Selective Winner's Curse

• Selective Winner’s Curse is the post-selection bias where effect size estimates are exaggerated due to conditioning on achieving statistical significance.
• The phenomenon occurs because selection based on threshold criteria, such as |b|/se > c, truncates the sampling distribution, leading to biased point estimates and undercoverage in confidence intervals.
• Mitigation strategies include using shrinkage methods, selective confidence interval adjustments, and separating selection from evaluation to reduce overestimation in adaptive settings.

Searching arXiv for recent and foundational papers on selective winner’s curse and closely related selective-inference settings. Selective winner’s curse is the post-selection distortion that arises when attention, publication, decision-making, or formal inference is conditioned on an estimate having “won” some data-dependent competition. In the canonical statistical version, one observes an unbiased estimator $b$ of a scalar effect $\beta$, with $b\sim N(\beta,se^2)$, and then conditions on a selection event such as $|b|/se>c$, typically $c=1.96$ for two-sided $5\%$ significance. After this conditioning, ordinary unconditional guarantees such as unbiasedness and nominal confidence-interval coverage no longer apply; the observed magnitude $|b|$ is systematically exaggerated, and confidence intervals can undercover, especially in low-power regimes (Zwet et al., 2020). More generally, the same mechanism appears whenever one selects the empirically best model, treatment, subgroup, policy, prompt, or instrument and then interprets the selected estimate as if it had not been chosen adaptively. Recent work has extended this idea from significance filtering to shared-validation model selection, randomized experiments, offline policy learning, adaptive LLM benchmarking, Mendelian randomization, and post-selection confidence procedures, clarifying that “winner’s curse” is often a structured selective-inference problem rather than a generic informal warning (Pena et al., 23 Feb 2026, Berman et al., 16 May 2026, Xu et al., 7 May 2026).

1. Formal statistical definition

In the formulation of van Zwet and Cator, the core object is the significance filter: the practice of conditioning on the event

$\frac{|b|}{se}>c,$

where $b$ is an unbiased estimator of a scalar effect $\beta$, $\beta$0 is known, and

$\beta$1

The parameter $\beta$2 may represent a difference in means, a regression coefficient, a log odds ratio, or a log hazard ratio (Zwet et al., 2020).

The central distinction is between unconditional sampling properties and conditional properties after selection. Unconditionally, $\beta$3 is unbiased for $\beta$4, and the standard normal-theory confidence interval has nominal coverage. Once one conditions on statistical significance, however, the sampling law is truncated by the event $\beta$5, and inference becomes selective. The consequence is not merely a signed bias problem but an exaggeration problem for the magnitude $\beta$6, which is why the paper focuses on $\beta$7 rather than $\beta$8 itself (Zwet et al., 2020).

The paper proves that even without selection, $\beta$9 is positively biased for $b\sim N(\beta,se^2)$0, because absolute value is convex. Under normality,

$b\sim N(\beta,se^2)$1

More importantly, under significance filtering,

$b\sim N(\beta,se^2)$2

is also positive for all $b\sim N(\beta,se^2)$3 and $b\sim N(\beta,se^2)$4, and this conditional bias is decreasing in $b\sim N(\beta,se^2)$5 and increasing in both $b\sim N(\beta,se^2)$6 and $b\sim N(\beta,se^2)$7 (Zwet et al., 2020).

A convenient exact representation is

$b\sim N(\beta,se^2)$8

where

$b\sim N(\beta,se^2)$9

This formula quantifies selective winner’s curse exactly in the normal-threshold model (Zwet et al., 2020).

The same idea appears in broader selection settings. In randomized experiments with two or more treatment arms, Berman, Zhang, and Zhao define the selective winner’s curse as

$|b|/se>c$0

where $|b|/se>c$1 is the empirically selected winner and $|b|/se>c$2 is the selected arm’s own true mean (Berman et al., 16 May 2026). This formulation isolates overstatement of the deployed winner’s performance, rather than comparison to the truly best arm.

2. Dependence on signal-to-noise ratio and power

A central theorem-level result is that selective winner’s curse is governed by the signal-to-noise ratio

$|b|/se>c$3

Van Zwet and Cator define the relative conditional bias

$|b|/se>c$4

and the exaggeration ratio

$|b|/se>c$5

They show that both depend on $|b|/se>c$6 and $|b|/se>c$7 only through the SNR, are decreasing in SNR, and increasing in the threshold $|b|/se>c$8 (Zwet et al., 2020).

