driver_isolation_reliability

The confidence columns every driver-isolation rating carries (WI-16a), at any grain.

The confidence columns every driver-isolation rating carries (WI-16a), at any grain. Given a raw rating, its standard error, the signal variance tau^2 of that rating at that grain and season, and the rating’s method score and grade (from the driver_isolation_method_scores seed), emits four columns: <prefix>_lambda reliability tau^2 / (tau^2 + SE^2), in [0, 1] <prefix>_shrunk_gain_s lambda * raw: the normal-normal posterior mean with prior mean 0 (every rating is field- or peer-centred). Equivalent to bayesian_shrinkage() with prior_weight = sigma_w^2 / tau^2 in effective laps when SE^2 = sigma_w^2 / n_eff. <prefix>_confidence_pct ROUND(100 * lambda * method_score); NULL without a lambda or before the method is validated. A declared policy, not an estimated quantity. <prefix>_trust_label ‘unvalidated’ (method score NULL), ‘suppress’ (grade F, no lambda, or confidence < 25), ‘thin’ (25-50), ‘indicative’ (50-75), ‘solid’ (>= 75); NULL when there is no raw value. Floors come before this: a raw value is already NULL below its sample floor, so the shrinkage only handles the range between “just enough” and “plenty”.

Signature

{{ driver_isolation_reliability(prefix, raw, se, tau2, method_score, method_grade) }}

Arguments

Argument	Description
prefix	column-name prefix, e.g. ‘pure_skill_5lap’
raw	expression for the raw rating (s/lap, positive = faster)
se	expression for its standard error (s/lap)
tau2	expression for the signal variance (s^2)
method_score	expression for the rating’s method score in [0, 1], or NULL
method_grade	expression for the rating’s grade (A/B/C/F, or ‘unvalidated’)

Usage

SELECT {{ driver_isolation_reliability('pure_skill', 'pure_raw', 'pure_se',
                                           'tau.pure_tau2', 'ms.method_score',
                                           'ms.grade') }}