Bayesian Marginal Structural Models for Analysis of Medication Use Among Critically Ill Older Patients

• Activity Number: 452
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Terrence Murphy*+ and Katy Araujo and Margaret A. Pisani
• Companies: and Yale and Yale School of Medicine
• Keywords: time dependent confounding ; marginal structural model ; Bayesian ; delirium ; haloperidol

Abstract

The observational analysis of medication use among older patients in a medical intensive care unit is complicated by clinical and statistical issues. Clinical issues include the concurrent use of multiple medications and imminent risk of death whereas statistical concerns include the potential for time dependent confounding between doses and symptoms. We emphasize the advantage of employing a marginal structural model to adjust for time dependent confounding and demonstrate additional benefits of doing this within the Bayesian context. In order to more closely approach causal inference, we establish temporal precedence by modeling the diagnosis of delirium on a given day in the ICU on the dose of the antipsychotic haloperidol administered the previous day, with weighting for time varying intubation and doses of the opioid fentanyl. Although weight models are frequentist, the high level of missing data in ICU data assumed to be missing at random complements their application within the Bayesian context. We contrast model results from weighted and unweighted models within both frequentist and Bayesian frameworks.