Bias and Its Remedy in Interval-Censored Time-to-Event Applications - Interval-Censored-Time-to-Event Data: Methods and Applications - page 318

is almost unbiased, \Cox.Right" is biased low by about 20% and the \Cox.Mid" is
biased low by about 10%. For the case of less frequent visits, \IntCox" can be biased
low by about 10% for higher censoring where the \Cox.Right" could be biased low
by about 50%. Generally, the bias increases as the probability of censoring increases.
FIGURE 11.1: Biases for all parameter estimates
This increasing pattern of bias is more distinctive for variance and MSE. Figure
11.2 illustrates this conclusion for mean squared error (MSE), and a similar figure
can be generated for variance. This figure is generated for the case of more frequent
visits (i.e., Month=1); MSE for \Month=3" is even larger. Obviously, MSE increases
rapidly when the probability of censoring increases.