Because power for the two-sided $|b|/se>c$9 test

$c=1.96$0

is a strictly increasing function of SNR, the same result can be restated as a power law: relative conditional bias decreases with power, and exaggeration decreases with power (Zwet et al., 2020). This formalizes the widespread observation that low-powered studies are the most severe environment for winner’s-curse inflation.

The same low-signal logic appears in several application domains. In online experimentation, hybrid-shrinkage work states explicitly that post-selection bias increases with sampling variability and the selection threshold and decreases as the true effect size grows (Mudd et al., 13 Mar 2026). In randomized experiments with selected treatment arms, Berman, Zhang, and Zhao state that winner’s-curse bias is largest in low signal-to-noise regimes, including small samples, large noise, near ties, and many-arm settings (Berman et al., 16 May 2026). In shared-validation model selection, the expected maximum of centered cumulative scores grows like $c=1.96$1, and the per-step selection premium is largest early, when competitors remain close (Pena et al., 23 Feb 2026).

This convergence across settings supports a general interpretation: selective winner’s curse is strongest when the selection rule admits winners primarily because noise can overturn small or uncertain underlying differences. When signals are large and ranking is stable, the same selection mechanism induces much less inflation.

3. Confidence-interval undercoverage after selection

Selective winner’s curse affects not only point estimates but also interval validity. In the normal-theory significance-filter setup, the usual $c=1.96$2 confidence interval

$c=1.96$3

has exact unconditional coverage

$c=1.96$4

After conditioning on significance at the same threshold, the relevant coverage becomes

$c=1.96$5

which need not equal $c=1.96$6 (Zwet et al., 2020).

Van Zwet and Cator prove that if the SNR satisfies

$c=1.96$7

then the usual interval undercovers after significance filtering. For a $c=1.96$8 interval, $c=1.96$9; the paper notes that when SNR equals this threshold, power is only slightly above $5\%$0, so the practical implication is that if a result is significant but power is about $5\%$1 or less, the ordinary interval undercovers (Zwet et al., 2020).

This undercoverage phenomenon reappears in field-level averaging. Under the additional assumptions that $5\%$2 has a decreasing density and is independent of $5\%$3, the paper proves that confidence intervals reported after significance filtering undercover on average across the field (Zwet et al., 2020). That result links selective winner’s curse to empirical settings with many small true signals.

Later work addresses interval construction more directly. The paper "Infer-and-widen, or not?" studies winner’s curse as a canonical selective-inference problem in the Gaussian means model

$5\%$4

with target $5\%$5. It argues that many existing methods fall into an infer-and-widen framework: they keep the selection-naive midpoint $5\%$6, which is biased upward after selection, and then widen symmetrically to recover coverage (Perry et al., 2024). The paper shows that in the winner’s-curse vignette, such intervals can be much wider than alternatives tuned to the same randomized selection event, and even an unattainable oracle infer-and-widen interval is often wider than simple direct selective methods (Perry et al., 2024).

In the Gaussian-randomized winner-selection setting

$5\%$7

the paper proves under equal means that

$5\%$8

showing explicitly that the naive midpoint remains biased even after randomization (Perry et al., 2024). The paper contrasts this with a selection-adjusted center based on

$5\%$9

which has conditional mean zero relative to the selected target (Perry et al., 2024).

A different line of work proposes confidence procedures that are valid for the winner without conditioning on a narrow parametric selection law. "A Flexible Defense Against the Winner’s Curse" treats the selected winner $|b|$0, where $|b|$1, by constructing a simultaneous confidence region and then projecting onto the selected coordinate (Zrnic et al., 2024). Its zoom correction adapts to the number of plausible competitors via suboptimality gaps

$|b|$2

and yields valid winner intervals

$|b|$3

while becoming close to uncorrected inference when the winner clearly stands out (Zrnic et al., 2024).

4. Dynamic, multiple, and procedure-level generalizations

Selective winner’s curse is no longer confined to a single scalar estimate. Several papers generalize it to dynamic, multiwinner, and procedure-level settings.

