Project New Values onto a Frozen Rank-Normal Reference

Projects new_vals onto the rank-based inverse-normal transform defined by train_vals, without ever refitting or observing new_vals when determining the transform. Each new value is treated as a single insertion into the frozen, sorted train_vals reference: its rank is the count of reference values at or below it (ties in train_vals share the same rank), and that rank is converted to a normal quantile using the same Blom-type constant (k = 0.375) as RNOmni::RankNorm().

Usage

project_rank_norm(train_vals, new_vals, k = 0.375)

Arguments

train_vals

Numeric vector, the frozen reference distribution (e.g. a baseline cohort's biomarker values). Must not contain NA.

new_vals

Numeric vector of new values to project onto train_vals. May be a single value.

k

Blom-type constant, passed through to match RNOmni::RankNorm()'s default.

Value

A numeric vector the same length as new_vals, giving each value's projected normal quantile.

Details

This is useful for scoring new samples (including a single new sample) against a fixed reference cohort without leaking information from the new samples into the reference's distribution – e.g. avoiding the train/test leakage that results from rank-normalizing a pooled reference+new-sample dataset together.

Note this does not reproduce RNOmni::RankNorm(train_vals) bit-for-bit when new_vals happens to equal train_vals: it always treats each new value as an additional observation inserted into the n-sample reference (using n + 1 in the denominator), rather than as one of the original n samples. The two are close but not identical.

See also

normalize_wss_biomarkers()

Examples

train_vals <- rnorm(200)
project_rank_norm(train_vals, new_vals = c(-1, 0, 1))
#> [1] -1.0271543 -0.1248768  0.8527665