Nonparametric Empirical Bayes Confidence Intervals

Abstract: Empirical Bayes methods can improve inference on unobservable individual effects by borrowing strength across units. This paper proposes nonparametric empirical Bayes confidence intervals (NP-EBCIs) for unobservable individual effects in a normal means model. The oracle intervals are constructed from posterior quantiles under a point-identified, fully nonparametric prior; feasible intervals replace these quantiles with nonparametric estimates. The NP-EBCIs are asymptotically exact in the sense that both their conditional and marginal coverage probabilities converge to the nominal level. The flexibility of this nonparametric construction has an unavoidable statistical cost. We demonstrate that posterior quantiles, unlike posterior means, inherit the severe ill-posedness of nonparametric deconvolution: the minimax optimal estimation rate is logarithmic. This logarithmic rate is minimax optimal for errors in the conditional coverage probability, and the resulting errors in the marginal coverage probability also vanish at the same logarithmic rate. Despite these slow asymptotic rates, simulations show that the NP-EBCIs remain close to nominal coverage when the prior is non-Gaussian, and deliver substantial length reductions relative to intervals that treat each unit in isolation.

• The paper introduces nonparametric empirical Bayes confidence intervals that directly use the full nonparametric prior instead of relying on Gaussian assumptions.
• It develops a rate-optimal kernel estimator that achieves logarithmic minimax rates for posterior quantile estimation in the face of severe ill-posedness.
• Empirical findings demonstrate that NP-EBCIs maintain nominal coverage and offer robust performance, outperforming traditional parametric and moment-based methods.

Nonparametric Empirical Bayes Confidence Intervals: Technical Summary

Introduction and Motivation

This paper develops a theory and methodology for constructing nonparametric empirical Bayes confidence intervals (NP-EBCIs) for unobservable individual effects in the normal means model. Departing from classical Cox-Morris-style EBCIs that require parametric (typically Gaussian) prior assumptions, or robust moment-based alternatives, the proposed approach directly exploits the nonparametric identification of the full prior, constructing intervals from posterior quantiles under a point-identified, entirely nonparametric prior $G$. Such NP-EBCIs are relevant in numerous applied settings—heterogeneous treatment effects, value-added modeling, and unobserved effects panel models—where accurate uncertainty quantification for latent unit-level parameters is critical.

Model and Interval Construction

The hierarchical setting is the normal means model with possible heteroskedasticity: for units $i=1,\ldots,n$, observed $Y_i \,|\, \theta_i \sim N(\theta_i, \sigma_i^2)$, with unobserved random effects $\theta_i \sim G$, where $G$ is unknown and potentially non-Gaussian. The target is an interval for each $\theta_i$ with prescribed conditional or marginal coverage.

Oracle Interval Construction

If $G$ is known, the oracle NP-EBCI is the equal-tailed posterior credible set:

$$\text{CI}_i^{NP*} = \big[ q_G(\alpha/2; D_i),\, q_G(1-\alpha/2; D_i) \big]$$

where $q_G(\tau; D_i)$ is the posterior $\tau$-quantile of $i=1,\ldots,n$0 under prior $i=1,\ldots,n$1 and observed data $i=1,\ldots,n$2.

Feasible NP-EBCI

Since $i=1,\ldots,n$3 is unknown, feasible NP-EBCIs plug in nonparametric estimators of the posterior quantiles, relying on nonparametric deconvolution techniques via the empirical characteristic function, kernel regularization, and leave-one-out resampling to mitigate the bias-variance tradeoffs endemic to ill-posed inverse problems.

Minimax Theory for Posterior Quantile Estimation

A central contribution is the derivation of a sharp minimax theory for posterior quantile estimation under nonparametric identification. Unlike the posterior mean—which admits nearly parametric rates for estimation in the normal means model—the quantile functional is non-smooth and inherits the severe ill-posedness of Gaussian deconvolution: the minimax optimal rate for estimating posterior quantiles, and thus attaining prescribed conditional/marginal coverage, is logarithmic in sample size. This substantial gap relative to the estimation of smooth functionals (e.g., posterior means) is formalized via lower bounds and the construction of a rate-optimal kernel estimator.

