Censoring and truncation are two main sources of incomplete data. For such cases, the actual values of a measurement x are observed if x satisfies a certain condition such as x≤c, where c is a pre-specified cut-off point. For censoring, the number of unobserved values is known, whereas such number is not reported for truncation. Comparisons of the accuracy of estimates of population parameter θ under censoring and truncation have been performed by calculating asymptotic variances or by inspection of the figures of log-likelihood functions.
In this paper, a different approach is proposed in assessing the stability of the estimates under censoring and truncation. We perturb the cut-off point c and examine the effect of the perturbation that influences the estimates. Specifically, we examine implicit functions θ=g(c) given by likelihood equations. It is observed that the function g(c) is increasing in c under censoring, but is decreasing under truncation. It is also observed that for censoring, the effect of c becomes small if the number of observed data, m, is large, but that for truncation, the number m does not contribute to make estimates stable. Two important cases, exponential and normal distributions will be examined in detail. Two numerical examples are also given to illustrate the situation. One of the inferences obtained from this paper is that the estimate under truncation is quite unstable even for large values of m.
View full abstract