Abstract
Bayesian bootstrapping (BBS, in short) in a broad sense, including nonparametric bootstrapping as a special case, is a versatile method for estimating the sampling distribution of an estimator T(Fn) of a population parameter T(F). Here, Fn is the empirical distributions based on observations of size n, F is the (unknown) population and T is a (known) functional with gentle characters typically written as a C-infinity function of moments defined in a neighborhood of the population moments. Assume that there exists an (unknown) functional TF such that T(F) = TF (U) for a (well known) distribution U, e.g., the uniform or normal distributions, then approximation to U by BBS from Un, i.e., a random sample of size n from U, should be our main concern, since T(Fn) and TF (Un) have the same sampling distribution. Algebraic and numerical works from this viewpoint have revealed that BBS with the Dirichlet prior Dir[n; c; :::; c] for c around 1/2 gives robust approximation to the target distribution. Numerical comparisons are also made with several resampling methods.
Author(s): Naoto Niki, Yoko Ono