A1418
Title: Second-order accurate percentile inference with the implicit bootstrap
Authors: Samuel Orso - University of Geneva (Switzerland) [presenting]
Mucyo Karemera - University of Geneva (Switzerland)
Stephane Guerrier - University of Geneva (Switzerland)
Maria-Pia Victoria-Feser - University of Bologna (Italy)
Min-ge Xie - Rutgers University (United States)
Abstract: Many complex testing and inference problems arise in models involving latent variables, censoring, misclassification, missingness, dependence, or robustness considerations. In such settings, obtaining a consistent estimator may require analytically involved corrections or computationally intensive procedures, and subsequent bootstrap inference can become even more demanding. The implicit bootstrap is a new percentile-type method designed to deliver accurate inference without requiring consistency of the underlying statistic. The method combines computationally simple auxiliary statistics with simulations from a chosen data-generating mechanism to implicitly recover the parameter information needed for inference. Asymptotically valid confidence intervals with second-order accuracy are established under appropriate conditions, matching some of the higher-order advantages of studentized bootstrap methods while retaining the invariance properties of percentile intervals. Numerical experiments illustrate the practical advantages of the method for complex models, especially when consistent estimators are difficult to derive, computationally costly, or unattractive for routine bootstrap inference.
