Multiple imputation methods for inference on cumulative incidence with missing cause of failure

• Published in: Biometrical journal Vol. 53; no. 6; pp. 974 - 993
• Main Authors: Lee, Minjung, Cronin, Kathleen A., Gail, Mitchell H., Dignam, James J., Feuer, Eric J.
• Format: Journal Article
• Language: English
• Published: Weinheim WILEY-VCH Verlag 01.11.2011; WILEY‐VCH Verlag; Wiley-VCH; Wiley - VCH Verlag GmbH & Co. KGaA
• Subjects: Applications; Asymptotic methods; Asymptotic properties; Biology, psychology, social sciences; Biometry - methods; Breast cancer; Breast Neoplasms - drug therapy; Breast Neoplasms - epidemiology; Cause-specific hazard function; Clinical Trials as Topic; Competing risks; Counting process; Exact sciences and technology; Failure analysis; Female; General topics; Hazards; Humans; Inference from stochastic processes; time series analysis; Intestine; Mathematics; Middle Aged; Missing at random; Probability and statistics; Probability theory and stochastic processes; Proportional hazards model; Risk; Sciences and techniques of general use; Statistical models; Statistics; Stochastic processes; Survival Analysis; Biometrics; Conditional distribution; Breast disease; Error estimation; Missing at random; Prediction theory; Competing risks; Covariate; Stochastic process; Statistical simulation; Survival data; Proportional hazards model; Hazard function; Competing risk; Life science; Prediction; Statistical estimation; Breast cancer; Statistical method; Cause-specific hazard function; Counting process; Observation data; Filtering theory; Asymptotic approximation; Multiple imputation
• ISSN: 0323-3847; 1521-4036; 1521-4036
• DOI: 10.1002/bimj.201000175

Analysis of cumulative incidence (sometimes called absolute risk or crude risk) can be difficult if the cause of failure is missing for some subjects. Assuming missingness is random conditional on the observed data, we develop asymptotic theory for multiple imputation methods to estimate cumulative incidence. Covariates affect cause‐specific hazards in our model, and we assume that separate proportional hazards models hold for each cause‐specific hazard. Simulation studies show that procedures based on asymptotic theory have near nominal operating characteristics in cohorts of 200 and 400 subjects, both for cumulative incidence and for prediction error. The methods are illustrated with data on survival after breast cancer, obtained from the National Surgical Adjuvant Breast and Bowel Project (NSABP).