Skip to contents

Identify Highly Correlated Variables

Source: R/findCorrX.R
findCorrX.Rd

Identifies variables in a correlation matrix that should be removed due to high pairwise correlations above a specified cutoff. The algorithm uses a greedy approach similar to caret::findCorrelation(), but extends it with flexible scoring methods and multiple output formats.

Usage

findCorrX(
  x,
  cutoff = 0.9,
  method = c("mean", "max", "median"),
  output = c("index", "names", "logical", "report"),
  verbose = FALSE
)

Arguments

x

a symmetric correlation matrix

cutoff

numeric threshold in (0, 1). Pairs with absolute correlation above this value are considered too highly correlated.

method

character string specifying how variable importance is scored. One of "mean", "max", or "median". Default is "mean".

output

character string specifying the return format:

  • "index" for indices of variables to remove (default)

  • "names" for column names of variables to remove

  • "logical" for a logical vector indicating removed variables

  • "report" for a detailed list of removed and retained variables and the decision log

verbose

logical; if TRUE, progress information is printed

Value

depending on output:

"index"

integer vector of column indices to remove

"names"

character vector of column names

"logical"

logical vector with one element per column

"report"

list with elements removed, kept, and log

Details

The function iteratively examines pairs of variables with correlations above cutoff. For each such pair, the variable with the higher overall correlation (based on method) is removed.

The scoring is computed once at the beginning and kept fixed throughout, ensuring deterministic and efficient behavior. Because the columns are then processed in decreasing score order, the variable removed from a pair is always the earlier - i.e. the higher-scoring - one.

This is a greedy heuristic and does not guarantee a globally optimal solution.

See also

Other assoc.continuous: corPart(), corPolychor(), hoeffdingD(), keepSig(), pearsonCor(), spearmanCor()

Examples

set.seed(123)
m <- matrix(rnorm(100), ncol = 5)
colnames(m) <- paste0("V", 1:5)
cmat <- cor(m)

findCorrX(cmat, cutoff = 0.8)
#> integer(0)
findCorrX(cmat, cutoff = 0.8, method = "max", output = "names")
#> character(0)