Analytically calibrated Box–Cox percentile limits for duration and event-time models

This paper proposes a unified approach to constructing confidence limits for a future percentile duration or event-time. The construction is based on an analytical calibration of the Box–Cox-type “plug-in” percentile limits (PL). The performance of the calibrated Box–Cox PL is investigated using Monte Carlo experiments. Comparisons are made with PLs that are specifically designed for a particular distribution such as Weibull and lognormal. Excellent performances of the calibrated Box–Cox PL are observed. Simulation based on other popular duration models such as gamma and inverse Gaussian reveal that the proposed PL is robust against distributional assumptions and that it performs much better than the distribution-free PL. An empirical illustration is also provided.