Approaches to Selecting “Time Zero” in External Control Arms with Multiple Potential Entry Points: A Simulation Study of 8 Approaches

Abstract

Background When including data from an external control arm to estimate comparative effectiveness, there is a methodological choice of when to set “time zero,” the point at which a patient would be eligible/enrolled in a contemporary study. Where patients receive multiple lines of eligible therapy and thus alternative points could be selected, this issue is complex. Methods A simulation study was conducted in which patients received multiple prior lines of therapy before entering either cohort. The results from the control and intervention data sets are compared using 8 methods for selecting time zero. The base-case comparison was set up to be biased against the intervention (which is generally received later), with methods compared in their ability to estimate the true intervention effectiveness. We further investigate the impact of key study attributes (such as sample size) and degree of overlap in time-varying covariates (such as prior lines of therapy) on study results. Results Of the 8 methods, 5 (all lines, random line, systematically selecting groups based on mean absolute error, root mean square error, or propensity scores) showed good performance in accounting for differences between the line at which patients were included. The first eligible line can be statistically inefficient in some situations. All lines (with censoring) cannot be used for survival outcomes. The last eligible line cannot be recommended. Conclusions Multiple methods are available for selecting the most appropriate time zero from an external control arm. Based on the simulation, we demonstrate that some methods frequently perform poorly, with several viable methods remaining. In selecting between the viable methods, analysts should consider the context of their analysis and justify the approach selected. Highlights There are multiple methods available from which an analyst may select “time zero” in an external control cohort. This simulation study demonstrates that some methods perform poorly but most are viable options, depending on context and the degree of overlap in time zero across cohorts. Careful thought and clear justification should be used when selecting the strategy for a study.