Adverse selection and stale pricing in clinical trials

What the trial is quoting, and against whom

An enrolment protocol is a price. It sets criteria — this cohort, this dose, these exclusions — and stands open until amended. Every patient screened against it, every site continuing to recruit under it, is trading against a quote fixed at the moment the protocol was written or last revised. The counterparty is the world: the adverse event that has already happened at another site, the interim analysis that has already crossed a boundary, the pharmacovigilance signal already filed in a database the enrolling site has not read. Whoever holds that later information holds a free option against the protocol's stale terms. In clinical trials, that option is not exercised for profit. It is exercised, unknowingly, against patient safety, and the bill arrives as a data safety monitoring board finding, a regulatory hold, or a cohort that must be partly discarded.

The streams and what they carry

Four feeds run continuously under any trial of reasonable size. Enrolment feeds report who has been screened, randomised, dosed, and where, updated as sites input case report forms — often with a lag of days because data entry is batched, not live. Safety signals arrive from multiple channels at different speeds: expedited serious adverse event reports required within 24 to 72 hours by regulation, routine adverse event logs updated weekly, and aggregate signal detection from pooled data that may only run monthly. Protocol amendments are the trial sponsor's own revisions — exclusion criteria tightened, a dose de-escalated, a monitoring test added — each one dated and version-controlled, in principle. Site telemetry covers the operational layer: screen-failure rates, deviation reports, query turnaround, drug accountability logs, the signals that tell you whether a site is executing the current protocol or an older one it printed out three amendments ago.

None of these streams stops when a decision is made. That is the structural fact the whole failure mode turns on. A signal can arrive the day after a cohort target is hit and still be relevant to every patient dosed since.

What is held: the belief, not the data point

The system's belief is the current inclusion/exclusion logic plus the current risk-benefit judgement for the population being enrolled — not a single number but a compound claim: "given evidence up to version 4.2, dated 11 March, patients meeting these criteria have an acceptable risk profile at this dose." That belief carries provenance if it is built properly: which safety database snapshot supported it, which interim analysis, which prior amendment it superseded. Provenance is what lets a trial monitor later ask "is this still the belief the world supports" rather than simply "is this still the belief on file."

Most trial infrastructure does not hold beliefs this way. It holds documents. The protocol PDF, version 4.2, sits in a document management system; the safety database sits in a separate pharmacovigilance system; the enrolment tracker sits in a third. Nothing forces the three to reconcile except a human reading all three and noticing a mismatch. That gap is where staleness lives.

The trigger the loop misses

A phase II oncology trial enrols against an exclusion criterion for moderate renal impairment, set at a threshold of eGFR 45. In week 14, a serious adverse event report from a different site in the same trial — a grade 4 renal toxicity in a patient at eGFR 52, just inside the current cutoff — is filed and coded correctly within the required 72 hours. It sits in the safety database. It should trigger a re-evaluation of the eGFR threshold. Whether it does depends entirely on whether anyone is watching the safety stream against the live enrolment stream in the same frame, on the same clock.

In practice the two run on separate cadences. The safety report goes to the sponsor's pharmacovigilance unit and eventually to the DSMB, which may meet quarterly or trigger only on pre-specified stopping rules. The enrolment feed keeps running on its own schedule, because enrolment targets are a separate operational priority with their own pressure — sites are paid per patient enrolled, sponsors are racing competitor trials for the same population, and nobody wants to pause recruitment on a single case pending confirmation. The belief that "eGFR 45 is an acceptable threshold" is never formally withdrawn. It is just left standing while a fact that invalidates it sits filed but unconnected.

What the trial monitor actually sees

The trial monitor is the person nominally responsible for holding these streams together, and the honest description of the job is that they see summaries, not signals. A monthly site visit report. A safety narrative distributed by the sponsor's medical monitor, often written for regulatory completeness rather than operational urgency. An enrolment dashboard showing screen-failure rates and randomisation counts, refreshed weekly. Nothing on the monitor's desk is timestamped against anything else on the desk. The eGFR case and the current enrolment count exist in different documents, produced by different teams, on different Tuesdays.

If the monitor is diligent and lucky, they cross-reference manually — read the safety narrative, notice the renal signal, check it against the current exclusion criterion, and escalate. If they are not, or if the volume of concurrent safety reports across a multi-site trial exceeds what one person can hold in working memory, the mismatch persists. Nobody made an error. Every individual system did its job. The failure is structural: no belief in the system carried a date that could be compared against another belief's date, so no comparison happened.

The cost, in months and patients

The characteristic failure is not a single bad decision. It is duration. The eGFR-52 signal is filed in week 14. The protocol amendment that tightens the threshold to eGFR 55 is not approved by the sponsor, submitted to the ethics committee, and rolled out to sites until week 27 — thirteen weeks, which is a realistic turnaround for amendment review, IRB resubmission, and site retraining, not a negligent one. In those thirteen weeks the trial enrols against the old, already-invalidated criterion. If the site accrual rate is eight patients a month across twelve sites, that is roughly 300 patients screened and perhaps 40 to 60 enrolled under a threshold the sponsor's own safety database had already disproven.

Some of those patients are fine. Some are not, and the ones who are not generate further adverse events that eventually force a larger, more disruptive amendment — a partial clinical hold, a re-consent exercise across every active site, a subset of the cohort excluded from the efficacy analysis because they were enrolled under superseded criteria and the statistical analysis plan cannot cleanly include them. The direct cost is measured in wasted screening (each screen failure costs a site $2,000 to $4,000 in coordinator time and lab work, unrecoverable), in re-consent administration, and in trial timeline: a hold or amendment of this kind routinely adds three to six months to a pivotal trial, at a fully loaded cost that sponsors estimate in the range of $600,000 to $800,000 per day of delay for a late-phase programme. The indirect cost is the one that matters more and prices less easily: exposure of patients to a risk the sponsor's own data had already flagged, which is the exact transfer the frozen-quote problem describes — value, here safety margin, moving from the party operating on old information to whoever held the newer fact and did not connect it in time.

What closing the loop would require, and what it would not fix

A structurally closed loop does not mean faster reporting. The 72-hour serious adverse event timeline is already fast enough; the eGFR signal was available in week 14 whether anyone looked or not. What was missing was a shared, dated representation: enrolment criteria and safety evidence living in the same timestamped ledger, so that a threshold and the evidence against it are comparable objects rather than documents in different systems on different clocks. That is a provenance problem, not a speed problem — closer to Copeland and Galai's resting quote than to a latency arms race. Making staleness visible, criterion dated against evidence dated, is what would have let the trial monitor see the mismatch in week 15 rather than the DSMB force it in week 27.

None of this removes the case for frozen, audited protocols. A trial that re-wrote its inclusion criteria every time a single case report arrived would be unrunnable, statistically incoherent, and impossible to consent patients into with any stability. The argument is narrower: the protocol can stay fixed between amendments, but the system watching it should never lose track of which evidence the fixed version has already been overtaken by. The amendment can still take thirteen weeks. What must not take thirteen weeks is noticing that it is needed.