Percentiles from a set of classes or from the raw data? an argument based on the trinomial distribution

Abstract

In a former study (Chatillon, Gelinas, Martin and Laurencelle, 1987), the authors arrived at the conclusion that for small to moderate sample sizes (n≦90), and for population distributions that are not too skewed nor heavy tailed, the percentiles computed from a set of 9 classes are at least as precise as the corresponding percentiles computed with raw data. Their proof was based essentially on Monte Carlo simulations. The present paper gives a different and complementary proof, based on an exact evaluation of the mean squared error. The method of proof uses the trinomial distribution in an interesting way.