OBJECTIVES: Progression-free and event-free survival are common primary endpoints in oncology trials. Initial response assessments conducted at fixed times (e.g., every 12 weeks) often lead to left-censoring. We sought to determine the impact of left-censoring in primary analyses and in ITCs through a simulation analysis.

METHODS: We simulated data from two trials with divergent time-to-event evaluation schedules. Our simulations compared 108 parameter permutations: true hazard ratio (HR; 0.25, 0.50, 0.75, or 1.00); sample size (50, 100, or 500); distribution (Weibull, Exponential, or Gompertz); and time intervals of evaluation for each trial with differences of 3, 6 and 12 weeks. Data were generated over 500 simulations for each combination of parameters using the simsurv package in R. Events occurring between timed evaluations were changed to the set evaluation time. We measured the mean squared error and mean bias of the log hazard ratios for both individual studies and ITCs to compare Cox proportional hazards regression to interval censoring regression.

RESULTS: The use of Cox regression, in the presence of left censoring, led to some degree of bias in the estimated HR of a given study. This varied from negligible to a difference of 0.059 on the log-hazard scale (i.e., HR of 1.061 as an estimate for 1.00). Similar results were observed using interval censoring, albeit ranging from negligible to 0.040 (HR 1.041 as an estimate for 1.00). Factors associated with larger bias were larger between-trial differences in time of evaluation and smaller sample sizes. Despite bias in the individual estimates, this did not translate to bias in the ITCs as all bias was negligible.

CONCLUSIONS: Our simulation study suggests that, despite biased estimates of HR in individual studies, ITCs are unaffected. This supports the use of Cox regression in place of interval censoring regression.