Futility and toxicity monitoring without halting for interim analyses.
e13579 Background: Many trial designs with futility or toxicity monitoring require that accrual halt to wait for currently enrolled patients to complete their assessment window. However, logistics often prevent this. In these cases, the performance of those designs might be compromised. This study examines the performance of 2 popular designs when enrollment is not halted at interim. Methods: Simulations were run to examine the effect of continuous enrollment on the operating characteristics (OCs) of a Simon’s 2-stage design for futility monitoring and a design using Bayesian posterior probabilities for toxicity monitoring. Both sets of 10,000 simulations examined the OCs when accrual rate was 0.5, 1.5, 3 and 5 patients/month with an assessment window of 30, 60 and 180 days. Results: The first scenario examined the OCs of a Simon design with 12 patients in the first stage and 21 at the end of the second stage. Regardless of accrual rate, the expected number of patients (EN0) increased and probability of early termination (PET0) decreased under the null hypothesis. Rate of change increased as assessment window increased. EN0 was 16 and PET0 was 54% when halting enrollment between stages. With continuous enrollment, EN0 ranged from 16-19, 17-21, and 18-21 patients for the 30-, 90- and 180-day assessment windows. PET0 ranged from 54%-50% with a 30-day assessment window. It halved to 24% with 3 patients/month enrolled and a 90-day window. PET0 was essentially 0 with a 180-day window and an enrollment of 3 patients/month. OCs for toxicity monitoring were examined for the early stopping rule Pr(toxicity rate > 0.3 | data) > 0.85 with toxicity rate ̃ beta(1, 1) with a maximum sample size of 20 and cohort size of 5. Expected number of patients (EN) increased and probability of early termination (PET) decreased as accrual rate increased, with rate of change increasing as assessment window increased. When the true probability of toxicity was 50% and enrollment halted between cohorts, EN was 10 patients and PET was 78%. EN was 17 and PET 54% with an assessment window of 30 days and 5 patients were enrolled per month. With a 90-day assessment window and 3 patients/month enrolled, EN was 16 and PET 59%. EN was 20 and PET was 12% with a 180-day assessment window and 3 patients were enrolled per month. Similar results were noted for cohorts of size 10 and a maximum number of 40 patients. Conclusions: The performance of designs that require halting enrollment while waiting for results of an interim analysis can be compromised by continuous accrual when assessment windows are lengthy and the accrual is fast. In these circumstances, consideration should be given to designs, such as Bayesian multiple imputation for delayed outcomes (Cai et al Stat Med 2014) and TOP2 (Lin et al JNCI 2019), that do not require accrual halt to make real-time interim analysis in the presence of pending patients, which protects patients from excessive toxicity or a futile intervention.