A Simulation Study of the Performance of the Prescription Time-Distribution Matching Method to Address Immortal Time Bias

• Activity Number: 624
• Type: Contributed
• Date/Time: Thursday, August 2, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Mohammad Ehsanul Karim*+ and Paul Gustafson and John Petkau and Yinshan Zhao and Afsaneh Shirani and Charity Evans and Elaine Kingwell and Mia Van Der Kop 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: Observational study ; survival analysis ; multiple sclerosis ; propensity scores ; causal inference ; simulation

Abstract

Immortal time bias is a common problem in observational studies of drug-effectiveness. An efficient solution to this problem is time-dependent drug exposure modelling. However these models are complicated and not always intuitive since findings are expressed as person-time rather than person-level. Prescription time-distribution matching (Zhou et al., 2005) is an alternative approach where drug exposure is treated as time-independent. Here time to treatment initiation is listed for treated patients and each untreated patient is assigned a randomly selected baseline from this list to ensure that the distribution of the baseline time for untreated patients matches to the drug initiation of treated patients. We conduct a simulation study to justify this method and to compare its performance with the Cox proportional hazard model with treatment being the time-dependent exposure. We also apply this method to a dataset of relapsing-onset multiple sclerosis patients to estimate the long-term effectiveness of beta-interferons with a time-to-disability event outcome. We further extended the model by using propensity score weighting when covariate balance is not achieved at the new baseline.