The Impact of Analysis Method and Model Specification for Handling Missing Covariate Data in Survival Analysis: a Case Study

• Activity Number: 534 - Contributed Poster Presentations: Section on Statistics in Epidemiology
• Type: Contributed
• Date/Time: Wednesday, August 1, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Evon Okidi* and Joseph W Hogan and Chanelle Howe
• Companies: Brown University and Brown University School of Public Health and Brown University
• Keywords: Missing data; Inverse Probability weighting; Multiple imputation; HIV/AIDS; Hypertension

Abstract

Missing data due to non-response are commonly handled using inverse probability weighting (IPW), which relies on the missing at random assumption and a correctly specified non-response model. We compare two IPW techniques for handling missing data when estimating the association between blood pressure and mortality among HIV+ people without AIDS. About 30% of data used to define AIDS were missing. Using 77,280 patient records from western Kenya, a Cox proportional hazards model was used to estimate the mortality hazard rate (HR) for systolic and diastolic blood pressure (SBP, DBP) categories. We compared two IPW methods to complete case analysis. Stabilized weights were estimated using logistic regression and regression tree (RT) models. We did multiple imputation (MI) of missing data using multinomial logistic regression models. Low SBP and DBP were associated with higher hazard of mortality for all approaches. Mortality HRs (95% Confidence Limit [CL]) for DBP< 60mmHg versus ?60mmHg using IPW via logistic and regression tree techniques were 3.19 (2.56, 3.97) and 3.32 (2.85, 3.82). IPW based on RT had the narrowest CLs likely because we made no assumptions about functional forms.