When the audit trail tells the truth too late
Case at a glance
In April 2026, FDA’s Center for Drug Evaluation and Research proposed to withdraw marketing approval for TAVNEOS (avacopan), the first drug approved on the strength of a single pivotal trial whose primary endpoint had been rescued through undisclosed post-unblinding readjudication. The case pairs a textbook data integrity failure with two distinct statutory grounds for withdrawal, and it ended with a rare retraction of the pivotal publication from the New England Journal of Medicine. It offers an unusually complete record of how a sponsor defeated its own control environment, how those controls were designed to stop exactly what happened, and how the failure surfaced years later through litigation rather than inspection.
The core lesson is narrow and durable: the integrity of a trial result depends less on the sophistication of the analysis than on the sequence and timing of database lock, unblinding, and adjudication. When a sponsor reopens a locked, unblinded database to change endpoint determinations, no subsequent statistical treatment can restore the interpretability of the result.
Thanks for reading AcceleraQA’s Substack! Subscribe for free to receive new posts and support my work.
Background: the drug, the disease, and the approval
FDA approved NDA 214487 for TAVNEOS (avacopan) 10 mg capsules on 7 October 2021, for adjunctive treatment of adult patients with severe active anti-neutrophil cytoplasmic autoantibody (ANCA)-associated vasculitis (AAV) in combination with standard therapy including glucocorticoids. AAV is a rare, life-threatening autoimmune condition, and much of its long-term morbidity comes from the glucocorticoids used to treat it. Avacopan, an oral C5a receptor antagonist, promised remission with reduced steroid exposure, which framed the clinical appeal and shaped the regulatory flexibility the drug later received.
Approval rested on one pivotal study, the phase 3 ADVOCATE trial (protocol CL010_168, NCT02994927). ChemoCentryx held the NDA at approval; Amgen acquired the product with its 2022 purchase of ChemoCentryx and later answered to FDA on the applicant’s behalf.
The approval was contested on its own terms before any misconduct came to light. FDA convened an 18-member advisory committee in May 2021 that split on whether the evidence established effectiveness, safety, and a favorable benefit-risk balance. The primary review team did not recommend approval. Approval proceeded because the Division Director of Rheumatology and Transplant Medicine, and the Office Director of Immunology and Inflammation, judged that the demonstrated superiority on sustained remission at week 52, together with the unmet need in a rare disease, justified additional flexibility. That superiority result was the manipulated one.
The pivotal trial: ADVOCATE design
ADVOCATE randomized 331 subjects with AAV in a 1:1 ratio to avacopan 30 mg twice daily or a protocol-specified prednisone taper. All subjects received background immunosuppression with rituximab or cyclophosphamide. Investigators measured disease activity using the Birmingham Vasculitis Activity Score (BVAS), a standardized clinician-reported outcome.
The trial carried two co-primary endpoints: remission at week 26 and sustained remission at week 52. A subject counted as being in remission at a time point with a BVAS of zero and no glucocorticoid use for AAV in the preceding four weeks. Sustained remission required remission at both week 26 and week 52 with no intervening relapse. The protocol and statistical analysis plan prespecified a missing-data rule that imputed subjects with unevaluable status as not in remission.
Three design features carry the weight of this case. First, FDA had told ChemoCentryx in 2016 that it would not approve the drug on a non-inferiority showing; the week 52 sustained remission endpoint had to demonstrate superiority over the control arm. Second, a blinded adjudication committee determined BVAS, remission, and sustained remission, governed by a BVAS and VDI Adjudication Committee Charter (the AC Charter). Third, the AC Charter and the statistical analysis plan tied adjudication to a controlled sequence: clean the data, lock the database, and only then unblind and analyze. The charter expressly prohibited readjudication of remission status after database unblinding and prohibited any changes in the electronic data capture (EDC) system after database lock.
Timeline of the failure
The manipulation occurred inside a two-week window in November 2019 and stayed concealed for more than five years.
On 4 November 2019, the Medpace lead biostatistician completed a pre-database-lock checklist confirming the datasets, the approved statistical analysis plan, and the validated statistical programs were ready. On 5 November 2019, Medpace data management signed the database lock checklist attesting that all data were entered, all edit checks run, all queries resolved, source data verified, and all records locked from editing. The database was then locked and unblinded on 5 November 2019, and the prespecified primary analysis was run.
That original analysis failed. Sustained remission at week 52 returned a two-sided p-value of 0.1025, above the 0.05 threshold the statistical analysis plan set for the controlled Type I error rate.
