Quick Start"

In corrselect: Correlation-Based and Model-Based Predictor Pruning

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 7,
  fig.height = 5,
  dev = "svglite",
  fig.ext = "svg"
)
library(corrselect)

Installation

# Install from CRAN
install.packages("corrselect")

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

Suggested packages (for extended functionality):

• lme4, glmmTMB: Mixed-effects models in modelPrune()

• WGCNA: Biweight midcorrelation (bicor)

• energy: Distance correlation

• minerva: Maximal information coefficient

What corrselect Does

corrselect identifies and removes redundant variables based on pairwise correlation or association. Given a threshold $\tau$, it finds subsets where all pairwise associations satisfy $|a_{ij}| < \tau$ (see vignette("theory") for mathematical formulation).

Interface Hierarchy

corrselect provides three levels of interface:

Level 1: Simple Pruning

corrPrune() - Removes redundant predictors based on pairwise correlation:

• Returns a single pruned dataset

• No response variable required

• Fast greedy or exact search

modelPrune() - Reduces VIF in regression models:

• Returns a single pruned dataset with response

• Iteratively removes high-VIF predictors

• Works with lm, glm, lme4, glmmTMB

Level 2: Structured Subset Selection

corrSelect() - Returns all maximal subsets (numeric data):

• Enumerates all maximal valid subsets satisfying threshold (see vignette("theory"))

• Provides full metadata (size, avg_corr, max_corr, min_corr)

• Exact or greedy search

assocSelect() - Returns all maximal subsets (mixed-type data):

• Handles numeric, factor, and ordered variables

• Uses appropriate association measures per variable pair

• Exact or greedy search

Level 3: Low-Level Matrix Interface

MatSelect() - Direct matrix input:

• Accepts precomputed correlation/association matrices

• No data preprocessing

• Useful for repeated analyses

Quick Examples

corrPrune(): Association-Based Pruning

data(mtcars)

# Remove correlated predictors (threshold = 0.7)
pruned <- corrPrune(mtcars, threshold = 0.7)

# Results
cat(sprintf("Reduced from %d to %d variables\n", ncol(mtcars), ncol(pruned)))
names(pruned)

Variables removed:

attr(pruned, "removed_vars")

How corrPrune() selects among multiple maximal subsets:

When multiple maximal subsets exist (which is common), corrPrune() (in exact mode) selects among them by, in order:

1. Largest subset size: Keeping more variables is preferred over keeping fewer

2. Lowest average absolute correlation: Among subsets of the same (largest) size, the one with the least redundancy is preferred

3. Alphabetically first variable names: A final tiebreaker for deterministic behavior when size and average correlation are both tied

Subset size is checked first, so a smaller subset is never preferred over a larger one even if its average correlation is lower.

To explore all maximal subsets instead of just the optimal one, use corrSelect() (see below).

modelPrune(): VIF-Based Pruning

# Prune based on VIF (limit = 5)
model_data <- modelPrune(
  formula = mpg ~ .,
  data = mtcars,
  limit = 5
)

# Results
cat("Variables kept:", paste(attr(model_data, "selected_vars"), collapse = ", "), "\n")
cat("Variables removed:", paste(attr(model_data, "removed_vars"), collapse = ", "), "\n")

corrSelect(): Enumerate All Maximal Subsets

results <- corrSelect(mtcars, threshold = 0.7)
show(results)

Inspect subsets:

as.data.frame(results)[1:5, ]  # First 5 subsets

Extract a specific subset:

subset_data <- corrSubset(results, mtcars, which = 1)
names(subset_data)

assocSelect(): Mixed-Type Data

# Create mixed-type data
df <- data.frame(
  x1 = rnorm(100),
  x2 = rnorm(100),
  cat1 = factor(sample(c("A", "B", "C"), 100, replace = TRUE)),
  ord1 = ordered(sample(1:5, 100, replace = TRUE))
)

# Handle mixed types automatically
results_mixed <- assocSelect(df, threshold = 0.5)
show(results_mixed)

# Verify all pairwise associations are below threshold
cat("Max pairwise association:", max(results_mixed@max_corr), "\n")

Protecting Variables

Use force_in to ensure specific variables are always retained:

# Force "mpg" to remain in all subsets
pruned_force <- corrPrune(
  data = mtcars,
  threshold = 0.7,
  force_in = "mpg"
)

# Verify forced variable is present
"mpg" %in% names(pruned_force)

Threshold Selection

Common thresholds: 0.5 (strict), 0.7 (moderate, recommended default), 0.9 (lenient).

Lower thresholds are stricter because they allow fewer variable pairs to coexist, resulting in smaller subsets. Higher thresholds permit stronger correlations, retaining more variables.

For detailed threshold selection strategies including visualization techniques, VIF guidelines, and sensitivity analysis, see vignette("advanced").

