Limitation of inverse probability-of-censoring weights in estimating survival in the presence of strong selection bias

Abstract

In time-to-event analyses, artificial censoring with correction for induced selection bias using inverse probability-of-censoring weights can be used to 1) examine the natural history of a disease after effective interventions are widely available, 2) correct bias due to noncompliance with fixed or dynamic treatment regimens, and 3) estimate survival in the presence of competing risks. Artificial censoring entails censoring participants when they meet a predefined study criterion, such as exposure to an intervention, failure to comply, or the occurrence of a competing outcome. Inverse probability-of-censoring weights use measured common predictors of the artificial censoring mechanism and the outcome of interest to determine what the survival experience of the artificially censored participants would be had they never been exposed to the intervention, complied with their treatment regimen, or not developed the competing outcome. Even if all common predictors are appropriately measured and taken into account, in the context of small sample size and strong selection bias, inverse probability-of-censoring weights could fail because of violations in assumptions necessary to correct selection bias. The authors used an example from the Multicenter AIDS Cohort Study, 1984-2008, regarding estimation of long-term acquired immunodeficiency syndrome-free survival to demonstrate the impact of violations in necessary assumptions. Approaches to improve correction methods are discussed.

Figure 1.

A) Taxonomy for follow-up in a cohort with artificial censoring. B) Taxonomy used to categorize data from the Multicenter AIDS Cohort Study, 1984–2008, when follow-up was administratively censored at 1996, as done in methods 1 and 2. C) Taxonomy used to categorize data from the Multicenter AIDS Cohort Study, 1984–2008, when follow-up was artificially censored at initiation of highly active antiretroviral therapy, as done in methods 3, 4, and 5. The numbers of artificially censored participants who were and were not observed to develop the event after meeting the defined criterion appear below the artificially censored category. NA, not applicable.

Figure 2.

A) Log cumulative hazard of acquired immunodeficiency syndrome (AIDS) among 467 seroconverters in the Multicenter AIDS Cohort Study, 1984–2008, based on the standard Kaplan-Meier (administrative KM) estimator and the generalized gamma (GG) distribution when the date of analysis was 1996. B) Log cumulative hazard for AIDS among 467 seroconverters the Multicenter AIDS Cohort Study, 1984–2008, based on the standard KM estimator (artificial KM), inverse probability-of-censoring weights (IPCW), and IPCW GG when follow-up was censored at highly active antiretroviral therapy initiation compared with the GG model when the date of analysis was 1996.