Marginal Structural Models for Investigating Long-Term Effectiveness of a Time-Dependent Treatment: An Application to Multiple Sclerosis

• Activity Number: 87
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Mohammad Ehsanul Karim*+ and Paul Gustafson and John Petkau and Afsaneh Shirani and Mia van der Kop and Elaine Kingwell and Yinshan Zhao and Charity Evans and Helen Tremlett
• Companies: University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia
• Keywords: causal inference ; observational study ; marginal structural models ; inverse probability weighting ; sensitivity analysis ; multiple sclerosis

Abstract

Beta-interferons (INFB) are often used to treat patients with relapsing-onset multiple sclerosis (MS). Access to the British Columbia-wide MS Database, one of the largest longitudinal cohorts of MS patients, has given us the opportunity to investigate the long term effectiveness of INFB with a time-to-event outcome. The first date a patient became eligible for INFB treatment was considered as the baseline, with relevant patient characteristics included as baseline covariates. Due to the lack of randomization, it was necessary to control for selection bias at baseline. Another feature of this study which introduced additional complexities into the model was that patients could drop out and be censored. Also, it is plausible that a patient's motivation to start or stop an INFB was influenced by time-varying confounders, such as frequency of relapses. In this complicated scenario, under certain assumptions, marginal structural models (MSM) offer a way to assess the causal effect of INFB, by estimating weights as a first step and then utilizing them in a time dependent Cox's model. A sensitivity analysis was also performed, to evaluate the suitability of the assumptions for causality.