Between 5 and 7 November 2019, unblinded ChemoCentryx personnel reviewed the tables, listings, and figures. Personnel then selected nine subjects for readjudication, including five avacopan subjects whose week 52 status would flip from “not in sustained remission” to “sustained remission.” According to the record FDA assembled, personnel confirmed before sending the subjects out that changing those five determinations would make the study result statistically significant. A blinded adjudicator, Dr. Jayne, then performed the readjudications.
The endpoint determinations were changed in the database, and the database was locked a second time on 20 November 2019. The primary analysis was rerun and now returned a two-sided p-value of 0.0132. That second analysis was the only primary endpoint result submitted in the NDA.
The concealment held until 29 May 2025, when plaintiffs in shareholder securities-fraud litigation filed the expert report of Marc Walton, M.D., Ph.D., which described the sequence. FDA issued an information request on 23 July 2025; Amgen’s 22 August 2025 response confirmed the key facts while defending the readjudications as appropriate error correction. FDA proposed withdrawal on 27 April 2026. The NEJM retraction followed on 29 June 2026.
What actually went wrong: the readjudication mechanism
The failure was not a fabrication of data and not a statistical sleight of hand in the analysis code. It was a controlled process run in the wrong order, with the outcome known in advance.
The table below reproduces the effect of the nine readjudications on the endpoint determinations. An asterisk marks the five avacopan flips at week 52 that changed the study outcome.
ET denotes early termination, with week 52 imputed as non-response. The three control readjudications occurred at week 26, where they could not affect the control arm’s week 52 sustained remission count, so their inclusion did not convert a targeted search into a neutral one.
The mechanism matters for anyone auditing adjudication workflows. The personnel who selected subjects were unblinded and knew the failing topline result. They searched for subjects whose status could plausibly be reversed, concentrating on patients with a BVAS of zero who had been scored as not in remission because of missing data or recent glucocorticoid use. The blinded posture of the final adjudicator did not cleanse the process, because the selection of who gets readjudicated is itself the point of bias. A blinded adjudicator asked to re-examine a hand-picked set of near-miss subjects will, on average, move some of them, and the selector already knew which direction mattered.
Control environment failures
This case is valuable precisely because the controls existed and were specific. The sponsor did not lack a control framework; it overrode one.
The AC Charter prohibited what happened, in three overlapping ways. It stated that a completed adjudication form would not be presented to the committee again at a later time. It permitted readjudication only when site-originated data changes after adjudication could affect the original decision, and it made Medpace responsible for tracking such post-adjudication data changes and circulating listings of them. And it prohibited entry of changes into EDC after database lock and prohibited blinded BVAS listings for review after unblinding. Because every remission determination rests on BVAS data, that last provision effectively barred any post-unblinding readjudication. There were no site-originated data changes that triggered these nine readjudications. The trigger was the failed result.
Database lock and unblinding were treated as reversible when the framework treated them as terminal. The entire logic of the lock-then-unblind sequence is that lock freezes the record and unblinding starts the clock on bias. Reopening a locked, unblinded database to alter endpoint determinations collapses that logic. The sponsor had multiple earlier and legitimate opportunities to correct any genuine errors: routine data cleaning before initial adjudication, the blinded data review meeting on 24 October 2019, and the quality-control checks immediately before the 5 November lock. One subject, avacopan 534–001, had actually been identified before unblinding as a non-responder under the missing-data rule, and was reclassified only afterward. The correct time to fix a misapplied imputation rule is before the lock, not after the result is known.
“Quality control” was the wrong label for the activity. Amgen defended the readjudications as appropriate steps to preserve data accuracy. The distinction that matters for practitioners is between quality control performed on blinded data before lock and outcome-driven revision performed on unblinded data after a failed analysis. The former protects the result; the latter destroys its interpretability. Blinded QC checks are ordinary and expected. Using unblinded results to decide which subjects to re-examine is not QC under any recognized definition.
Data integrity analysis through the ALCOA+ lens
The endpoint values in the final database were arguably accurate in a narrow sense; the deeper failures were of contemporaneity, originality, and the integrity of the record as a whole.
Contemporaneous. The endpoint determinations submitted to FDA were not the determinations reached contemporaneously at lock. They were later revisions, made after the result was known, and the record was presented as though the later state was the only state.
Get Chris Gent’s stories in your inbox
Join Medium for free to get updates from this writer.
Original. The original analysis and its failing p-value were the true first record of the trial’s outcome. That original record was suppressed. The NDA disclosed only the revised state.
Attributable and traceable. The changes were traceable in the sense that an audit trail and the underlying documents eventually allowed reconstruction, which is how the Walton report and Amgen’s own response could establish the sequence. This is the paradox of the case: the audit trail did its job forensically, years later, but nothing in the routine review surfaced the two-lock sequence because the sponsor’s narrative concealed it. A robust audit trail is necessary but not sufficient; someone has to read it against the sponsor’s representations.
