A quadratic bootstrap method and improved estimation in logistic regression

Abstract

This paper presents a quadratic one-step bootstrap method for binary response data. Rather than resampling from the original sample, the proposed method resamples summands appearing in the quadratic approximation of the estimates. It enjoys the same computational simplicity as its linear analogue while being more accurate. Moreover it allows the construction of a bias corrected estimator and improved confidence intervals. A small simulation study illustrates the improved finite sample behaviour for binary response data.