Random Survival Forests for Interval-Censored Outcomes in the Presence of Imperfect Diagnostic Tests

• Activity Number: 457
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Hui Xu*+ and Xiangdong Gu and Raji Balasubramanian
• Companies: and and University of Massachusetts, Amherst
• Keywords: random survival forests ; Survival analysis ; High dimensional data ; self-report

Abstract

In many epidemiologic settings,, such as the Women's Health Initiative (WHI), the occurrence of a silent event such as diabetes is ascertained through error-prone procedures such as self-reports. For this setting, we propose a modification to the Random Survival Forests (Ishwaran, H., 2008) (RSF) algorithm by incorporating a likelihood function that accounts for the error-prone nature of the diagnostic tests. To evaluate the performance of our proposed algorithm, we simulated datasets of 100 subjects and 100 features per subject, of which 5 were assumed to be true biomarkers that influence the risk of the event of interest through a proportional hazards model. The parameter settings were selected to mimic diabetes self-reported outcomes in the WHI. Averaging across 100 simulated datasets, the proportion of times the true biomarkers were ranked among the top 5 features was 0.658 and 0.718, by RSF and our approach, respectively. The proposed algorithm will be applied to GWAS data from the WHI (approximately 12,000 subjects), where the outcome of interest is diabetes that is ascertained through self-report.