Complete and consistent. The clinical study report disclosed and documented a set of minor post-lock corrections to vasculitis damage index scores, complete with original and revised listings. It did not disclose the far more consequential post-lock, post-unblinding readjudication of primary endpoint data. A record that documents trivial changes while omitting outcome-determining ones is neither complete nor consistent, and the selective transparency is itself evidence of intent.
The case is a reminder that ALCOA+ compliance is not satisfied by a defensible final dataset. It requires that the record faithfully represent the sequence of decisions that produced that dataset, including the decisions that failed.
The two statutory grounds for withdrawal
FDA proposed withdrawal on two independent grounds under section 505(e) of the Federal Food, Drug, and Cosmetic Act.
Lack of substantial evidence of effectiveness under section 505(e)(3) and 21 CFR 314.150(a)(2)(iii). New information, evaluated with the evidence available at approval, showed no substantial evidence from adequate and well-controlled investigations. FDA’s argument runs through 21 CFR 314.126: an adequate and well-controlled study must analyze results adequately and must take adequate measures to minimize bias among subjects, observers, and analysts. The reanalysis did neither. The statistical analysis plan never planned for repeated analyses of the primary endpoint, so the second analysis inflated Type I error without any multiplicity control. FDA’s conclusion is categorical: the reported p-value of 0.0132 is uninterpretable, no further blinded or unblinded reanalysis can salvage the study because each additional analysis compounds the false-positive risk, and the only interpretable analysis is the original one from the 5 November lock, which failed. The agency framed the alternative as an incentive it would not tolerate, in which a sponsor caught after approval simply reanalyzes until a positive result appears.
Untrue statements of material fact under section 505(e)(5) and 21 CFR 314.150(a)(2)(iv). This ground does not depend on the effectiveness analysis. The applicant represented a single database lock on 20 November 2019 and never disclosed the 5 November lock, the failed original analysis, or the readjudication. In response to a December 2020 FDA information request that probed precisely whether the analysis plan had been finalized before unblinding, the applicant characterized the freeze as a single event on 20 November 2019. The clinical study report likewise stated that treatment assignments remained blinded until after the database had been cleaned and locked, which was false with respect to the only lock date the NDA disclosed. FDA tied these representations to specific content requirements in 21 CFR 314.50 and to the applicant’s certification that the submission was true and accurate. The materiality is direct: the Division Director testified, in effect, that he relied on the reported superiority and on the assumption that trial-quality procedures had been followed, and would not have recommended approval had he known otherwise.
The safety overlay
Effectiveness and honesty were the grounds for withdrawal, but the safety picture removed any counterweight. Hepatotoxicity was a known risk at approval. Through October 2024, FDA’s review of its Adverse Event Reporting System, the sponsor’s global safety database, and the literature identified 76 cases of drug-induced liver injury associated with avacopan, of which 74 reported a serious outcome, including 54 hospitalizations and 8 deaths. Seven cases reported biopsy-confirmed vanishing bile duct syndrome, a serious adverse event that was unexpected at approval. The required postmarketing study (PMR 4155–1), meant to characterize these risks and confirm efficacy, had enrolled only 21 of a planned 300 patients. With no demonstrated efficacy benefit, FDA concluded there was nothing to balance against a serious and partly unexpected safety profile.
Detection and unraveling
Three features of how this surfaced deserve attention from anyone who designs oversight systems.
It surfaced through litigation, not inspection. The securities-fraud discovery process, not a GCP inspection or a sponsor self-disclosure, produced the Walton report. Routine review had missed it because the concealment was competent: the NDA told an internally consistent story about a single lock.
The court result and the regulatory result diverged. On 15 August 2025 the court granted summary judgment for ChemoCentryx and dismissed the shareholder suit without addressing the manipulation allegations. FDA treated that dismissal as irrelevant to the NDA and the FD&C Act. A finding for the defendant on securities-fraud elements does not resolve whether an NDA contains untrue statements of material fact.
The scientific record corrected itself last. On 29 June 2026, the New England Journal of Medicine retracted the ADVOCATE publication. The two academic authors requested the retraction, stating that they had learned only through the FDA investigation, after publication, that primary endpoint assessments in nine patients had been readjudicated after database lock and unblinding, and that this had not been disclosed. The retraction underscores that co-authors on a pivotal publication may not know how the sponsor produced the numbers they put their names to.
