Confirm a frozen pattern basis on independent observations
Source:R/pattern_inference.R
pattern_confirm.RdConfirm a frozen pattern basis on independent observations
Usage
pattern_confirm(
fit,
dataset,
design,
block_var = NULL,
inference = confirmation_plan(),
observation_ids,
discovery_ids,
nuisance = NULL,
feature_ids,
preprocessing_id,
subject_id = NULL
)
# S3 method for class 'pattern_confirmation'
print(x, ...)Arguments
- fit
A pattern fit, view, or global result with a refit.
- dataset
Confirmation observations by all input features, or an MVPA dataset whose training partition contains only confirmation observations.
- design
Confirmation targets (factor or numeric matrix), or an MVPA design with those targets in its training partition.
- block_var
Row-aligned independent run/block labels. Required for block-based error models; these labels do not add nuisance intercepts.
- inference
- observation_ids
Globally meaningful, unique confirmation row IDs.
- discovery_ids
Complete discovery row IDs in the same namespace, including all rows used for selection, tuning, or preprocessing. Required: the legacy fit's positional train/test IDs are not adequate provenance.
- nuisance
Numeric nuisance columns, without an intercept. Categorical nuisance variables must be coded by the caller using model.matrix.
- feature_ids
Unique IDs for all input columns, in their input order.
- preprocessing_id
A stable identifier for measurement units and the preprocessing recipe; matching strings are a caller assertion, not proof.
- subject_id
Optional unique participant ID, required for group analysis.
- x
A confirmation result to print.
- ...
Reserved.
Value
A pattern_confirmation: original-feature-unit unpenalized
estimates and SEs, t and omnibus F tests, Holm-adjusted p values, complete
within-feature coefficient covariance (factorized for independent errors),
frozen basis and provenance. Sign-flip plans also return bootstrap p and
max-statistic adjusted p values. Nonestimable sampling distributions are NA.
Details
Regresses every input feature on frozen target scores and nuisance columns, including an intercept. Discovery feature screening is not repeated or used to select the confirmation hypothesis family. All target scores must be estimable after nuisance adjustment. Independent Gaussian homoskedastic errors give exact t/F tests; block methods are approximate. Holm correction covers all feature-component tests, and separately all omnibus tests. Component columns depend on the frozen basis; omnibus tests are invariant to nonsingular changes of score coordinates. Rank selection is not a rank hypothesis test and this API does not report a supported population rank. IDs detect overlap but the caller must establish genuine independence, including no shared preprocessing fit or correlated repeated measurements.