Sample Size Re-estimation: catching a trial that is the wrong size
Every trial is sized before it begins based on two estimates: how large the effect size will be, and the variance from patient to patient. Both come from earlier, smaller studies, and both have the potential to be wrong in the same directions. Assumed benefits tend to be too optimistic, because only promising phase 2 results lead to a phase 3. Assumed variability tends to be too low.
When either is wrong, the trial runs to completion underpowered. It enrolls the number of patients the original calculation called for, which is no longer enough to demonstrate an effect that may genuinely be there. The drug works; the trial was simply too small to show it. The result neither confirms nor rules out a benefit, and the asset is usually abandoned.
Sample size re-estimation is the established answer. What has limited it in practice is that correcting for a smaller-than-expected benefit requires an unblinded look, which means a scheduled interim analysis and a committee review. Most trials have at most one such look, often placed too late to help. Of 152 completed adaptive trials, fewer than 10% actually increased their sample size.
Sample size re-estimation removes the bottleneck. Inside the secure enclave, blinded to all humans, it checks the outcome variability at every, blinded, scheduled look across a window set before the trial, and assesses the treatment effect once, at a pre-specified point. If either check finds the trial undersized, it makes available to the Data Monitoring Committee a recommended sample size, never below the planned size and never above a maximum agreed in advance.
Two things matter more here than the mechanics:
It recommends; the DMC decides. The platform does not resize anything. It identifies that a decision point you pre-specified has been reached, and hands the committee a number along with the reasoning behind it in plain language.
It will not try to rescue a trial that is not worth rescuing. If the treatment is not showing enough promise to justify more patients, the platform makes no recommendation and leaves the trial to our futility monitoring instead.
Because a recommendation can change the size of the trial it is running on, the use of Pre.sentient’s platform is always pre-specified in the protocol and the statistical analysis plan as part of an adaptive design, with the final analysis using a pre-agreed combination test so the false-positive rate stays where it should be.
Using this method, in a benchmark of 72,000 synthetic phase 3 trials, Pre.sentient’s method resized 86.4% of underpowered trials, against 52.3% for the best conventional method tested, and its recommendations arrived earlier: a median of 67% of planned enrollment rather than 75%, which across a typical recruitment period is roughly one to two months. Across 100 phase 3 trials with a real treatment effect, it converted 13 of the 15 that the planned sample size would have lost, at a cost of around 141 additional patients on average for a rescued trial.
Sponsor simulations: test the platform on your own terms
Earlier this year we released simulation tooling so regulators could verify the performance of our algorithms themselves, using their own data, without us seeing the results. With this release, sponsors can do the same. Sponsors with the appropriate role can generate simulated trial data, run it through a real platform configuration, and compare what the algorithms concluded against what would actually have happened.
Two things this is useful for:
Seeing the performance characteristics for yourself. Rather than reading our figures on how early futility, success or an undersized trial is detected, you generate the trials and produce the figures. Simulation is supported for Success and Futility and Strata modules.
Comparing against your current approach. Run the same simulated trials through your existing interim analysis plan and see what each would have concluded, and when.
Availability
If you would like to speak to us about how the Pre.sentient Real Time Trial Analysis platform can assist in your trial, please get in touch using the contact form below.