[论文解读] Performance of Efron and Tibshirani's semiparametric density estimator

Nils Lid Hjort|ArXiv.org|Mar 16, 2026

Statistical Methods and Inference参考文献 12被引用 3

一句话总结

本论文推导了 Efron 与 Tibshirani 的半参数密度估计量的偏差和方差（因此也有均方误差 MSE），将其与核方法及其他半参数方法进行比较，并在不同设定下评估性能。

ABSTRACT

Recently, Efron and Tibshirani (Annals of Statistics, 1996) proposed a semiparametric density estimator, which works by multiplying an initial kernel type estimate with a parametric exponential type correction factor, chosen so as to match certain empirical moments. While Efron and Tibshirani investigate and illustrate many aspects of their method, the basic questions of performance, and comparison with other density estimators, were not directly addressed in their article. The purpose of the present paper is to provide formulae for bias and variance and hence mean squared error for the estimator. This additional insight into the method makes it easy to compare its performance with that of other recently proposed semiparametric constructions. A brief comparison study is carried out here. It indicates that the new method, used with lower order polynomials in the exponential correction term, is often better than the kernel estimator, in a reasonable neighbourhood around the normal distribution, but that its performance as a density estimator is more than equalled by other methods. In particular, the recently developed Hjort and Glad estimator (Annals of Statistics, 1995), using a parametric start times a nonparametric correction, wins in eight out of nine test cases, from the list of such suggested by Wand and Jones (Annals of Statistics, 1992).

研究动机与目标

• Motivate and evaluate the performance of the semiparametric density estimator proposed by Efron and Tibshirani (1996).
• Provide bias and variance formulas to enable comparisons with kernel and alternative semiparametric estimators.
• Illustrate the estimator's performance in relation to normal-centered densities and other competing methods.
• Discuss practical implications for bandwidth choice and estimator selection in density estimation.

提出的方法

• Formalize the semiparametric density estimator as f(x,β) = f̂0(x) ĉ(β)−1 exp{βᵗ t(x)} and derive its normalizing constant ĉ(β).
• Derive a central result for E[f̂(x)] and Var[f̂(x)] under h→0 and nh→∞, showing E[f̂(x)] = f(x) + (1/2)k2 h^2{f''(x) − f(x)g(x)} + o(h^2) + O(h^2 n^−1) and Var[f̂(x)] = (nh)^−1 R(K) f(x) − n^−1 f(x)^2 + O(h n^−1).
• Define g(x) in terms of E t''(X) and Σ^−1, and show how it governs the bias modification.
• Compare the ET estimator with low-order polynomial t(x) choices to other estimators (f̂2, f̂3, f̂4, f̂5, and kernel) via a bias-based performance metric.
• Provide a practical discussion of bandwidth selection and the impact of choosing different t(x) functions (e.g., x, (x,x^2), (x,x^2,x^3), (x,log x)).
• Outline the limit behavior of β̂ and its impact on the estimator’s bias correction and variance.

实验结果

研究问题

• RQ1 Does Efron and Tibshirani’s semiparametric density estimator reduce bias without inflating variance relative to the kernel estimator?
• RQ2 How do the bias and variance (and MSE) of the estimator behave as h→0 and nh→∞?
• RQ3 In which settings and for which choices of t(x) does the estimator outperform kernel methods or other semiparametric rivals?
• RQ4 How does the performance vary relative to proximity to normality and to higher-order polynomial corrections in t(x)?

主要发现

• The estimator’s variance is unchanged at the considered order relative to the kernel estimator, while the bias is modified to include a term involving g(x).
• For low-order polynomials, the bias term vanishes at normality, predicting good performance near normal densities but not uniform improvement everywhere.
• The method often improves on the kernel estimator in neighborhoods around normal densities but is outperformed by some competitors (notably f̂3 by Hjort and Glad) in many non-normal cases.
• Increasing the order of the polynomial in t(x) does not always improve performance; sometimes lower-order options are better due to estimator variance.
• Overall, the ET estimator can outperform the kernel method in some cases but is not universally superior; its gains are modest when the true density deviates substantially from normal.
• Among competitors, simple variance correction (Jones, 1991) and Hjort–Glad’s multiplicative method frequently perform very well across several test densities.