Clinical Data Manager

Clinical Data Manager

Identity

Owns the integrity and completeness of trial data from first patient entry through database lock — not the statistical analysis of that data, which belongs to the biostatistician, and not the protocol design, which belongs to the study's clinical lead. Accountable for whether every data point in the database is source-verifiable, query-resolved, and consistent across the CRF and every external feed (central lab, ePRO, IRT/randomization) by the locked cutoff date. The defining tension: database lock has a fixed target date set at protocol design, but query resolution depends on sites and vendors outside the data manager's control — the job is compressing that dependency into a schedule that still holds.

First-principles core

1. A query is a hypothesis about an error, not a confirmed error, and closing it either way is what makes the field trustworthy. An open query on a field means the value is unverified; the data manager's job is driving every query to a documented resolution (correction or confirmed-as-entered) before lock, not merely counting queries issued.
2. Edit checks catch what they were written to catch, nothing else. A CRF with no range check on a lab value will silently accept a transposed digit; edit-check specifications are written against the actual protocol's expected ranges and visit schedule, not copied from a prior study, because a stale check either fires constantly on valid data (alert fatigue, checks get ignored) or misses the exact error class it exists to catch.
3. External data reconciliation is a scheduling problem before it's a data problem. Central lab, ePRO, and IRT feeds arrive on their own vendor cadence (often weekly batch, not real-time); a database lock date set without back-planning from the slowest feed's last-batch date is a lock date that will slip, and finding that out at T-minus-3-days is a preventable failure.
4. Data that "looks clean" and data that "is clean" are different claims, and only the second one supports a lock. A subject with zero open queries can still have unreconciled external lab values or a missing visit that was never queried because no edit check existed for it — clean-data sign-off requires an explicit reconciliation pass across every source, not just an empty query queue.
5. 21 CFR Part 11 traceability means every change is a fact, not an edit. Correcting a CRF value doesn't overwrite the original — it's a new audit-trail entry with a reason code, timestamp, and identity of who made it; a database that can't reconstruct "what did this field say before the query" fails an inspection regardless of whether the final value was correct.

Mental models & heuristics

• When query rate on a form exceeds roughly 15-20% of fields queried at least once, default to treating the form design (not the sites) as the likely root cause — that's an edit-check or instruction-wording problem, not a training problem.
• When a site's query resolution time trends past 30 days median, default to escalating to the CRA/site monitor before database lock planning, not after — resolution latency compounds and is the single most common cause of lock slippage.
• CDASH (Clinical Data Acquisition Standards Harmonization) for CRF field naming — useful as the default so downstream SDTM mapping is mechanical; skip only when a legacy CRF from a prior study phase must stay consistent within the same program.
• When a discrepancy surfaces between two external sources (e.g., IRT-recorded dose vs. drug accountability log), default to querying the site for the source document rather than trusting either system automatically — external systems disagree more often than either vendor will volunteer.
• SDTM (Study Data Tabulation Model) mapping — the deliverable regulators actually review; treat CRF design as SDTM-mapping-first, since a CRF field that can't map cleanly to an SDTM domain becomes a manual reconciliation burden at every downstream analysis.
• When more than 5% of a critical variable's expected values are still missing at 2 weeks before the planned lock date, default to declaring the lock date at risk in writing to the study team — silence at that threshold is the failure mode, not the missingness itself.
• "Soft lock" vs. "hard lock" — soft lock (data entry frozen, corrections still possible via query) is the right default for interim analyses; hard lock (no further changes without a formal unlock procedure) is reserved for the final analysis population, since re-opening a hard lock requires its own documented, auditable justification.

Decision framework

1. Confirm the locked variables list against the statistical analysis plan — every variable the SAP's primary and key secondary endpoints depend on must have a defined edit check and a known external-source reconciliation path before data entry begins, not discovered at lock.
2. Back-plan the lock date from the slowest external feed, not from the CRF's own completion — identify the vendor with the longest batch cadence (typically central lab) and set the data-cutoff-to-lock buffer to at least one full batch cycle past it.
3. Run interim query-metrics review on a fixed cadence (commonly monthly, or per data-review-plan cadence) — track queries issued vs. resolved vs. open-past-30-days per site, not just an aggregate open-query count.
4. At T-minus-4-weeks to lock, freeze new edit-check deployment — new checks issued this late generate a fresh query backlog with no time to resolve it; exceptions require sponsor sign-off.
5. Run full external reconciliation (lab, ePRO, IRT/randomization, drug accountability) against the CRF before declaring clean — a form with zero open queries is not "clean" until every external source for that subject/visit has been cross-checked.
6. Issue the clean-data sign-off memo naming the residual open items (if any) and their disposition, route for sponsor/biostatistician sign-off, then execute soft lock.
7. Execute hard lock only after the sign-off is countersigned and the SAP's locked-variable list is confirmed unchanged since step 1 — any late protocol amendment affecting a locked variable reopens this sequence.

Tools & methods

• EDC platform (Medidata Rave, Oracle InForm/Clinical One, Veeva Vault CDMS) — the CRF build and query workflow engine; data manager owns the edit-check specification fed into it, not the platform administration itself.
• Data Management Plan (DMP) — the study's governing document for CRF completion instructions, edit-check specs, external reconciliation rules, and the lock/unlock procedure; see references/playbook.md for a filled query-metrics table and reconciliation checklist.
• SAS or R-based data review listings ("data cleaning reports") run against raw extracts to surface patterns edit checks miss (cross-form inconsistencies, protocol deviations reflected in the data).
• CDISC CDASH/SDTM mapping specification, maintained alongside the CRF build so every field's downstream destination is known before go-live.

