Diagnosing bootstrap success (Q1370526)

The bootstrap method, proposed by \textit{B. Efron} [Ann. Stat. 7, 1-26 (1979; Zbl 0406.62024)], is now being used in many different areas. Even though it was thought at the beginning that the bootstrap would be a panacea in statistical data analysis, many examples have been found later which showed the failure of the bootstrap method. That is, the bootstrap distribution fails to converge to the distribution of the original statistic. Two such examples are presented in this paper. Some more examples are found by \textit{J. Shao} and \textit{D. Tu} [The jackknife and bootstrap. (1995)]. In this paper, for estimators based on a sample of \(n\) iid random elements, the situations where bootstrapping (parametric or nonparametric) would fail to converge correctly are characterized theoretically. Based on these theoretic results, a graphical procedure of diagnosis is proposed to detect the failure of the bootstrap. When the original bootstrap is found not to converge correctly, two approaches are also suggested to modify the original bootstrap such that the modified bootstrap distributions would converge correctly.