corrselect

every uncorrelated subset of predictors

corrselect removes redundant predictors before modelling, either by keeping only predictors whose pairwise correlations do not exceed a threshold or by dropping predictors with a high variance inflation factor (VIF) from a regression model.

The correlation search treats each predictor as a node of a graph and joins two predictors when their correlation is at or below the threshold. Every set of predictors that can be kept together is then a clique of this graph. corrselect enumerates the maximal ones, the subsets to which no further predictor can be added, with the Bron–Kerbosch and Eppstein–Löffler–Strash algorithms in C++. It returns either one pruned data frame or every maximal subset, so you can choose the subset that keeps the variables your analysis needs.

corrselect supports:

• Pruning in one call. corrPrune() returns one data frame. It searches exactly up to 100 predictors (max_exact_p) and switches to a greedy search in C++ beyond that.
• Choosing among every valid subset. corrSelect() returns every maximal subset, ranked by size and then by average correlation.
• Mixed-type data. Numeric, factor and ordered columns are compared with a measure suited to each pair: Pearson correlation for two numeric columns, the correlation ratio eta for a numeric column and a factor, and Cramér’s V for two factors.
• Fixed variables and grouped data. force_in keeps named predictors in every result, and by requires the threshold to hold within each group.
• Model-based pruning. modelPrune() refits a model after each removal, by VIF or condition number, with built-in engines for linear models, generalized linear models, lme4 and glmmTMB, and custom engines for any other modelling package.

In exact mode the result is the same on every run, and the largest subset it returns keeps at least as many predictors as any greedy removal at the same threshold. Theory and Formulation defines the threshold graph and the algorithms.

Installation

Install the released version from CRAN. Windows and macOS get a binary package and need no build tools:

install.packages("corrselect")

The development version compiles C++ code. Install the build tools for your operating system first:

Operating system	Build tools
Windows	Rtools matching your R version
macOS	Run xcode-select --install in Terminal
Debian / Ubuntu	Install r-base-dev

Check that R can find the compiler:

pkgbuild::has_build_tools(debug = TRUE)

Then install corrselect from GitHub:

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

corrselect needs no system libraries beyond the compiler. Its C++ code links only to Rcpp, which is installed from CRAN with the package.

Usage

1. Prune a data frame to one subset with corrPrune().
2. List every maximal subset with corrSelect() and take one out as a data frame with corrSubset().
3. Prune the predictors of a regression model by VIF with modelPrune().

The input is a data frame of candidate predictors, with numeric, factor or ordered columns. Correlations are computed on the rows without missing values, and the returned data frame keeps every row. The threshold defaults to 0.7.

library(corrselect)

pruned <- corrPrune(mtcars)
names(pruned)
#> [1] "mpg"  "drat" "qsec" "gear" "carb"

Five of the eleven columns of mtcars remain, and no pair of them has a correlation above 0.7. With eleven predictors the search is exact.

Several subsets usually satisfy the same threshold. corrSelect() lists all maximal ones. corrSubset() returns one of them, the top-ranked by default:

res <- corrSelect(mtcars)
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
#>   [ 4] wt, qsec, am, carb                0.373  0.692     4
#>   [ 5] wt, qsec, gear, carb              0.388  0.656     4
#>   ... (10 more combinations)

best <- corrSubset(res, mtcars)
head(best, 3)
#>                mpg drat  qsec gear carb
#> Mazda RX4     21.0 3.90 16.46    4    4
#> Mazda RX4 Wag 21.0 3.90 17.02    4    4
#> Datsun 710    22.8 3.85 18.61    4    1

modelPrune() takes a model formula instead. It removes the predictor with the highest VIF, refits, and repeats until no VIF exceeds 5:

fit <- modelPrune(mpg ~ cyl + disp + hp + wt, data = mtcars)
names(fit)
#> [1] "mpg" "cyl" "hp"  "wt"

disp is removed, and the response stays in the returned data frame.

Keeping variables and holding the threshold across groups

force_in keeps predictors in the result whatever else is removed. by computes the correlations within each level of a grouping column, and the threshold has to hold in every group:

corrPrune(mtcars, threshold = 0.7, force_in = c("hp", "wt"))

cols <- c("x1", "x2", "x3", "x4", "x5", "site")
long <- longitudinal_example[, cols]
corrPrune(long, threshold = 0.6, by = "site")

group_q relaxes the grouped rule to a quantile of the per-group correlations. Advanced Topics shows it on simulated data.

