Skip to contents

Select Variable Subsets with Low Association (Mixed-Type Data Frame Interface)

Source: R/assocSelect.R
assocSelect.Rd

Identifies combinations of variables of any common data type (numeric, ordered factors, or unordered) factors—whose pair-wise association does not exceed a user-supplied threshold. The routine wraps MatSelect() and handles all pre-processing (type conversion, missing-row removal, constant-column removal) for typical data-frame/tibble/data-table inputs.

Usage

assocSelect(
  df,
  threshold = 0.7,
  method = NULL,
  force_in = NULL,
  method_num_num = c("pearson", "spearman", "kendall", "bicor", "distance", "maximal"),
  method_num_ord = c("spearman", "kendall"),
  method_ord_ord = c("spearman", "kendall"),
  ...
)

Arguments

df

A data frame (or tibble / data.table). May contain any mix of:

  • numeric / integer (treated as numeric)

  • ordered factors

  • unordered factors (character vectors are coerced to factors)

threshold

Numeric in \((0,1]\). Maximum allowed pair-wise absolute association. Default 0.7.

method

Character; the subset-search algorithm. One of "els" or "bron-kerbosch". If NULL (default) the function selects automatically: ELS when force_in is supplied, otherwise Bron–Kerbosch.

force_in

Optional character vector or column indices specifying variables that must appear in every returned subset.

method_num_num

Association measure for numeric–numeric pairs. One of "pearson" (default), "spearman", "kendall", "bicor", "distance", or "maximal".

method_num_ord

Association measure for numeric–ordered pairs. One of "spearman" (default) or "kendall".

method_ord_ord

Association measure for ordered–ordered pairs. One of "spearman" (default) or "kendall".

...

Additional arguments passed unchanged to MatSelect() (e.g., use_pivot = TRUE for Bron–Kerbosch).

Value

A CorrCombo object containing:

  • all valid subsets,

  • their summary association statistics,

  • metadata (algorithm used, rows kept, forced-in variables, etc.).

The object’s show() method prints the association metrics that were actually used for this data set.

Two additional attributes are attached to the returned object:

assoc_methods_used

Named list mapping each variable-type-pair combination that actually occurs in df (e.g. "numeric_numeric", "numeric_factor") to the association method used for it.

assoc_methods_all

Named list mapping every possible variable-type-pair combination to its resolved method, regardless of whether that combination occurs in df.

Details

A single call can therefore screen a data set that mixes continuous and categorical features and return every subset whose internal associations are “sufficiently low” under the metric(s) you choose.

Rows containing NA are dropped with a warning; constant columns (a single distinct value) are excluded with a warning.

The default association measure for each variable-type combination is:

numeric – numeric

method_num_num (default "pearson")

numeric – ordered

method_num_ord

numeric – unordered

"eta" (the correlation ratio \(\eta = \sqrt{\eta^{2}}\) of a one-way ANOVA)

ordered – ordered

method_ord_ord

ordered – unordered

"cramersv"

unordered – unordered

"cramersv"

Every measure above is a correlation magnitude in \([0,1]\), so threshold means the same thing for every variable-type pair. \(\eta\) is the multiple correlation between the numeric variable and the factor and equals the absolute point-biserial correlation for a two-level factor; Cramer's V equals the absolute phi coefficient for a 2x2 table. A binary variable therefore gets the same association whether it is supplied as a 0/1 numeric column or as a two-level factor. External packages are required for "bicor" (WGCNA), "distance" (energy), and "maximal" (minerva); an informative error is thrown if they are missing.

See also

corrSelect(), MatSelect(), corrSubset()

Examples

set.seed(42)
df <- data.frame(
  height = rnorm(15, 170, 10),
  weight = rnorm(15, 70, 12),
  group  = factor(rep(LETTERS[1:3], each = 5)),
  score  = ordered(sample(c("low","med","high"), 15, TRUE))
)

## keep every subset whose internal associations <= 0.6
assocSelect(df, threshold = 0.6)

## use Kendall for all rank-based comparisons and force 'height' to appear
assocSelect(df,
            threshold       = 0.5,
            method_num_num  = "kendall",
            method_num_ord  = "kendall",
            method_ord_ord  = "kendall",
            force_in        = "height")