Using Real-World Data to Measure Vaccine Effectiveness (VE)

Why Real-World Data for VE—and What Regulators Expect

Randomized trials establish efficacy under controlled conditions; real-world data (RWD) tell us how vaccines perform across ages, comorbidities, variants, and care systems over months or years. Post-authorization, decision makers want to know: Does protection wane? Do boosters restore it? Which subgroups (e.g., adults ≥65 years, the immunocompromised) need earlier re-dosing? RWD—immunization registries, EHR/claims, laboratory systems, and vital records—lets us answer these questions at scale. But credibility hinges on methods and documentation: explicit protocols and SAPs; auditable data pipelines; bias diagnostics (propensity scores, negative controls); and transparency about laboratory performance and manufacturing quality context. When lab results define outcomes, include analytical capability—e.g., RT-PCR LOD 25 copies/mL and LOQ 50 copies/mL (illustrative), or ELISA IgG LOD 3 BAU/mL and LOQ 10 BAU/mL—so case adjudication is reproducible. To pre-empt “non-biological” confounders in reviewer discussions, keep a short appendix with representative PDE (e.g., 3 mg/day for a residual solvent) and cleaning MACO limits (e.g., 1.0–1.2 µg/25 cm²) demonstrating stable manufacturing hygiene.

Regulators also expect ALCOA (attributable, legible, contemporaneous, original, accurate) for data transformations and outputs, and computerized-system controls (21 CFR Part 11 and EU Annex 11): role-based access, audit trails, validated backups, and time synchronization between sources. Build governance that connects clinical, epidemiology, statistics, safety, and quality—monthly boards reviewing KPIs, pre-declared decision thresholds, and version-locked code. For practical checklists to align SOPs and analysis artifacts, see PharmaRegulatory.in, and mirror terminology used by the European Medicines Agency in post-authorization guidance.

Core VE Designs with RWD: Cohort, Test-Negative, and Case-Control

Cohort designs. Follow vaccinated and comparator groups over time using Cox or Poisson models. Represent time since vaccination (TSV) via restricted cubic splines or pre-specified intervals (0–3, 3–6, 6–9, 9–12 months). Estimate hazard ratios (HR) or incidence-rate ratios (IRR) and convert to VE = (1−HR)×100% or (1−IRR)×100%. Adjust for calendar time, geography, and variant periods; include prior infection and booster status as time-varying covariates. Example (dummy): Adjusted HR for hospitalization 0.35 at 0–3 months → VE 65%; 0.58 at 6–9 months → VE 42%.

Test-Negative Design (TND). Restrict to symptomatic testers; cases are test-positives, controls test-negatives. TND reduces healthcare-seeking bias but assumes similar exposure/testing propensities. Always stratify by symptom criteria and testing policy periods, and run falsification checks (e.g., pre-rollout “VE” ≈ 0%).

Case-control. Useful for rare outcomes (ICU, death). Sample controls densely in time (risk-set sampling) and match on age, sex, geography, and calendar time; analyze with conditional logistic regression. Whatever the design, pre-declare subgroup analyses (≥65, immunocompromised), outcome tiers (ED visit, hospitalization, ICU, death), and decision thresholds that trigger communications or label updates.

Data Linkage & Quality: Turning Heterogeneous Sources into Analysis-Ready Sets

VE lives or dies on linkage. Combine immunization registries (dose dates, products, lots) with EHR/claims (encounters, comorbidities), laboratories (PCR/antigen/serology), and vital statistics (deaths). Use privacy-preserving linkage (hashing, third-party matching) and log deterministic/probabilistic match keys. Build an ETL with validation gates: impossible intervals (dose 2 before dose 1), duplicate vaccinations, outcome-date sanity checks, and cross-source concordance (admit/discharge vs diagnosis timestamps). Version-lock code and containerize (e.g., Docker) so re-runs reproduce hashes. Maintain a data dictionary and MedDRA/ICD-10 mapping under change control. Archive raw snapshots with checksums to satisfy ALCOA’s “original.”

Outcome adjudication must be explicit. Define laboratory thresholds and specimen rules (e.g., accept PCR Ct ≤ 35; resolve discordant antigen/PCR with repeat testing). If using biomarkers in severity tiers, declare the assay performance in the SAP: potency or infection assays with LOD/LOQ values. Keep a short “quality context” memo in the TMF with representative PDE and MACO examples to document that manufacturing and cleaning controls stayed in-spec while clinical effectiveness varied.

Governance, KPIs, and Decision Rules

Stand up a monthly Safety/Effectiveness Board to review dashboards and decide actions. Pre-define KPIs: cohort coverage (% registry-linked to EHR), lag from data cut to dashboard, capture of prior infection, VE at key TSV intervals, and subgroup VE. Quality KPIs include ETL error rate, linkage success, audit-trail review completion, and reproducibility checks (code hash). Establish decision rules such as: “If hospitalization VE in ≥65 years drops >10 points over a quarter with overlapping variant periods and no quality confounder, then recommend booster timing update and prepare HCP comms.” File minutes and decisions with supporting outputs in the TMF.

