Early Termination of Experiments in Nonparametric Predictive Comparisons

Coolen and van der Laan (2001) presented nonparametric predictive inference (NPI) for comparison of groups of units, aimed at selection of the ‘best group’, where ‘best’ relates to maximum value of a real-valued random quantity corresponding to each unit, e.g. its failure time. NPI is a statistical approach based on few assumptions in addition to data, made possible by the use of lower and upper probabilities to quantify uncertainty. We extend the results by Coolen and van der Laan by considering experiments that are terminated before all units have failed, and the effect of this on the inferences. It is shown that continuing an experiment, as long as not all units have failed, will never lead to increased imprecision, which will indeed decrease if further failures are observed.