CE2 Effect of Different Variance Estimation Methods with Inverse Probability Treatment Weights (IPTW) on Comparative Effectiveness Measure in Multiple Sclerosis


      The Inverse Probability Treatment Weighting (IPTW) method provides marginal treatment effects that are more generalizable than other propensity score (PS) methods. The objective of this study was to compare the treatment effects from three different variance estimation methods with stabilized IPTWs on comparative effectiveness between oral fingolimod and injectable Disease Modifying Agents (DMA) users in Multiple Sclerosis (MS).


      This longitudinal retrospective study used adults(≥18 years) with MS diagnosis (ICD-9-CM:340) and a DMA prescription from the IBM MarketScan Commercial Claims and Encounters Database from 2010–2012. Patients were classified into fingolimod or injectable users based on their initial DMA prescription. The composite endpoint (time-to-relapse/DMA switch) was assessed during the one-year follow-up period after DMA initiation. The stabilized IPTW-Cox Proportional Hazards regression model was used to evaluate the composite endpoint with three different variance estimators – (i)Naïve, (ii)Robust sandwich-type, and (iii)Bootstrapping(200 replications). Patients who died/were lost from follow-up due to the lack of insurance coverage were censored.


      The new DMA user study cohort consisted of 1,700 MS patients who were initiated with oral(15.82%) or injectable(84.18%) DMAs during 2010-2011. The proportion of patients who had a composite endpoint in fingolimod and injectable DMA users was 16.72% and 27.16%, respectively. The stabilized IPTW-Cox model with naïve and bootstrapping variance estimators revealed that oral fingolimod users were superior to injectable DMAs in reducing the risk of composite endpoint (Naïve estimator: Adjusted Hazards Ratio [aHR]-0.67, 95%CI:0.51-0.87; Bootstrapped estimator: aHR-0.68, 95%CI:0.39-0.97). However, the findings were not significant in the IPTW-Cox model with robust sandwich estimator(aHR-0.67, 95%CI:0.43-1.03).


      The analyses revealed that the significance of treatment effect estimates could vary depending on the choice of variance estimation method. Hence, researchers should pay attention to the selection of variance estimation method with small samples in addition to handling of extreme weights while using IPTWs for time-to-event analyses.