Robust and Efficient Empirical Bayes Confidence Intervals using $γ$-Divergence

Although parametric empirical Bayes confidence intervals of multiple normal means are fundamental tools for compound decision problems, their performance can be sensitive to the misspecification of the parametric prior distribution (typically normal distribution), especially when some strong signals are included. We suggest a simple modification of the standard confidence intervals such that the proposed interval is robust against misspecification of the prior distribution. Our main idea is using well-known Tweedie's formula with robust likelihood based on $\gamma$-divergence. An advantage of the new interval is that the interval lengths are always smaller than or equal to those of the parametric empirical Bayes confidence interval so that the new interval is efficient and robust. We prove asymptotic validity that the coverage probability of the proposed confidence intervals attain a nominal level even when the true underlying distribution of signals is contaminated, and the coverage accuracy is less sensitive to the contamination ratio. The numerical performance of the proposed method is demonstrated through simulation experiments and a real data application.