SENSITIVITY ANALYSIS FOR INFORMATIVE CENSORING

In Chapters 2 and 6, we discussed the relatively intractable problem of informative censoring. Let's quickly review the problem. Suppose that just before some particular time t, there are 50 individuals who are still at risk of an event. Of those 50 individuals, 5 are censored at time t. Suppose further that 20 of the 50 at risk have covariate values that are identical to those of the 5 who are censored. We say that censoring is informative if the 5 who are censored are a biased subsample of the 20 individuals with the same covariate values. That is, they have hazards that are systematically higher or lower than those who were not censored. Informative censoring can lead to parameter estimates that are seriously biased.

When censoring is random (that is, not under the control of the investigator), it's usually not difficult to imagine scenarios that would lead to informative censoring. Suppose, for example, that you're studying how long it takes rookie policemen to be promoted, and those who quit are treated as censored. It doesn't take much insight to suspect that those who quit before promotion have, on average, poorer prospects for promotion than those who stay. Unfortunately, there's no way to test this hypothesis. You can compare the performance records and personal characteristics of those who quit and those who stayed, but these are all things that would probably be included as covariates in your model. Remember that what we're concerned about is residual informativeness, after the effects of covariates have been taken into account. And even if we could discriminate between informative and noninformative censoring, there are no standard methods for handling informative censoring. The best that can be done by way of correction is to include as covariates any factors that are believed to affect both event times and censoring times.