Mixed-type data

assocSelect() enumerates maximal subsets of a data frame with numeric, factor and ordered columns, with the measure chosen per pair:

set.seed(1)
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() uses the same measures, so one call handles all-numeric and mixed data. For two numeric columns, measure selects one of six measures, three of which need a suggested package; the corrPrune() reference lists them.

Pruning with other modelling packages

modelPrune() has built-in engines for linear models, generalized linear models, lme4 and glmmTMB. With lme4 and glmmTMB only the fixed effects are candidates for removal. Any other package plugs in as a custom engine, a list of two functions: one that fits the model and one that returns a score per fixed effect. Advanced Topics builds engines for INLA and mgcv.

corrPrune or modelPrune?

	corrPrune()	modelPrune()
Needs a model	No	Yes
Based on	Pairwise correlation or association	Model diagnostics (VIF, condition number)
Works without a response	Yes	No
Mixed models	No	Yes (lme4, glmmTMB)
Suited to	Exploratory analysis, many predictors	Regression workflows, VIF reduction

Start with corrPrune() to reduce the number of predictors, then use modelPrune() to remove the remaining collinearity within the model you fit.

Exact and greedy search

corrPrune() chooses the search mode from the number of predictors: exact enumeration up to max_exact_p (100 by default) and the greedy C++ search beyond that. Exact mode returns the largest subset that satisfies the threshold. Greedy mode builds one subset by successive removals and runs in time quadratic in the number of predictors, so it handles wide data without that guarantee. mode sets either one:

corrPrune(mtcars, threshold = 0.7, mode = "exact")
corrPrune(mtcars, threshold = 0.7, mode = "greedy")

Compared with caret::findCorrelation()

caret’s findCorrelation() removes predictors one at a time: it takes the most correlated pair and drops the member with the larger mean absolute correlation, until no pair exceeds the cutoff. On mtcars at the same threshold, with caret 7.0.1:

drop <- caret::findCorrelation(cor(mtcars), cutoff = 0.7)
names(mtcars)[-drop]
#> [1] "qsec" "gear" "carb"

exact <- corrPrune(mtcars, threshold = 0.7, mode = "exact")
names(exact)
#> [1] "mpg"  "drat" "qsec" "gear" "carb"

Comparison with Alternatives runs both on a larger dataset and sets corrselect beside Boruta, glmnet and manual VIF removal.

Documentation

Start with Quick Start: each function on a small example, with its main arguments.

• Complete Workflows: Real-World Examples: ecological, survey, high-dimensional and mixed-model data pruned and modelled end to end.
• Comparison with Alternatives: corrselect beside caret, Boruta, glmnet and manual VIF removal on the same data.
• Advanced Topics: exact and greedy search with timings, tie-breaking, grouped pruning and custom engines.
• Theory and Formulation: the threshold graph, maximal cliques and the enumeration algorithms.
• Full reference: every function.

Getting help

Questions and bug reports go to the GitHub issues. A bug report is easiest to act on with a small reproducible example, for instance made with reprex.

Contributing

Pull requests are welcome; see the contributing guide.

Support

“Software is like sex: it’s better when it’s free.” — Linus Torvalds

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

corrselect is licensed under MIT.

Citation

If you use corrselect, please cite the article in the Journal of Open Source Software and the package version you used.

Colling G (2026). "corrselect: Fast and flexible predictor pruning for
data analysis and modeling." _Journal of Open Source Software_,
*11*(118), 9539. doi:10.21105/joss.09539
<https://doi.org/10.21105/joss.09539>.
<https://joss.theoj.org/papers/10.21105/joss.09539>.

Colling G (2026). _corrselect: Correlation-Based and Model-Based
Predictor Pruning_. doi:10.32614/CRAN.package.corrselect
<https://doi.org/10.32614/CRAN.package.corrselect>. R package version
3.3.0, <https://gillescolling.com/corrselect/>.

BibTeX:

@Article{colling2026corrselectjoss,
  title = {corrselect: Fast and flexible predictor pruning for data analysis and modeling},
  author = {Gilles Colling},
  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},
}

@Manual{colling2026corrselect,
  title = {corrselect: Correlation-Based and Model-Based Predictor Pruning},
  author = {Gilles Colling},
  year = {2026},
  note = {R package version 3.3.0},
  doi = {10.32614/CRAN.package.corrselect},
  url = {https://gillescolling.com/corrselect/},
}

Download the citation as BibTeX or RIS.