For hands-on SOP templates covering protocols, ETL validation, and inspection-ready reports, see pharmaValidation.in. Public terminology for post-authorization evidence can be cross-checked on the EMA website.

Modeling Waning & Boosters: Time-Since-Vaccination Done Right

Waning is not a single slope—it varies by age, risk, variant, and outcome. Treat time since vaccination (TSV) as a primary exposure. In Cox models, use restricted cubic splines (3–5 knots) or stepped intervals (0–3, 3–6, 6–9, 9–12 months). Interact TSV with age bands and immunocompromised status. For boosters, apply a biologically plausible grace period (e.g., 7–14 days post-booster) and model booster status as a time-varying covariate. Adjust for calendar time via strata or splines to absorb variant waves and policy changes; include prior infection as a time-varying variable. Report absolute risks (per 100,000 person-months) alongside VE to support policy decisions.

Bias control. Guard against immortal-time bias by aligning person-time precisely around dose dates and grace periods. Use propensity-score weighting/matching with calendar-time strata and geography to reduce confounding by indication. Deploy negative control outcomes (e.g., ankle sprain) and exposures (future vaccination date) to detect residual bias. In TND, vary symptom definitions and exclude occupational screens to test robustness. Where outcomes depend on assays, keep method transparency visible—e.g., RT-PCR LOD 25 copies/mL; LOQ 50 copies/mL—and preserve chain-of-custody. Tie everything back to ALCOA: version-locked code, timestamped cuts, and immutable raw snapshots.

Case Study (Hypothetical): A National VE Program that Drove a Booster Decision

Context. A country links registries, EHR, labs, and vital stats for 2.5 M adults. Findings (dummy). Hospitalization VE in ≥65 years: 68% at 0–3 months post-primary, 52% at 3–6 months, 36% at 6–9 months. Booster lowers HR to 0.28 (VE 72%) in months 0–3 post-booster, stabilizing at VE 60% by months 3–6. TND sensitivity analyses show VE within ±3 points; cohort and case-control designs converge on similar estimates. Negative controls are null; falsification in pre-rollout months ≈0% VE. Labs document analytical capability; adjudication rules are transparent. Quality appendix shows representative PDE 3 mg/day and MACO 1.0–1.2 µg/25 cm²; no manufacturing or cold-chain anomalies are linked to outcome spikes.

Action. The board applies pre-declared rules: “>10-point drop in ≥65s over a quarter with consistent bias checks → recommend booster at 6 months.” HCP materials are updated; an eCTD supplement compiles protocol/SAP, dashboards, and a reproducibility package (container hash, code, parameter files). Public comms explain denominators, absolute risks, and limits. The system continues monthly, ready to detect further waning or variant-specific changes.

Deliverables & Inspection Readiness: Make ALCOA Obvious

Create a simple crosswalk in the TMF: SOP → data cuts → code → outputs → decisions → labels/comms. For each cycle, file (1) protocol/SAP (and addenda), (2) data-cut memo (sources, versions, date), (3) analysis report with TSV curves and subgroup tables, (4) bias diagnostics (balance plots, negative controls), (5) reproducibility pack (code, containers, hashes), and (6) board minutes with decisions. Keep one internal link handy for your teams’ SOPs and validation templates—practitioners often adapt patterns from PharmaSOP.in—and cite a single external reference for public expectations; the ICH Quality Guidelines page is a concise touchstone to align vocabulary on validation and data integrity across functions.

How to Detect Safety Signals After Vaccine Licensure

What “Signal Detection” Means—and the Architecture You Need

After licensure, millions of doses transform rare safety events from theoretical risks into observable data. A signal is a hypothesis—a statistically and clinically plausible association between a vaccine and an adverse event that warrants verification. Detecting it reliably requires a layered architecture: (1) passive spontaneous reports (e.g., national ICSRs) for early pattern recognition, (2) active denominated data (claims/EHR networks) for rate estimation, and (3) targeted follow-up for clinical adjudication. The system must connect methods to governance: a PV System Master File (PSMF), SOPs for coding/triage/escalation, and a standing multidisciplinary review (safety clinicians, epidemiologists, statisticians, quality). Documentation lives in the TMF with ALCOA discipline—attributable, legible, contemporaneous, original, accurate—so an inspector can trace any decision back to raw data and time-stamped actions.

Your design question is not “which method is best?” but “how do we make weak evidence in one stream corroborate in another?” Typical flow: disproportionality screens (PRR, ROR, EBGM) flag vaccine–event pairs in spontaneous reports; observed-versus-expected (O/E) analyses check whether counts in a short, biologically relevant window exceed background; sequential monitoring (e.g., MaxSPRT) controls false positives while watching weekly; and confirmatory designs—self-controlled case series (SCCS) or cohorts—quantify risk. Around the analytics, you must enforce clean inputs: MedDRA version control, ICSR de-duplication, stable case definitions (Brighton Collaboration), and causality recording (WHO-UMC). Finally, keep manufacturing/handling context visible so non-biological drivers are excluded: representative PDE (e.g., 3 mg/day residual solvent) and cleaning MACO (e.g., 1.0–1.2 µg/25 cm²) examples help demonstrate state-of-control while safety is assessed.

