Considerations in Applying Marginal Structural Models to Analyze Longitudinal Naturalistic Data

• Activity Number: 263
• Type: Topic Contributed
• Date/Time: Tuesday, August 4, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Section on Health Policy Statistics
• Author(s): Ouhong Wang*+
• Companies: Amgen, Inc.
• Keywords: Time Dependent Confounding ; Marginal Structural Models ; Observational Study ; Naturalistic Data

Abstract

While clinical trials in a controlled environment provide the basis for regulatory approval, large-scale long-term observational databases are usually the means for long-term safety assessment. The choice of statistical methods is critical in dealing with such databases as the data almost without exception suffer from confounding and sometimes even time-dependent confounding. In this talk the marginal structural models (MSM) approach using inverse probability of treatment weights is illustrated in analyzing safety information using such a database. The association between treatments and safety signals is less confounded and, under the assumption of no model misspecification, may even have a causal interpretation. Important considerations such as the ETA assumption violation, model selection, and weight truncation are discussed.