Inverse probability‐of‐censoring weights for the correction of time‐varying noncompliance in the effect of randomized highly active antiretroviral therapy on incident AIDS or death
- 3 April 2009
- journal article
- research article
- Published by Wiley in Statistics in Medicine
- Vol. 28 (12) , 1725-1738
- https://doi.org/10.1002/sim.3585
Abstract
In 1996–1997, the AIDS Clinical Trial Group 320 study randomized 1156 HIV-infected U.S. patients to combination antiretroviral therapy (ART) or highly active ART with equal probability. Ninety-six patients incurred AIDS or died, 51 (4 per cent) dropped out, and 290 (=51 + 239, 25 per cent) dropped out or stopped their assigned therapy for reasons other than toxicity during a 52-week follow-up. Such noncompliance likely results in null-biased estimates of intent-to-treat hazard ratios (HR) of AIDS or death comparing highly active ART with combination ART, which were 0.75 (95 per cent confidence limits [CL]: 0.43, 1.31), 0.30 (95 per cent CL: 0.15, 0.60), and 0.51 (95 per cent CL: 0.33, 0.77) for follow-up within 15 weeks, beyond 15 weeks, and overall, respectively. Noncompliance correction using Robins and Finkelstein's (RF) inverse probability-of-censoring weights (where participants are censored at dropout or when first noncompliant) yielded estimated HR of 0.46 (95 per cent CL: 0.25, 0.85), 0.43 (95 per cent CL: 0.19, 0.96), and 0.45 (95 per cent CL: 0.27, 0.74) for follow-up within 15 weeks, beyond 15 weeks, and overall, respectively. Weights were estimated conditional on measured age, sex, race, ethnicity, prior Zidovudine use, randomized arm, baseline and time-varying CD4 cell count, and time-varying HIV-related symptoms. Noncompliance corrected results were 63 and 13 per cent farther from the null value of one than intent-to-treat results within 15 weeks and overall, respectively, and resolve the apparent non-proportionality in intent-to-treat results. Inverse probability-of-censoring weighted methods could help to resolve discrepancies between compliant and noncompliant randomized evidence, as well as between randomized and observational evidence, in a variety of biomedical fields. Copyright © 2009 John Wiley & Sons, Ltd.Keywords
Funding Information
- NIH (NIAID R03-AI 071763)
This publication has 30 references indexed in Scilit:
- Estrogen Therapy and Coronary-Artery CalcificationNew England Journal of Medicine, 2007
- Comparison of Dynamic Treatment Regimes via Inverse Probability WeightingBasic & Clinical Pharmacology & Toxicology, 2006
- Low-Fat Dietary Pattern and Risk of Invasive Breast CancerJAMA, 2006
- Effect of acyclovir on herpetic ocular recurrence using a structural nested modelContemporary Clinical Trials, 2005
- Sense and sensitivity when correcting for observed exposures in randomized clinical trialsStatistics in Medicine, 2004
- Determinants of survival following HIV-1 seroconversion after the introduction of HAARTThe Lancet, 2003
- Administrative and artificial censoring in censored regression modelsStatistics in Medicine, 2001
- Marginal Structural Models to Estimate the Joint Causal Effect of Nonrandomized TreatmentsJournal of the American Statistical Association, 2001
- Relation of pooled logistic regression to time dependent cox regression analysis: The framingham heart studyStatistics in Medicine, 1990
- Statistics and Causal InferenceJournal of the American Statistical Association, 1986