Slope estimation in the presence of informative right censoring: modeling the number of observations as a geometric random variable.

A method is proposed for the estimation of rate of change from incomplete longitudinal data where the number of observations made for each subject is assumed to vary depending on the level of the response variable. The proposed method involves a random slope model, in which the number of observations is modeled as a geometric distribution with its mean dependent on the individual subject's rate of change. The method adjusts for informative right censoring and provides estimates of the slopes of individual subjects as well as of the population. Under noninformative right censoring these estimators of the slopes are equivalent to Bayes estimators (Fearn, 1975, Biometrika 62, 89-100). The simulation study demonstrates that, in cases where the censoring process is informative, the proposed estimator is more efficient than either the unweighted or weighted estimator of slope. The method is illustrated by the analysis of renal transplant data.