In shared-validation model selection, "A Selection Premium Decomposition for the Expected Maximum of Random Walks" considers $|b|$4 candidate models evaluated on the same validation sample. Let

$|b|$5

In the equal-mean null case, $|b|$6 is exactly the expected optimism created by ex post winner selection. The paper proves the decomposition

$|b|$7

where

$|b|$8

is the per-step selection premium (Pena et al., 23 Feb 2026). The structural properties of $|b|$9 show that the premium is nonnegative, vanishes in the one-model case, is translation invariant, and is maximized at ties (Pena et al., 23 Feb 2026). Thus selection-induced optimism is largest when competitors are close and fades as one model separates.

The same paper extends the decomposition to heterogeneous means via

$\frac{|b|}{se}>c,$0

yielding

$\frac{|b|}{se}>c,$1

This separates the winner’s observed score into a deserved component from the true best mean and an excess component due to continued noisy selection (Pena et al., 23 Feb 2026).

The paper also proves a universal bias concentration law: $\frac{|b|}{se}>c,$2 so the first $\frac{|b|}{se}>c,$3-fraction of observations generates asymptotically a $\frac{|b|}{se}>c,$4-fraction of total selection bias (Pena et al., 23 Feb 2026). This identifies a front-loaded temporal structure absent from static winner’s-curse formulations.

A separate generalization concerns multiple winners. "Inference on Multiple Winners" studies $\frac{|b|}{se}>c,$5 random selections

$\frac{|b|}{se}>c,$6

with simultaneous inference targets $\frac{|b|}{se}>c,$7. The paper emphasizes that this setting combines selective bias and multiple-inference distortion, making naive post-selection intervals invalid and projection methods overly conservative (Petrou-Zeniou et al., 2024). Its two-step method first models likely winners through lower bounds on population score gaps and then constructs simultaneous confidence intervals only over the likely-winner set. The resulting confidence set has formal simultaneous validity and, according to the paper, reduces over-coverage error by up to $\frac{|b|}{se}>c,$8 relative to projection methods (Petrou-Zeniou et al., 2024).

Adaptive LLM benchmarking supplies a procedure-level analogue. "Towards Reliable LLM Evaluation: Correcting the Winner’s Curse in Adaptive Benchmarking" argues that once benchmark items are reused inside prompt or program search, the benchmark score of the ex post winning artifact is no longer an estimate of the fresh-data performance of the full tune-then-deploy rule (Xu et al., 7 May 2026). The paper proposes SIREN, a repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap. Under a fixed-shortlist smooth-selector regime, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid (Xu et al., 7 May 2026). This reframes winner’s curse at the level of adaptive evaluation procedures rather than individual selected artifacts.

5. Applications across scientific and decision domains

The language of selective winner’s curse now appears across several specialized fields, but the underlying mechanism remains selection on noisy estimates followed by naive interpretation.

Online experimentation and A/B testing

In high-throughput experimentation systems, selected experiments are often those with sufficiently positive observed effects. "Breaking the Winner’s Curse with Bayesian Hybrid Shrinkage" models experiment $\frac{|b|}{se}>c,$9 through

$b$0

with selection or launch represented generically by

$b$1

The paper distinguishes unconditional unbiasedness of face-value estimates from post-selection bias after launch, emphasizes that low-power regimes are most vulnerable, and proposes an empirical-Bayes Bayesian Hybrid Shrinkage model with local experiment-specific shrinkage factors (Mudd et al., 13 Mar 2026). Under

$b$2

the posterior mean becomes

$b$3

when $b$4 (Mudd et al., 13 Mar 2026). The paper’s central practical claim is that this shrinkage reduces post-selection overestimation while remaining robust to prior misspecification through the local scale $b$5 (Mudd et al., 13 Mar 2026).

Randomized experiments and treatment-arm choice

In multi-arm randomized experiments, selective winner’s curse refers to bias relative to the deployed arm’s own mean. Berman, Zhang, and Zhao distinguish

$b$6

from

$b$7

with regret

$b$8

and identity

$b$9

This yields seven evaluation targets: bias, MSE, and interval coverage for both global and selective winner’s curse, plus mean regret (Berman et al., 16 May 2026). The paper concludes that no method dominates uniformly: plug-in performs best when treatment differences are large, cross-fitting performs best when treatments are similar, resampling methods often do well for MSE at moderate differences, and adaptive empirical likelihood yields asymptotically valid confidence intervals across settings (Berman et al., 16 May 2026).