Interface Selection Guide

| Scenario | Function | Key Parameters | |----------|----------|----------------| | Quick dimensionality reduction | corrPrune() | threshold, mode | | Model-based refinement | modelPrune() | limit (VIF threshold), engine | | Enumerate all maximal subsets | corrSelect() | threshold | | Mixed-type data | assocSelect() | threshold | | Precomputed matrices | MatSelect() | threshold, method | | Protect key variables | Any function | force_in |

Quick Reference

corrPrune()

Removes redundant predictors based on pairwise correlation.

corrPrune(data, threshold = 0.7, measure = "auto", mode = "auto",
          force_in = NULL, by = NULL, group_q = 1, max_exact_p = 100)

| Parameter | Description | Default | |-----------|-------------|---------| | data | Data frame or matrix | required | | threshold | Maximum allowed correlation | 0.7 | | measure | Correlation type: "auto", "pearson", "spearman", "kendall", "bicor", "distance", "maximal" | "auto" | | mode | Algorithm: "auto", "exact", "greedy" | "auto" | | force_in | Variables that must be retained | NULL | | by | Column name(s) for grouped pruning | NULL | | group_q | Quantile for aggregating group correlations (0-1] | 1 | | max_exact_p | Max predictors for exact search when mode = "auto" | 100 |

Returns: Data frame with pruned variables. Attributes: selected_vars, removed_vars.

modelPrune()

Iteratively removes predictors with high VIF from a regression model.

modelPrune(formula, data, engine = "lm", criterion = "vif",
           limit = 5, force_in = NULL, max_steps = NULL, ...)

| Parameter | Description | Default | |-----------|-------------|---------| | formula | Model formula (e.g., y ~ .) | required | | data | Data frame | required | | engine | "lm", "glm", "lme4", "glmmTMB", or custom | "lm" | | criterion | "vif" or "condition_number" | "vif" | | limit | Maximum allowed diagnostic value | 5 | | force_in | Variables that must be retained | NULL |

Returns: Pruned data frame. Attributes: selected_vars, removed_vars, final_model.

corrSelect()

Enumerates all maximal subsets satisfying correlation threshold (numeric data).

corrSelect(df, threshold = 0.7, method = NULL, force_in = NULL,
           cor_method = "pearson", ...)

| Parameter | Description | Default | |-----------|-------------|---------| | df | Data frame (numeric columns only) | required | | threshold | Maximum allowed correlation | 0.7 | | method | Algorithm: "bron-kerbosch", "els" | auto | | cor_method | "pearson", "spearman", "kendall", "bicor", "distance", "maximal" | "pearson" | | force_in | Variables required in all subsets | NULL |

Returns: CorrCombo object with properties: subset_list, avg_corr, min_corr, max_corr.

assocSelect()

Enumerates all maximal subsets for mixed-type data (numeric, factor, ordered).

assocSelect(df, threshold = 0.7, method = NULL, force_in = NULL,
            method_num_num = "pearson", method_num_ord = "spearman",
            method_ord_ord = "spearman", ...)

| Parameter | Description | Default | |-----------|-------------|---------| | df | Data frame (any column types) | required | | threshold | Maximum allowed association | 0.7 | | method_num_num | Numeric-numeric: "pearson", "spearman", etc. | "pearson" | | method_num_ord | Numeric-ordered: "spearman", "kendall" | "spearman" | | method_ord_ord | Ordered-ordered: "spearman", "kendall" | "spearman" |

Returns: CorrCombo object.

MatSelect()

Direct matrix interface for precomputed correlation/association matrices.

MatSelect(mat, threshold = 0.7, method = NULL, force_in = NULL, ...)

| Parameter | Description | Default | |-----------|-------------|---------| | mat | Symmetric correlation/association matrix | required | | threshold | Maximum allowed value | 0.7 | | method | Algorithm: "bron-kerbosch", "els" | auto | | force_in | Variables required in all subsets | NULL |

Returns: CorrCombo object.

corrSubset()

Extracts a specific subset from a CorrCombo result.

corrSubset(res, df, which = "best", keepExtra = FALSE)

| Parameter | Description | Default | |-----------|-------------|---------| | res | CorrCombo object from corrSelect/assocSelect/MatSelect | required | | df | Original data frame | required | | which | Subset index or "best" (largest size, then lowest avg correlation) | "best" | | keepExtra | Include non-numeric columns in output? | FALSE |

Returns: Data frame containing only the selected variables.

Troubleshooting

"No valid subsets found" error - Threshold too strict: all variable pairs exceed it

• Solution: Increase threshold or use force_in to keep at least one variable

VIF computation fails in modelPrune() - Perfect multicollinearity (R² = 1) present

• Solution: Use corrPrune(threshold = 0.99) first to remove near-duplicates

Forced variables conflict - Variables in force_in are too highly correlated with each other

• Solution: Increase threshold or reduce force_in set

Slow performance with many variables - Exact mode is exponential for large p

• Solution: Use mode = "greedy" for p > 25

For comprehensive troubleshooting with code examples, see vignette("advanced"), Section 5.

See Also

• vignette("workflows") - Complete real-world workflows (ecological, survey, genomic, mixed models)

• vignette("advanced") - Algorithmic control and custom engines

• vignette("comparison") - Comparison with caret, Boruta, glmnet

• vignette("theory") - Theoretical foundations and formulation

• ?corrPrune, ?modelPrune, ?corrSelect, ?assocSelect, ?MatSelect

Session Info

sessionInfo()

corrselect documentation built on July 19, 2026, 1:06 a.m.