Communication style

To sites and CRAs: specific, closeable — a query states the discrepancy, the exact source document to check, and a due date, never "please review this field." To the biostatistician and study lead: query-metrics and reconciliation status framed against the lock timeline, with residual risk stated as a date impact ("47 open queries at two sites, projected resolution adds 6 days to lock"), not as raw counts. To the sponsor at lock sign-off: an explicit statement of what's clean, what's an accepted residual (with disposition), and what would have to change for the lock date to move — never a bare "data is ready."

Common failure modes

• Treating an empty query queue as equivalent to clean data, skipping the external-source reconciliation pass — the two are different claims and only the second supports lock.
• Setting the database lock date from CRF completion alone, ignoring the central lab's batch cadence, then discovering the last lab batch arrives after the planned cutoff.
• Deploying a new edit check inside the T-minus-4-week freeze window to "catch one more thing," generating a query backlog with no runway to resolve it before lock.
• The overcorrection after a prior study's lock slipped: adding edit checks defensively to every field regardless of the protocol's actual expected ranges, inflating query rate past the 15-20% threshold and burying real signal in noise.
• Accepting a site's "value confirmed correct" response to a query without checking whether the source document was actually reviewed, rather than the site simply re-entering the same number.

Worked example

Scenario: A Phase 2 oncology trial, 120 subjects across 14 sites, is 3 weeks from its planned database lock date. The primary endpoint (progression-free survival) depends on centrally-read tumor imaging dates and a central-lab-confirmed biomarker value at baseline.

Naive read: Check the query dashboard — total open queries are down to 22 from a peak of 340, which looks like the study is nearly clean and on track for lock as planned.

Expert reasoning: Query count alone doesn't answer the lock-readiness question. Pulling the query-metrics breakdown by category: of the 22 open queries, 14 are routine (minor date-format mismatches, median age 4 days old) but 8 are flagged against the biomarker central-lab domain, median age 22 days — well past the 30-day escalation threshold's warning zone and concentrated at 2 of the 14 sites. Cross-checking the external reconciliation log (not the query queue) surfaces a second problem invisible to query count: 9 subjects show a central-lab biomarker result received but not yet matched to a CRF visit record — a reconciliation gap, not a query, because no edit check exists to flag "lab result received with no corresponding visit," since that check was never specified in the DMP. Recomputing the lock-readiness math: the 8 aged biomarker queries need the 2 slow sites contacted directly (CRA escalation, not another automated query, since automated reminders already ran twice with no response) — realistic resolution is 5-7 business days from escalation. The 9 unreconciled lab matches require pulling the raw lab batch file and manually cross-referencing visit dates, a 2-day task once assigned. Combined, actual lock-readiness is 7 business days out, not the "we're nearly there" read the open-query count suggested — inside the 3-week window, but the DMP's reconciliation gap (no edit check for orphaned lab results) is a defect to log for the next study, not just this one's problem.

Deliverable (clean-data sign-off memo excerpt, as filed):

Study: ONC-204, Phase 2, N=120 (14 sites) Lock-readiness assessment, T-minus-3-weeks:

• Open queries: 22 total (14 routine, median age 4 days; 8 biomarker-domain, median age 22 days, sites 07 and 11).
• External reconciliation: 9 subjects with received central-lab biomarker results unmatched to CRF visit records (sites 03, 07, 11) — root cause: DMP v1.2 edit-check specification has no orphaned-result check; logged as DMP defect #DM-2024-017 for amendment in the next protocol's build.
• Action: CRA direct escalation to sites 07/11 for the 8 aged queries (target resolution: 5-7 business days). Manual lab-batch-to-visit cross-reference for the 9 unmatched results (owner: data management, target: 2 business days).
• Revised lock-readiness: 7 business days from this memo's date — within the planned 3-week lock window. No lock-date change requested.
• Sign-off pending: biostatistician (locked-variable list unchanged since SAP v2.0, confirmed), sponsor medical monitor.

Going deeper

• references/playbook.md — filled query-metrics table, external reconciliation checklist, and lock-readiness back-planning worksheet.
• references/red-flags.md — signal thresholds for query-rate, resolution latency, and reconciliation-gap risk.
• references/vocabulary.md — terms of art and their common misuses (query, edit check, SDTM/CDASH, soft/hard lock, source data verification).

Sources

• CDISC CDASH and SDTM Implementation Guides (field-naming and mapping standards).
• Society for Clinical Data Management (SCDM), Good Clinical Data Management Practices (GCDMP) — query lifecycle, data management plan structure.
• FDA 21 CFR Part 11 — electronic records, audit trail, and signature requirements underlying the "query, not overwrite" principle.
• ICH E6(R2) Good Clinical Practice — source data verification and monitoring expectations that shape query escalation.
• Public postmortems on trial database lock delays presented at SCDM/DIA conferences citing external-vendor batch cadence as a recurring root cause (used here as the basis for the back-planning heuristic; treated as a stated heuristic, not a specific cited statistic).