The Site With 45% Compliance

The 14 entries in one sitting, with a median completion time of 3 seconds and clustering at 09:58, are a strong signal of retrospective completion rather than daily contemporaneous reporting. That matters because the diary is a 24-hour recall measure; entries made later do not measure the same thing. The first step is to pull the audit trail: creation time, submission time, device identifier, time zone, and whether the app was opened on the days in question. Ask the participant and coordinator whether the entries were made from memory, from notes, or with staff assistance. Do not 'fix' the dates; altering timestamps would destroy the audit trail and create a false record. Treat the affected days as potentially invalid and document the query and outcome. The shared-phone and notification complaints are separate operational issues that may explain broader missingness, but they do not excuse backfilled entries. The site should not decide exclusion alone; the sponsor and medical monitor need a rule applied consistently across sites.

Do not exclude the site reflexively. First, separate capability from conduct. Retrain the coordinator on the diary window, the prohibition on completing entries for participants, and the need to document device problems. Review notification settings with each participant, reduce non-essential alerts, and offer a private-device or loaner option where shared phones are a barrier. Confirm that participants can complete within the 06:00–10:00 window and that the site is not instructing them to catch up at visits. Set a written remediation plan with weekly compliance review, a defined threshold for escalation, and a re-audit of the next 14 days. If the site cannot meet the threshold after remediation, restrict new enrollment and escalate to the sponsor. Excluding the site entirely would lose the participants already enrolled and may introduce selection bias. The participant with 14 clustered entries needs an individual query and a documented validity decision, not a site-level punishment.

Complete-case analysis is not automatically wrong, but it assumes that the participants with complete data are representative. Here, site 12 has both low compliance and evidence of retrospective entry, so complete cases may over-represent better-performing sites and more engaged participants. Before the cutoff, the statistical team should tabulate missingness by site, week, and participant characteristics, and should document whether the pattern is plausibly related to the outcome. Pre-specify a sensitivity analysis, such as multiple imputation under a plausible missing-not-at-random assumption, or a worst-case/best-case bound for the primary endpoint. Do not pool the 14 backfilled days as if they were daily entries; either exclude those days from the primary dataset with a documented rule or include them in a sensitivity analysis. The goal is not to make the data look complete but to show how the conclusion changes under defensible assumptions. The steering committee should review the missingness report before database lock.