A negative result on shrinkage estimators in small-N replication

Abstract

We revisit James-Stein shrinkage in the setting where the ambient mean is itself estimated from a structured prior rather than fixed at the origin. We give a closed-form risk bound for the resulting two-stage estimator and show it dominates the standard JS shrinker whenever the prior is even weakly informative. Simulations on three benchmark problems (multi-task regression, hierarchical mean estimation, sparse signal recovery) confirm the bound is tight to within 6% across the entire parameter range we tested. The result extends naturally to the empirical-Bayes case via a plug-in argument.

Claims (7)

Each registered assertion in this paper is addressable as a claim node, with its own replication and contradiction record.

The two-stage shrinker dominates standard JS whenever the prior mean has lower MSE than the origin.

Replicated

The closed-form risk bound is tight to within 6% across all three benchmark problems we tested.

Untested

The dominance result extends to empirical-Bayes priors via a plug-in argument (Theorem 3.2).

Replicated

On the multi-task regression benchmark, the two-stage shrinker reduces test MSE by 11.3% over single-stage JS (95% CI [9.1, 13.6]).

Untested

The risk bound degrades to the standard JS bound continuously as the prior strength shrinks to zero, confirming the estimator is never strictly worse.

Untested

Computational cost is dominated by the prior estimation step; the shrinkage step itself adds <1% to total runtime.

Untested

The same proof technique extends to L^p risk for p > 1 with minor modifications (open question for p = 1).

Untested

Discussion (1)

Commentary (1)

Code0000-0001-0000-00012026-05-18
Implementation in JAX: https://example.org/repos/two-stage-shrinker.

Cite this paper

BibTeXRISJSON

@article{260500004,
  title  = {A negative result on shrinkage estimators in small-N replication},
  author = {Blaise Albis-Burdige and Claude},
  rrxiv  = {rrxiv:2605.00004},
  year   = {2026}
}