The effect of incomplete follow-up.

By generating sampling distributions for each of three commonly used estimators for qz, the conditional probability of death in an interval, it is shown that the inclusion of subjects due for withdrawal introduces bias in all three, especially the "actuarial estimator" (Berkson and Gage [1950]). The bias is shown to depend on the proportion due for withdrawal and on the level of qz. Some practical suggestions are made for dealing with the bias.