Regulatory and legal posture
The April 2026 letter is a proposal to withdraw and a notice of opportunity for a hearing under 21 CFR 314.200, not a final withdrawal. To contest it, ChemoCentryx had to file a notice of participation and hearing request within 30 days of Federal Register publication, and to submit the supporting data and analyses within 60 days. The standard for obtaining a hearing is demanding: a request cannot rest on allegations or denials and must present specific facts showing a genuine and substantial issue of material fact. If the submission does not clear that bar, the Commissioner enters summary judgment and denies the hearing. The letter was signed by Tracy Beth Høeg, M.D., Ph.D., Acting Director of CDER.
Lessons for sponsors and for eClinical and CRO vendors
For sponsors, the operative principle is that database lock and unblinding are one-way doors. Build the process so that reopening a locked, unblinded database for anything touching an endpoint is impossible without a documented, justified, prospectively defined exception that is reported to the agency. Treat any post-unblinding endpoint change as reportable by default. Recognize that a defensible final dataset does not establish integrity; the sequence that produced it does. And understand that selective disclosure, documenting minor corrections while omitting outcome-determining ones, reads as consciousness of wrongdoing rather than as diligence.
For eClinical and CRO vendors, the case sharpens several design obligations. Adjudication and EDC systems should enforce the charter’s timing rules technically, not merely document them, so that no endpoint determination can be entered or changed after the recorded lock and unblinding events. Audit trails must capture lock, unblind, and re-lock as first-class events with immutable timestamps, because in this case the two-lock sequence was the whole story and it lived in exactly those records. Data management checklists that attest a database is clean and locked acquire evidentiary weight; the 4 and 5 November checklists here became the documentary spine of FDA’s timeline. And the vendor that tracks post-adjudication data changes holds the tripwire: the AC Charter assigned Medpace responsibility for listing site-originated data changes precisely so that readjudication could be justified only by a real change, which is the check that would have flagged nine readjudications with no triggering change behind them.
The unifying point for both audiences is that automated audit trail review, continuous validation, and integrity analytics earn their value at exactly this seam. A system that reconciles lock and unblind timestamps against endpoint-change events, and that reads the sponsor’s disclosures against its own audit trail, is the mechanism that turns a competent concealment into a detectable one before approval rather than five years after it.
Works cited
Amgen, Inc. Response to FDA Requests Dated Jul. 23, 2025. NDA 214487, 22 Aug. 2025.
ChemoCentryx, Inc. BVAS and VDI Adjudication Committee Charter. Protocol CL010_168, version 4.0, NDA 214487, 21 June 2019.
ChemoCentryx, Inc. Clinical Study Report CL010_168. NDA 214487, 1 June 2020.
ChemoCentryx, Inc. Protocol CL010_168. Amendment 4.0, 18 Jan. 2016.
ChemoCentryx, Inc. Statistical Analysis Plan. Version 2.0, 28 Oct. 2019.
Jayne, David R. W., et al. “Avacopan for the Treatment of ANCA-Associated Vasculitis.” New England Journal of Medicine, vol. 384, no. 7, 2021, pp. 599–609. doi:10.1056/NEJMoa2023386. Retracted.
Love, Sharon, et al. “Making a Distinction Between Data Cleaning and Central Monitoring in Clinical Trials.” Clinical Trials, 2021.
Rubin, Eric J. “Retraction: Jayne DRW et al. Avacopan for the Treatment of ANCA-Associated Vasculitis, N Engl J Med 2021;384:599–609.” New England Journal of Medicine, 29 June 2026. doi:10.1056/NEJMe2608684.
United States, Food and Drug Administration, Center for Drug Evaluation and Research. Proposal to Withdraw Marketing Approval; Notice of Opportunity for a Hearing. NDA 214487, Docket No. FDA-2026-N-1321, 27 Apr. 2026.
United States, Food and Drug Administration. ICH Guidance for Industry E6(R3) Good Clinical Practice. Sept. 2025.
United States, Food and Drug Administration. ICH Guidance for Industry E9 Statistical Principles for Clinical Trials. Sept. 1998.
Walton, Marc. Expert Report. 25 Sept. 2024, Homyk v. ChemoCentryx, Inc., Case №4:21-cv-03343 (N.D. Cal.).
Warrington, Kenneth J. “Avacopan for the Treatment of ANCA-Associated Vasculitis.” Editorial. New England Journal of Medicine, vol. 384, no. 7, 2021, pp. 664–665. doi:10.1056/NEJMe2033621.
Additional references cited in the FDA withdrawal letter and internal to NDA 214487, including the applicant’s responses to CDER information requests, the advisory committee record, the summary review and approval documentation, and the postmarketing requirement documentation, are enumerated in the letter’s reference list (Refs. 13–20) and are not independently public.
