Issues in Methodological Strategies for Marginal Structural Models with Large Data Sets

• Activity Number: 425
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Bret Zeldow* and Jason Roy
• Companies: and University of Pennsylvania
• Keywords: marginal structural models ; probability weighting ; truncation ; big data ; comparative effectiveness

Abstract

Parameters in marginal structural models are estimated using inverse-probability-of-treatment weighted estimators. Common ways to estimate the probability of treatment include logistic regression models or machine learning. In longitudinal datasets with long follow-up time, weights for the select individuals can grow to unwieldy levels if the probability of the individual's observed treatment trajectory is small, even when using stabilized weights. In practice, many analyses deal with this issue by choosing a truncation point beyond which all weights are held constant. In this work, we examine estimation of parameters from marginal structural models in large datasets, focusing on sensitivity due to various methodological strategies such as the choice and flexibility of treatment and censoring models, model validation, and weight truncation. Our motivating example is a dataset of HIV/hepatitis C co-infected individuals where we look at the risk of liver decomposition from the use of select antiretroviral agents. The study included over 4900 patients who were followed up to 170 months.