| The likelihood ratio statistic (l.r.s.) is a commonly used element of parametric statistical inference; its distribution in large samples is, under the null hypothesis, approximately chi-squared. Bartlett adjustments are corrections for the l.r.s. to be used with small samples to yield statistics whose asymptotic properties more nearly match those of a chi-squared random variable. Most of what has been written deals with theoretical aspects of Bartlett adjustments such as their derivation and the asymptotic properties of the corrected statistic.;For censored survival data models, while the usual asymptotic results for the maximum likelihood estimate, the score function and the l.r.s. do not depend on the specific censoring mechanism beyond some mild restrictions, the form of the Bartlett adjustments does.;We aim to look more practically at adjustments for parametric models for censored survival data by studying their magnitude under a variety of circumstances and by observing their effect in practice. We do this for single sample and two sample exponential and Weibull models. A correction based on the Kaplan-Meier estimate of the censoring time distribution is proposed; this correction can be calculated without assuming a specific censoring mechanism. A computer program written for Mathematica is presented which calculates Bartlett adjustments using Lawley's method. |
