README

Hand it a data frame and a threshold. corrselect finds every maximal set of predictors whose pairwise correlations all stay under that threshold, by graph enumeration (Bron–Kerbosch / Eppstein–Löffler–Strash) in C++, and returns whichever view you need: a single pruned data frame, or every valid subset to choose from.

library(corrselect)

# prune to one low-correlation set, in a single call
corrPrune(mtcars, threshold = 0.7)

# or enumerate *every* maximal low-correlation set
corrSelect(mtcars, threshold = 0.7)

Every valid choice, ranked

corrSelect() returns every maximal low-correlation subset, ranked by size then average correlation:

res <- corrSelect(mtcars, threshold = 0.7)
res
#> CorrCombo object
#> -----------------
#>   Method:      bron-kerbosch
#>   Correlation: pearson
#>   Threshold:   0.700
#>   Subsets:     15 maximal subsets
#>   Data Rows:   32 used in correlation
#>   Pivot:       TRUE
#>
#> Top combinations:
#>   No.  Variables                          Avg    Max    Size
#>   ------------------------------------------------------------
#>   [ 1] mpg, drat, qsec, gear, carb       0.416  0.700     5
#>   [ 2] cyl, drat, qsec, gear, carb       0.434  0.700     5
#>   [ 3] mpg, drat, vs, gear, carb         0.466  0.700     5
#>   ... (12 more combinations)

corrSubset(res, mtcars, which = "best")   # pull the top-ranked subset back out as a data frame

Keeping variables you already trust, across groups

force_in protects variables that must survive pruning regardless of what else gets dropped; by requires the threshold to hold separately inside every group:

# hp and wt are kept no matter what else is removed
corrPrune(mtcars, threshold = 0.7, force_in = c("hp", "wt"))

# threshold holds within every level of `site`, checked group by group
corrPrune(longitudinal_example[, c("x1", "x2", "x3", "x4", "x5", "site")],
          threshold = 0.6, by = "site")

Mixed-type data

assocSelect() picks the right association measure for each pair of columns (Pearson for numeric-numeric, the correlation ratio eta for numeric-factor, Cramer’s V for factor-factor), and enumerates maximal subsets under all of them at once:

df <- data.frame(
  height = rnorm(30, 170, 10),
  weight = rnorm(30, 70, 12),
  group  = factor(sample(c("A", "B"), 30, TRUE)),
  rating = ordered(sample(c("low", "med", "high"), 30, TRUE))
)

assocSelect(df, threshold = 0.6)

corrPrune() runs the same mixed-type dispatch under the hood, so a single pruning call works whether the input is all-numeric or a mix of numeric, factor, and ordered columns. Six numeric-numeric measures are available, selectable via measure / method_num_num: "pearson", "spearman", "kendall", "bicor" (WGCNA), "distance" (energy), and "maximal" (minerva).

Model-based pruning with any engine

modelPrune() removes predictors by VIF or condition number, refitting after each removal. Built-in engines cover lm, glm, lme4, and glmmTMB; anything else plugs in as a two-function engine:

# linear model, VIF threshold
modelPrune(mpg ~ cyl + disp + hp + wt, data = mtcars, limit = 5)

# any modeling package, via a custom engine (here: INLA)
inla_engine <- list(
  name = "inla",
  fit  = function(formula, data, ...) {
    INLA::inla(formula, data = data, family = "gaussian", ...)
  },
  diagnostics = function(model, fixed_effects) {
    scores <- model$summary.fixed[, "sd"]          # posterior SD as the badness score
    setNames(scores, rownames(model$summary.fixed))[fixed_effects]
  }
)
modelPrune(y ~ x1 + x2, data = df, engine = inla_engine, limit = 0.5)

corrPrune or modelPrune?

	`corrPrune()`	`modelPrune()`
Needs a model?	No	Yes
Based on	Pairwise correlation / association	Model diagnostics (VIF)
Works without a response?	Yes	No
Mixed models?	No	Yes (`lme4`, `glmmTMB`)
Best for	Exploratory analysis, large `p`	Regression workflows, VIF reduction

Use corrPrune() first to cut dimensionality, then modelPrune() for final cleanup inside a modeling framework.

Exact where it matters, greedy where it scales

corrPrune() picks a search mode for you: exact enumeration up to max_exact_p predictors (100 by default), and a greedy C++ backend beyond that. Exact mode guarantees the largest subset satisfying the threshold; greedy mode trades that guarantee for speed on wide data. Both are available directly, so you can pin one explicitly:

corrPrune(mtcars, threshold = 0.7, mode = "exact")   # guaranteed-maximal, small/medium p
corrPrune(mtcars, threshold = 0.7, mode = "greedy")  # fast, approximate, any p

caret::findCorrelation() solves this with greedy iterative removal, which is order-dependent and can vary between runs on the same data. corrPrune(mode = "exact") solves the same threshold constraint by maximal-clique enumeration, retaining at least as many variables and returning the same answer every run. The Comparison vignette walks through this and three other alternatives (Boruta, glmnet, manual VIF removal) side by side on the same dataset.

What’s in the box

Installation

install.packages("corrselect")            # CRAN

install.packages("pak")                   # development version
pak::pak("gcol33/corrselect")

Documentation

Support

I’m a PhD student who builds R packages in my free time because I believe good tools should be free and open. I started these projects for my own work and figured others might find them useful too.

If this package saved you some time, buying me a coffee is a nice way to say thanks. It helps with my coffee addiction.

License

Citation

@article{corrselect,
  author  = {Colling, Gilles},
  title   = {corrselect: Fast and flexible predictor pruning for data analysis and modeling},
  journal = {Journal of Open Source Software},
  year    = {2026},
  volume  = {11},
  number  = {118},
  pages   = {9539},
  doi     = {10.21105/joss.09539},
  url     = {https://joss.theoj.org/papers/10.21105/joss.09539}
}

corrselect