Offline policy learning and data-driven decisions

"Beating the Winner’s Curse via Inference-Aware Policy Optimization" studies policy learning where counterfactual outcome models are first estimated and then optimized over to choose a policy. The paper argues that policies selected for high estimated reward may fail downstream significance tests because optimization exploits prediction error (Bastani et al., 20 Oct 2025). Its inference-aware policy optimization instead trades off expected gain

$\beta$0

against expected inferential strength

$\beta$1

where $\beta$2 is the IPW standard error. The Pareto frontier between value and inference is characterized explicitly, and the method selects policies that are more likely to survive downstream statistical validation (Bastani et al., 20 Oct 2025).

A more negative result appears in "Winner’s Curse Drives False Promises in Data-Driven Decisions: A Case Study in Refugee Matching." There, model-based policy evaluation is shown to produce substantial optimism even when models are accurate, stable, calibrated, trained on randomized data, correctly specified, and combined with sample splitting (Bastani et al., 9 Feb 2026). In a synthetic refugee-matching environment designed so that no policy can improve expected employment over random assignment, model-based methods report gains of around $\beta$3 even though the true effect is zero (Bastani et al., 9 Feb 2026). This supports the view that selective winner’s curse can persist despite several diagnostics often invoked as safeguards.

Mendelian randomization

Mendelian randomization has produced one of the most mathematically explicit literatures on winner’s curse as post-selection bias. In two-sample summary-data MR, SNPs are often selected for large observed exposure associations, and the same selected estimates are then reused in IVW or Egger estimation. This is selective winner’s curse in a particularly transparent form.

The paper "Breaking the Winner’s Curse in Mendelian Randomization: Rerandomized Inverse Variance Weighted Estimator" shows that standard selection by

$\beta$4

induces post-selection inflation in $\beta$5, which then biases standard IVW toward zero (Ma et al., 2023). Its solution is rerandomized selection using

$\beta$6

followed by a Rao–Blackwellized estimator

$\beta$7

which satisfies

$\beta$8

The resulting RIVW estimator is asymptotically normal and corrects both selective winner’s curse and measurement-error bias (Ma et al., 2023).

That logic is extended to MR-Egger in "Correction for Weak IV Bias and Winner’s Curse in Mendelian Randomization Egger Regression: Rerandomized Egger estimator," which distinguishes weak-IV attenuation from winner’s curse and constructs REgger by combining denominator debiasing with rerandomized selection and Rao–Blackwellization (Su et al., 13 Jul 2025). Under directional pleiotropy, REgger remains asymptotically normal and achieves better precision than IVW-family methods that do not model a nonzero average pleiotropic effect (Su et al., 13 Jul 2025).

A further extension incorporates sample structure. "Simultaneously accounting for winner’s curse and sample structure in Mendelian randomization: bivariate rerandomized inverse variance weighted estimator" notes that if

$\beta$9

then exposure-side selection propagates bias to the outcome side as well (Liu et al., 6 Mar 2026). The paper constructs a Rao–Blackwellized outcome estimator

$\beta$00

and proves

$\beta$01

This shows that selective winner’s curse can propagate through correlated statistics, turning a univariate correction problem into a genuinely bivariate one (Liu et al., 6 Mar 2026).

Adaptive LLM benchmarking

In LLM evaluation, adaptive prompt and program search induces winner’s curse because benchmark items are reused inside tuning. SIREN addresses this by freezing the post-search shortlist, splitting benchmark items repeatedly into scoring and evaluation subsets, and bootstrapping item-level influence contributions over a finite budget grid (Xu et al., 7 May 2026). Real tuning experiments on MMLU-Pro show that winner-based reporting is optimistic and can reverse deployment conclusions, whereas SIREN remains close to the finite-sample reporting target (Xu et al., 7 May 2026).

6. Conceptual variants, limits, and reversals

Selective winner’s curse is often presented as universally adverse, but recent work qualifies that claim.

