Skip to contents

Test frozen component association and incremental decoding value

Usage

pattern_component_tests(fit, confirmation, dataset, design, calibration = NULL)

Arguments

fit

The discovery fit used for confirmation.

confirmation

A pattern_confirmation from that fit.

dataset, design

The exact confirmation data and targets used by pattern_confirm. Content hashes are checked before computing scores.

calibration

Optional list with dataset, design, and observation_ids. Its IDs must be included in the declared discovery set. Full and leave-one-component-out ordinary least-squares decoding heads are fitted on these rows alone. Without calibration only association is tested.

Value

A list with an association table and, when calibration is supplied, an incremental-loss table, fitted decoding heads, and row-level losses.

Details

Association is the partial correlation of each frozen decoded score with its corresponding frozen target score, adjusting for confirmation nuisance columns (not for other components). Tests use the confirmation error model and Holm correction over components.

Incremental gain is reduced-head minus full-head squared prediction error, averaged over raw target columns, then within independent blocks, then equally across blocks. Categorical targets use one-hot squared loss; linear heads are not probability models. Positive gain favors retaining the component. The paired block-mean t test is approximate; sign-flip plans use a centered, studentized block wild bootstrap, also approximate. These tests concern separately fitted decoding heads, not the original Gaussian posterior decoder or a supported latent rank. Component tests depend on the frozen basis. Score degeneracy is reported for association and rejected when a decoding head cannot be estimated.