Skip to contents

Predict from a Random Forest Meta-Analysis

Usage

# S3 method for class 'mars_rf'
predict(
  object,
  newdata = NULL,
  type = c("response", "working", "matrix"),
  re_form = c("conditional", "population"),
  matrix_action = c("warn", "error", "repair", "none"),
  ...
)

Arguments

object

Fitted mars_rf object.

newdata

Data frame of raw predictors. NULL uses the prepared training data. Correlation matrix prediction additionally needs the internal effect and study ID columns named in object$effect_id and object$study_id.

type

response returns predictions on the effect scale; working returns the forest fitting scale; matrix returns complete named correlation matrices.

re_form

conditional includes available known-group BLUPs; population omits every random-effect contribution. Unknown groups receive zero BLUPs.

matrix_action

warn checks complete correlation matrices and reports invalid predictions; error rejects them; none skips checks; repair explicitly projects complete matrices using Matrix::nearPD(corr=TRUE).

...

Additional arguments are rejected.

Value

A numeric vector or a list of correlation matrices. Matrix results include a diagnostics attribute with eigenvalues, repair indicators and maximum changes. Repaired vector results include matrix_diagnostics.

Details

Back-transforming Fisher-z predictions ensures pairwise bounds but does not ensure joint matrix validity. Raw-scale extrapolation can violate bounds. Repair requires each group to contain every effect ID exactly once; checks on vector predictions inspect only complete groups. Transformations and EM conditional moments are frozen at training. These are point predictions, without prediction intervals or confirmatory inference.