Statistical Analysis of Dependent Current Status Data with Application to Tumorigenicity Experiments - Interval-Censored-Time-to-Event Data: Methods and Applications - page 125

and death without tumor are not significant, but that high-dose treatment

has a very significant effect on the death of subjects with tumor.
Additionally, Figures 5.2 and 5.3 correspondingly display the estimates of

1 ; 2 ; 3 ; and obtained in each iteration of the EM algorithm. These show
that the convergence rates of the regression parameters are faster than that

of , irrespective as to the choice of their initial values.

5.6 Discussion

This chapter outlined approaches to analyze current status data with de-

pendent censoring. A three-state model was applied and the correlation is

incorporated with frailty effects. Two distributions, normal and gamma dis-

tribution, were considered and have similar results. There are still open issues

for the suggested method. One issue is to extend bivariate current status data

with informative censoring. A more complicated multi-state model and a bi-

variate frailty were applied. We need to make cautious remarks related to the

suggested algorithm. As often occurred for the EM algorithm, inappropriate

initial values make a convergence rate too slow. In particular, small initial val-

ues of baseline hazards cause trouble in the calculation of the inverse matrix.