Pharmaceutical Audit Processes, Data Integrity, And GxP

What's being tested

Interviewers probe your ability to detect, quantify, and defend conclusions about data integrity within a regulated GxP environment using statistical reasoning and modeling. They want to see how you translate noisy operational signals into audit‑grade evidence: experimental design for sampling, anomaly detection with controlled false‑positive risk, and model validation/documentation that meets regulatory expectations. Capital One cares because regulated workflows require traceable, statistically defensible decisions when data quality impacts risk, compliance, or patient safety.

Core knowledge

GxP: stands for Good Practices (e.g., GMP/GCP/GLP); it requires traceability, accountability, and demonstrable controls over data creation, modification, and retention, not infrastructure design.
ALCOA+: data must be Attributable, Legible, Contemporaneous, Original, Accurate (plus Complete, Consistent, Enduring, Available) — use these to define evidence requirements for any analytic claim.
Audit trail: treat system logs as a signal source — timestamps, user IDs, and hash/diff comparisons are features; verify immutability assumptions before counting on them.
Statistical process control (SPC): use Shewhart (mean ± 3σ), CUSUM ( $S_t = \max(0, S_{t-1} + x_t - \mu - k)$ ) and EWMA ( $z_t=\lambda x_t+(1-\lambda)z_{t-1}$ ) to detect shifts in measurement systems with tunable detection delay vs false alarm tradeoffs.
Hypothesis testing & error rates: explicitly report type I/II errors, control familywise error or FDR (Benjamini–Hochberg) when screening many variables; compute power with $n = \left(\frac{z_{1-\alpha/2}+z_{1-\beta}}{\delta/\sigma}\right)^2$
Sampling design for audits: prefer stratified random sampling by risk strata; compute sample size to detect a defect rate $p$ with confidence: $n \approx \frac{z^2 p(1-p)}{d^2}$ and adjust for finite population.
Anomaly detection: choose between statistical models (z‑scores, robust MAD, control charts), density methods (isolation forest, LOF) or supervised classifiers; prefer interpretable signals and calibrated scores for audit trails.
Missing data & imputation: test MCAR/MAR/MNAR assumptions; avoid naive mean imputation for audit evidence — use multiple imputation and report sensitivity.
Outlier handling: prefer robust estimators (median, MAD) or winsorization with documented rules; never drop records without documenting decision criteria.
Model validation under GxP: require fit‑for‑purpose documented validation set, held‑out test, performance thresholds, artifact reproducibility, and versioned code/seed; keep acceptance criteria pre‑specified.
Causal vs correlation: when alleging root cause, favor causal inference (difference‑in‑differences, interrupted time series, instrumental variables) and enumerate identifying assumptions for auditors.
Documentation & reproducibility: produce runnable artifacts (R/Python notebooks in Jupyter) with data provenance, seed, environment, metric definitions, and concise audit writeup.

Worked example — "Design an anomaly detection approach to find data‑integrity issues in lab assay results"

First 30 seconds: clarify scope — are we monitoring a single lab instrument, batch runs, or an aggregated database? Confirm acceptable false positive rate and regulatory reporting timeline. Skeleton answer pillars: (1) define concrete detection targets (e.g., impossible values, timestamp irregularities, sudden mean shifts), (2) choose statistical detectors per target (rules + control charts + isolation forest for multivariate anomalies), (3) design sampling and alert thresholds to control false alarms (specify α and detection delay), and (4) validation and audit documentation (held‑out batch, synthetic injected faults, runbook for triage). A key tradeoff: stricter thresholds reduce missed issues but raise investigations — quantify operational cost per false positive vs cost of undetected integrity breach. Close by saying: if more time, I'd run adversarial tests (inject realistic corruptions), build a confusion‑matrix with cost weighting, and automate reproducible reports that map each alert to ALCOA+ evidence.

A second angle — "Create a statistical sampling plan for a regulatory audit of clinical trial CRF data"

Same core concepts apply but framed toward sampling: define risk strata (sites, CRF pages, high‑risk fields), determine target defect rate to detect with required power, and choose stratified random sampling over simple random sampling to reduce sample size while increasing sensitivity for risky strata. Use acceptance‑sampling logic for batch/site adjudication and prespecify decision rules (e.g., if defects > x then 100% review). Validate via simulation: simulate realistic corruption patterns and compute operating characteristic curves to ensure the plan meets regulatory expectations.

Common pitfalls

Pitfall: trusting raw system timestamps/logs as fully trustworthy evidence.

Assuming immutability without verifying log cadence, clock skew, or aggregation windows leads to false conclusions; always cross‑check with independent signals (e.g., instrument serial numbers, file checksums) and document confidence.

Pitfall: optimizing models for detection rate without controlling false discovery.

A model with high recall but uncontrolled false positives will generate audit noise; predefine acceptable FDR, tune thresholds on validation data, and report ROC/precision‑recall and expected alerts/day.

Pitfall: vague documentation and post‑hoc thresholds.

Changing thresholds after seeing results undermines regulatory defensibility; instead predefine hypotheses, acceptance criteria, and hold out test data or use synthetic injection to prove sensitivity.

Connections

Interviewers may pivot to model governance (change control, versioning), causal inference for root cause, or statistical monitoring at scale (operationalizing SPC across thousands of signals). Be ready to discuss tradeoffs between interpretability and sensitivity.