Inverse-probability-of-censoring weights — morie_censoring_at_risk

\(w_i = 1/\hat G(t_i)\) for the uncensored, where G is the Kaplan-Meier estimate of the CENSORING distribution. Each surviving subject stands in for the ones like them who were censored, which is what restores the original population.

Usage

morie_censoring_at_risk_weight(time, censor, at = NULL, stabilize = TRUE)

Arguments

time: follow-up times.
censor: 1 if censored, 0 if the event was observed.
at: optional single time at which to evaluate G, instead of each subject's own time.
stabilize: multiply by the mean of G.

Value

list with weights, G, ess, max_weight_share.

Details

The weights explode as the censoring survivor approaches zero at long follow-up, so a handful of late survivors can end up carrying the estimate. The fix is to truncate the time axis at a point where G is still comfortably positive, not to weight through administrative censoring and hope.

References

Robins, J. M. and Rotnitzky, A. (1992). Recovery of information and adjustment for dependent censoring using surrogate markers. In AIDS Epidemiology, 297-331. Birkhauser.

Examples

set.seed(1)
tt <- rexp(100); cc <- rbinom(100, 1, 0.3)
round(morie_censoring_at_risk_weight(tt, cc)$ess, 1)
#> [1] 68.3