Statistical analysis of heaped duration data

This paper shows how heaping of duration data, e.g. caused by rounding due to memory effects, can be analyzed. If the data are heaped Cox's partial likelihood approach, which is often used in survival analysis, is no longer appropriate. We show how this problem can be overcome by considering the problem as a missing data problem. A variant of Cox's Proportional Hazard Model is constructed that takes heaping into account, and is estimated by maximum likelihood using the EM algorithm, with many nuisance parameters, simultaneously for all parameters. Ingredients of our method are application of the EM algorithm, Cox regression and nonparametric maximum likelihood calculation with `predicted' data in each M step. An example from practice, where jackknife is used to estimate the variances, illustrates the power of the new methodology.