Check design matrix for multicollinearity

Convenience helper to quickly flag highly correlated regressors. This is a pairwise check only: constant columns (including intercepts) are excluded, and ok = TRUE does not establish full rank, estimability or good contrast precision. Dependencies involving several columns, such as task regressors summing to an intercept in a no-rest design, can pass this check. Use check_estimability() on the full design, including baseline/nuisance regressors, to inspect those dependencies.

Usage

check_collinearity(X, threshold = 0.9)

Arguments

X

A numeric design matrix (or an event_model).

threshold

Absolute correlation above which a pair is flagged. Default 0.9.

Value

A list with elements: ok (logical), pairs (data.frame with offending pairs and their correlations). Invisibly returns the same list.

Examples

# Create a simple event model
des <- data.frame(
  onset = c(0, 10, 20, 30),
  run = 1,
  cond = factor(c("A", "B", "A", "B"))
)
sframe <- fmrihrf::sampling_frame(blocklens = 40, TR = 1)
emodel <- event_model(onset ~ hrf(cond), data = des, block = ~run,
                      sampling_frame = sframe)

# Check for multicollinearity
res <- check_collinearity(design_matrix(emodel), threshold = 0.95)
if (!res$ok) print(res$pairs)