P43 A Case Study and Simulation to Compare Different Indirect Treatment Comparison Methods Under Varying Access to Individual Patient Data


      To compare alternative methods for unanchored indirect comparisons of two interventions for overall survival when individual patient data (IPD) is available for both studies (IPD-IPD analyses), and when IPD is only available for one study and aggregate data (AD) for the other study (IPD-AD analyses).


      A case study comparing nivolumab with standard of care (SoC) in 3L small cell lung cancer was performed using the CheckMate032 trial and the Flatiron Health database to compare unanchored comparisons without population adjustment, IPD-IPD adjustment analyses using inverse probability treatment weighting (IPTW), regression adjustment (RA), and doubly robust methods (DRMIPD-IPD), and IPD-AD analyses using matching-adjusted indirect comparison (MAIC), simulated treatment comparison (STC), and a novel DRMIPD-AD. Relative treatment effects were expressed with Cox hazard ratios (HRs) and differences in restricted mean survival time (DRMST). A simulation study was performed evaluating the performance of the alternative methods in different scenarios.


      Conditional HRs from RA differed from the marginal HRs from IPTW due to non-collapsibility. Therefore, estimates of DRMST in months were used as the relative treatment effect measure for the case study, which provides a collapsible estimand for DR estimates and comparison between methods DRMST with nivolumab versus SoC was higher for RA (3.22) than IPTW (2.72) and DRMIPD-IPD (2.71). MAIC, STC, and DRMIPD-AD estimates were similar (2.88; 2.92; 2.88) but with more uncertainty. The simulations found that when covariates differed substantially between studies, IPTW performed poorly in comparison to RA and DRM.


      Analysts should be aware of differences between marginal and conditional treatment effects when performing indirect comparisons of time-to-event outcomes; DRMST may be considered. Regression-based or DRM models may be preferable over other adjustment methods in cases with limited overlap in covariates between studies indirectly compared. Having IPD for both studies is preferable over having IPD for only one.