Work

Sequential randomization tests using e-values for clinical trial monitoring across endpoint types

Year: 2026

Type: article

Abstract: Background In randomized trials, repeated interim analyses of accumulating data must keep the Type I error controlled. Sequential analyses that use e-values allow continuous assessment of evidence while controlling their own Type I error. We present e-RT, a randomization-based e-value method for continuous trial monitoring, intended mainly as a non-binding ancillary review trigger for data safety monitoring boards, used alongside prespecified primary stopping boundaries. Methods At each update, the endpoint data are reviewed before the current randomized assignment enters the betting-martingale wealth calculation. Under the null hypothesis, the assignment cannot be predicted from the observed data, so the wealth process is a nonnegative martingale and Ville’s inequality controls the monitor’s Type I error at any time. We describe variants for binary, event-only, and continuous endpoints. A charter prespecifies the alert’s role: process-only review, standalone stopping rule, or path to an efficacy stop. We evaluated operating characteristics in simulations against calibrated group-sequential designs matched on sidedness and Type I error, and illustrate the method by reanalyzing two completed mechanical ventilation trials. Results The calibrated one-sided design showed 11.6 percentage points higher power than the design-calibrated binary e-RT variant, although e-RT signaled earlier, at a null crossing rate below the nominal 5%. Under a treatment effect, an e-RT alert occurred before 91.8% of scheduled stops, at a median lead of 776 patients (in the process-only role, an alert that leaves the primary analysis’s Type I error unchanged). Letting every crossing stop the trial raised the null stop rate to 7.2%; a confirmation-bar repair returned it to 5.6% while preserving power, and an alpha-partition repair is bounded by construction at a modest power cost. Conclusions e-RT offers a transparent surveillance method for randomized trials that remains valid at any time, whatever the wager policy, as long as wagers stay predictable. Under a prespecified charter, it can alert data safety monitoring boards to accumulating evidence between scheduled interim analyses. Trial registration Not applicable: this manuscript describes a statistical monitoring method; reanalyses used deidentified completed-trial data from the National Heart, Lung, and Blood Institute biorepository (BioLINCC).

Source: BMC Medical Research Methodology

DOI: 10.1186/s12874-026-03048-z

Authors: Fernando G. Zampieri, Arthur M. Albuquerque, Aaditya Ramdas

Institutions: University of Alberta, Canadian VIGOUR Centre, Universidade Federal do Rio de Janeiro, Carnegie Mellon University

Language: English

Topic: Statistical Methods in Clinical Trials

Subfield: Statistics and Probability

Field: Mathematics

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: gold

Estimated APC paid: $3,090