile(x,
    0.025), old_max = quantile(x, 0.975), new_min = 0, new_max = 1)
 }
)

Arguments

data	data.frame with group columns to summarize by. Can be grouped (see dplyr::group_by()), in which case the function is applied group-wise. This is not to be confused with `group_cols`.
group_cols	Names of columns with group identifiers to summarize columns in `data` by.
cat_cols	Names of categorical columns to summarize. Each categorical level is counted per group. To distinguish between levels with the same name from different `cat_col` columns, we prefix the count column name for each categorical level with parts of the name of the categorical column. This amount can be controlled with `max_cat_prefix_chars`. Normalization when `include_normalized` is enabled: The counts of each categorical level is normalized with log(1 + count).
num_cols	Names of numerical columns to summarize. For each column, the mean and sum is calculated per group. Normalization when `include_normalized` is enabled: Each column is normalized with `num_normalize_fn` before calculating the mean and sum per group.
id_cols	Names of factor columns with IDs to summarize. The number of unique IDs are counted per group. Normalization when `include_normalized` is enabled: The count of unique IDs is normalized with log(1 + count).
summarize_size	Whether to summarize the number of rows per group.
include_normalized	Whether to calculate and include the normalized summary in the output.
rank_weights	A named vector with weights for averaging the rank columns when calculating the `SD_rank` column. The name is one of the balancing columns and the number is its weight. Non-specified columns are given the weight 1. The weights are automatically scaled to sum to 1. When summarizing size (see `summarize_size`), name its weight "size". E.g. c("size" = 1, "a_cat_col" = 2, "a_num_col" = 4, "an_id_col" = 2).
cat_levels_rank_weights	Weights for averaging ranks of the categorical levels in `cat_cols`. Given as a named list with a named vector for each column in `cat_cols`. Non-specified levels are given the weight 1. The weights are automatically scaled to sum to 1. E.g. list("a_cat_col" = c("a" = 3, "b" = 5), "b_cat_col" = c("1" = 3, "2" = 9))
num_normalize_fn	Function for normalizing the `num_cols` columns before calculating normalized group summaries. Only used when `include_normalized` is enabled.

Value

list with two/three data.frames:

Groups

A summary per group.

`cat_cols`: Each level has its own column with the count of the level per group.

`num_cols`: The mean and sum per group.

`id_cols`: The count of unique IDs per group.

Summary

Statistical descriptors of the columns in `Groups`.

Contains the mean, median, standard deviation (SD), interquartile range (IQR), min, and max measures.

Especially the standard deviations and IQR measures can tell us about how balanced the groups are. When comparing multiple `group_cols`, the group column with the lowest SD and IQR can be considered the most balanced.

Normalized Summary

(Disabled by default)

Same statistical descriptors as in `Summary` but for a "normalized" version of the group summaries. The motivation is that these normalized measures can more easily be compared or combined to a single "balance score".

First, we normalize each balance column:

`cat_cols`: The level counts in the original group summaries are normalized with with log(1 + count). This eases comparison of the statistical descriptors (especially standard deviations) of levels with very different count scales.

`num_cols`: The numeric columns are normalized prior to summarization by group, using the `num_normalize_fn` function. By default this applies MinMax scaling to columns such that ~95% of the values are expected to be in the [0, 1] range.

`id_cols`: The counts of unique IDs in the original group summaries are normalized with log(1 + count).

Contains the mean, median, standard deviation (SD), interquartile range (IQR), min, and max measures.

Author(s)

Ludvig Renbo Olsen, [e-mail]

See Also

Other summarization functions: ranked_balances(), summarize_group_cols()

Examples

# Attach packages
library(groupdata2)
library(dplyr)

set.seed(1)

# Create data frame
df <- data.frame(
  "participant" = factor(rep(c("1", "2", "3", "4", "5", "6"), 3)),
  "age" = rep(sample(c(1:100), 6), 3),
  "diagnosis" = factor(rep(c("a", "b", "a", "a", "b", "b"), 3)),
  "score" = sample(c(1:100), 3 * 6)
)
df <- df %>% arrange(participant)
df$session <- rep(c("1", "2", "3"), 6)

# Using fold()

## Without balancing
set.seed(1)
df_folded <- fold(data = df, k = 3)

# Check the balances of the various columns
# As we have not used balancing in `fold()`
# we should not expect it to be amazingly balanced
df_folded %>%
  dplyr::ungroup() %>%
  summarize_balances(
    group_cols = ".folds",
    num_cols = c("score", "age"),
    cat_cols = "diagnosis",
    id_cols = "participant"
  )

## With balancing
set.seed(1)
df_folded <- fold(
  data = df,
  k = 3,
  cat_col = "diagnosis",
  num_col = 'score',
  id_col = 'participant'
)

# Now the balance should be better
# although it may be difficult to get a good balance
# the 'score' column when also balancing on 'diagnosis'
# and keeping all rows per participant in the same fold
df_folded %>%
  dplyr::ungroup() %>%
  summarize_balances(
    group_cols = ".folds",
    num_cols = c("score", "age"),
    cat_cols = "diagnosis",
    id_cols = "participant"
  )

# Comparing multiple grouping columns
# Create 3 fold column that only balance "score"
set.seed(1)
df_folded <- fold(
  data = df,
  k = 3,
  num_fold_cols = 3,
  num_col = 'score'
)

# Summarize all three grouping cols at once
(summ <- df_folded %>%
  dplyr::ungroup() %>%
  summarize_balances(
    group_cols = paste0(".folds_", 1:3),
    num_cols = c("score")
  )
)

# Extract the across-group standard deviations
# The group column with the lowest standard deviation(s)
# is the most balanced group column
summ %>% ranked_balances()

groupdata2 documentation built on July 9, 2023, 6:46 p.m.