Refit a coxtrans model with hard constraints

Extracts the constraint structure (active set, sharing pattern, prior constraints) from a penalized coxtrans fit, encodes them as hard constraints via a link matrix, and solves an unpenalized stratified Cox model in the reparametrized space. Returns debiased MLE coefficients and valid standard errors for target-group features.

Usage

refit(object, ...)

# S3 method for class 'coxtrans'
refit(object, ...)

Arguments

object

A fitted coxtrans object.

...

Additional arguments (unused).

Value

An object of class refit.coxtrans with components:

coefficients

Matrix (p x 5) with columns coef, exp(coef), se(coef), z, Pr(>|z|). Rows correspond to original features; inactive features have all zeros except Pr(>|z|) = 1.

coxph_fit

The underlying coxph object, or NULL.

n

Total number of observations used in refit.

nevent

Number of events.

active_set

Integer vector of active feature indices.

Details

The refit procedure:

1. Build the link matrix \(L\) from the penalized fit, encoding shared coefficients, prior constraints, and group-specific parameters.

2. Construct reparametrized design \(Z = \mathrm{bdiag}(X_1, \ldots, X_K) L_{[:, \text{nonzero}]}\).

3. Fit coxph(Surv ~ Z + strata(group)) — unpenalized stratified Cox with hard constraints built into the design.

4. Map estimates back to target coefficients via \(L\), compute standard errors via delta method.