Pulse Documentation

name: op-reg-mod-resample
description: Spec-level resampling modifier (jackknife / bootstrap) that composes with an OLS or GLM fit; replaces analytical std errors with empirical estimates. Non-empty value forces the buffered path.
kind: operator
category: REG
operator: REG_RESAMPLE
type: reference
applies_to: process, compose, predict
examples_tags: [regression, resampling, jackknife, buffered-pipeline]

Regression operators emit coefficient + diagnostics; no Response.Components. This modifier overwrites StdErrors / PValues in Response.Regressions[i]; the point estimate stays the full-data fit.

Params

Top-level resample on RegressionSpec plus two bootstrap knobs. A wrapper, not a fit: the host fit runs once per replicate (jackknife = n refits; bootstrap = bootstrap_iters) and β samples are aggregated.

Inputs

Inherits target / predictors from the host RegressionSpec. No additional field inputs.

Output

• StdErrors[name] — jackknife √((n−1)/n · Σ(β⁽⁻ⁱ⁾ − β̄)²); bootstrap sample std of replicates.
• PValues[name] — jackknife pivotal; bootstrap percentile-method two-sided.
• Resample echoes back; Coefficients unchanged. For l1 / elasticnet OLS, bootstrap suppresses PROCESSING_REGRESSION_APPROXIMATE_SE.

Gotchas

• Any non-empty resample forces buffered — RegressionSpec.Streamable() → false. Predict reports the downgrade.
• Jackknife is O(n) refits; on large cohorts prefer bootstrap with moderate replicates.
• Composes with selection — the resample wraps the selected-feature fit.
• rng_seed = 0 is deterministic per request but cohort-dependent; pin non-zero for cross-cohort reproducibility.
• Never weighted: any row weight in force on the regression (slot, request or Options.DefaultWeight) is PULSE_WEIGHT_UNSUPPORTED with details.reason; "weight": null on the slot runs it unweighted.

See

• pulse_examples_search tags=[jackknife]
• Skills: regression-modeling, op-reg-ols, op-reg-glm, op-reg-mod-selection