First, the exact inferential target matters. In randomized experiments, selective bias relative to the selected arm’s true mean differs from bias relative to the truly best arm, and neither is equivalent to regret (Berman et al., 16 May 2026). Methods that are effective for one target can perform poorly for another, so “correcting the winner’s curse” is not a single-objective problem (Berman et al., 16 May 2026).

Second, the inferential framework matters. Van Zwet and Cator explicitly contrast frequentist and Bayesian perspectives. Frequentistically, $\beta$02 is unbiased before selection, but $\beta$03 is biased upward after significance filtering; Bayesianly, a shrinkage estimator

$\beta$04

is posterior-unbiased conditional on the observed data and mitigates winner’s curse (Zwet et al., 2020). In their phrase, shrinkage “lifts the winner’s curse” (Zwet et al., 2020). Later empirical-Bayes work adopts the same basic logic at scale, though typically without exact selective frequentist guarantees (Mudd et al., 13 Mar 2026).

Third, the sign of the informational effect can reverse in economic models. "The Winner’s Bliss in Common-Value Auctions under Horizontal Differentiation" studies a two-bidder first-price common-value auction with binary state $\beta$05, bidder types $\beta$06, and probability $\beta$07 (Chen et al., 7 Jun 2026). In that model,

$\beta$08

Hence winning is adverse when $\beta$09, but favorable when $\beta$10 (Chen et al., 7 Jun 2026). The paper calls the latter regime winner’s bliss. This does not negate selective winner’s curse in statistical inference; rather, it shows that “winning is bad news” is not a universal principle once the selection event also conveys favorable information about state-preference alignment (Chen et al., 7 Jun 2026).

Finally, most formal results remain model-specific. Common assumptions include Gaussian summary laws, known or consistently estimated standard errors, i.i.d. or independent increments, fixed shortlist sizes, finite candidate sets, and simple threshold or argmax selection rules. Several papers explicitly note that they do not solve arbitrary post-selection inference, arbitrary temporal dependence, or fully open-ended adaptive search (Pena et al., 23 Feb 2026, Xu et al., 7 May 2026, Mudd et al., 13 Mar 2026). This suggests that the theory of selective winner’s curse is mature in several canonical models but still fragmented across applications.

7. Practical implications

Across domains, several practical implications recur.

Significant estimates are often too large on average: conditioning on significance or empirical superiority systematically inflates reported effect sizes, model scores, or policy values (Zwet et al., 2020, Pena et al., 23 Feb 2026).

Low power and near ties are the most dangerous regimes: if power is low or alternatives are close, winners are disproportionately selected because of favorable noise realizations (Zwet et al., 2020, Berman et al., 16 May 2026).

Naive confidence intervals are unreliable after selection: nominal intervals often undercover unless they explicitly model selection, re-center appropriately, or use simultaneous/projection-based corrections (Zwet et al., 2020, Perry et al., 2024, Zrnic et al., 2024).

Separation of selection and evaluation helps but is not universally sufficient: cross-fitting, repeated splits, nested evaluation, and held-out testing reduce bias, but several papers show that sample splitting alone may not remove all optimism if structural model error remains aligned across stages (Berman et al., 16 May 2026, Bastani et al., 9 Feb 2026, Xu et al., 7 May 2026).

Shrinkage is often necessary, not merely optional: in both theoretical and production settings, shrinkage is presented as a way to counteract the bias that selection itself creates, rather than merely trading bias for variance (Zwet et al., 2020, Mudd et al., 13 Mar 2026).

Selection-aware reporting should target the operational procedure: in adaptive benchmarking and policy learning, the relevant object is frequently the performance of the tune-then-deploy rule or decision pipeline, not the same-data score of the ex post winner (Bastani et al., 20 Oct 2025, Xu et al., 7 May 2026).

The correct correction depends on the target: selective winner’s curse, global winner’s curse, regret, selected-parameter inference, population-maximum inference, and procedure-level validation are distinct tasks. A plausible implication is that no single correction paradigm can dominate across all of them (Berman et al., 16 May 2026, Petrou-Zeniou et al., 2024).

Selective winner’s curse is therefore best understood not as a single pathology with a single fix, but as a family of post-selection distortions produced by conditioning inference on success, extremeness, or empirical superiority. The unifying principle is straightforward: once a result is reported because it won, its naive estimate no longer inherits the unconditional guarantees that held before the competition.