• Minimax Lower Bound: For Sobolev priors $i=1,\ldots,n$4 with smoothness $i=1,\ldots,n$5, the minimax risk for posterior quantile estimation is $i=1,\ldots,n$6.
• Attainability: A one-step direct kernel estimator for the quantile, regularized in the Fourier domain, attains this rate.

This sharp ill-posedness is a consequence of the non-smooth, locally discontinuous nature of the quantile functional, precluding the smoothing and bias-cancellation properties available for posterior means.

Coverage Properties: Conditional and Marginal

The paper establishes that feasible NP-EBCIs are asymptotically exact under both conditional (data-dependent) and marginal (population) coverage criteria, but, crucially, the convergence of coverage errors to nominal levels proceeds only at the logarithmic rate dictated by the minimax theory above.

• Conditional Coverage: The difference between actual and nominal conditional coverage probabilities of feasible NP-EBCIs vanishes at the logarithmic minimax rate but cannot be improved asymptotically by any data-driven construction.
• Marginal Coverage: Similarly, the marginal (empirical Bayes) coverage error converges at the same logarithmic rate.

This identifies an intrinsic efficiency-robustness trade-off:

• Parametric (e.g., Cox-Morris) EBCIs achieve parametric rates when the prior is correctly specified but can badly undercover under misspecification.
• Moment-based robust EBCIs guarantee coverage over restricted prior classes but may be overly conservative.
• NP-EBCIs, using the full nonparametric prior, achieve asymptotic exactness but face slow convergence due to the difficulty of deconvolution.

Monte Carlo Analysis

Empirical evidence via simulation demonstrates that in moderate and large samples, feasible NP-EBCIs:

• Maintain empirical marginal coverage close to the nominal level, even in non-Gaussian and discrete prior settings.
• Realize substantial interval length reductions relative to the naive unit-wise $i=1,\ldots,n$7-interval.
• Outperform moment-robust methods (e.g., AKP EBCI) in terms of interval efficiency and are far less prone to parametric undercoverage than Cox-Morris intervals.
• The price for nonparametric robustness manifests primarily as slightly wider intervals rather than substantial losses in coverage.

The tradeoff between efficiency (interval width) and robustness across prior shapes and signal-to-noise regimes is explicit in finite-sample results.

Extensions and Practical Implementation

The methodology extends to heteroskedastic and precision-dependent settings, with practical implementation including kernel selection and bandwidth tuning via out-of-fold nonparametric MLE estimation. The discrete/finite-sample structure of the NPMLE necessitates careful regularization to prevent undercoverage from overly concentrated posteriors.

Implications and Future Directions

Theoretical Implications

The results illuminate a fundamental dichotomy for empirical Bayes inference in mixture models:

• Functional Regularity is crucial: posterior means (smooth) allow for nearly parametric estimation; posterior quantiles (non-smooth) are subject to the full force of ill-posedness.
• When the inferential goal demands exact finite-sample/data-dependent coverage with minimal modeling assumptions, one must accept the slow convergence rates inherent to deconvolution.

Practical Implications

Applied researchers in empirical microeconomics, biostatistics, and related domains are now equipped to construct valid, nonparametric, shrinkage-based confidence sets for individual effects, without sacrificing nominal coverage. The nonparametric approach reduces reliance on unverifiable distributional assumptions at the cost of interval width and sample size requirements.

Future Research

Immediate extensions include:

• Non-Gaussian measurement error models (supersmooth and conditionally identified mixtures).
• Nonparametric empirical Bayes inference for other functionals—e.g., extreme/conditional tail probabilities, or quantile ranking in compound selection problems.
• Adaptive estimation and optimal regularization in misspecified or semi-parametric mixture contexts (partial identification).
• Robustification strategies (e.g., uncertainty quantification for estimation of $i=1,\ldots,n$8) that balance finite-sample conservatism and regularization-induced bias.

Conclusion

This work establishes a comprehensive minimax and implementation theory for nonparametric empirical Bayes confidence intervals in normal means models. NP-EBCIs offer an exact, robust framework for uncertainty quantification in heterogeneous effect estimation, with careful attention to the inherent efficiency loss imposed by statistical ill-posedness. These tools have immediate relevance for empirical practice and provoke new lines of questioning regarding the limits of nonparametric inference in high-dimensional latent variable models.

Reference: "Nonparametric Empirical Bayes Confidence Intervals" (2605.08551)