Improved reference-interval estimation

Abstract

We used two standard methods (the common percentile and the log-power gaussian transformation) and two novel methods (weighted percentile and smoothed spline interpolation) to estimate the 2.5th and 97.5th percentiles of 1000 sets of data from eight diverse distributional forms, generated by Monte Carlo simulation. For each distributional form we derived an estimate of optimal performance. Although none of the four proposed methods closely approximated the optimal performance bound, the spline interpolation method and the weighted percentile method were superior to the two standard methods in accurately estimating percentiles.