Disproportionality 101: PRR, ROR, and Empirical Bayes (EBGM)

Spontaneous reporting systems are rich in narratives but poor in denominators. To screen for unusual reporting patterns, use disproportionality statistics. The Proportional Reporting Ratio (PRR) compares the proportion of a specific Preferred Term among reports for your vaccine versus all others; a typical screen is PRR ≥2 with χ² ≥4 and at least 3 cases. The Reporting Odds Ratio (ROR) offers similar insight with confidence intervals; a 95% CI excluding 1 suggests elevation. Empirical Bayes approaches (e.g., EBGM) shrink noisy estimates toward the overall mean, stabilizing small counts; focus on the lower bound (e.g., EB05 >2) to avoid chasing noise. Statistics do not make a signal by themselves—apply clinical triage: time-to-onset, demographic clustering, and mechanistic plausibility. Document versioned data cuts, coding conventions, and deduplication rules in the TMF.

Build your SOP so screen hits trigger a multi-disciplinary review within a fixed cadence (e.g., weekly). Ensure narratives are adjudicated to Brighton levels where applicable (e.g., myocarditis, anaphylaxis). If diagnostics contribute to “rule-in,” declare their performance so decisions are transparent (e.g., high-sensitivity troponin I LOD 1.2 ng/L; LOQ 3.8 ng/L). For adaptable SOP templates and validation checklists that align with GDP/CSV expectations, see PharmaSOP.in. For public regulator terminology and safety expectations you should mirror in submissions, consult the European Medicines Agency.

Observed vs Expected (O/E): Getting Denominators and Windows Right

O/E asks whether the number of events observed after vaccination exceeds what would be expected from background incidence, given the person-time at risk. Build background rates by age, sex, geography, and calendar time from pre-campaign years; adjust for seasonality (splines or month fixed effects). Choose biologically plausible risk windows (e.g., anaphylaxis Day 0–1; myocarditis Days 0–7 and 8–21). Example calculation (dummy): 1,200,000 doses administered to males 12–29 in one week; background myocarditis 2.1 per 100,000 person-years; expected in 7 days ≈ 1,200,000 × (7/365) × (2.1/100,000) ≈ 0.48. If six adjudicated Level 1–2 cases are observed, O/E ≈ 12.5—an elevation that justifies confirmatory analytics. File the worksheet with assumptions, rate sources, and sensitivity analyses (alternative backgrounds, different lags) to your TMF.

Pre-specify how to handle boosters, dose intervals, prior infection, and competing risks. Keep lot/handling context close at hand. If an excursion or shelf-life question arises, cite representative PDE and MACO controls to show the product remained within manufacturing hygiene expectations while you evaluate temporal patterns.

Sequential Monitoring & Rapid Cycle Analysis: Watching Week by Week

When vaccines roll out rapidly, you need near-real-time surveillance that controls false positives. Rapid Cycle Analysis (RCA) applies repeated looks at accumulating data with statistical boundaries (e.g., MaxSPRT) that preserve overall type I error. Choose cadence (weekly), risk windows, and comparators (historical vs concurrent). Simulate operating characteristics before launch so stakeholders understand power and expected time-to-signal under plausible relative risks (e.g., RR 1.5, 2.0, 4.0). Define “stop/go” criteria in the protocol—e.g., cross the boundary for myocarditis in males 12–29 during Days 0–7, then initiate SCCS and clinical adjudication. Document software versions, parameter files, and outputs with checksums; inspectors will ask how boundaries were set and whether the code that ran matches the code in your validation pack.

RCA does not replace clinical review. Every boundary crossing should trigger case-level adjudication (Brighton levels), causality assessment (WHO-UMC), and a check for data or process artifacts (coding changes, batch updates). Keep a signal log with timestamps, decisions, and owners; file minutes from review boards. Align terminology and escalation thresholds with your Risk Management Plan and labeling sections to avoid inconsistent messaging.

Confirmatory Designs: SCCS and Cohorts That Survive Audit

Self-Controlled Case Series (SCCS) compares incidence in post-vaccination risk windows with control windows within the same individuals, controlling for fixed confounders by design. Specify pre-exposure periods to avoid bias (healthcare-seeking before vaccination), adjust for seasonality, and handle time-varying confounders (infection waves). Cohort studies (vaccinated vs concurrent/historical comparators) are intuitive but demand rigorous confounding control: high-dimensional propensity scores, negative controls, and sensitivity to unmeasured confounding. Pre-state primary endpoints, analysis sets, and missing-data rules; register code and lock it under change control. Example (dummy SCCS output): IRR 4.6 (95% CI 2.9–7.1) for myocarditis Days 0–7 and 1.8 (1.1–3.0) for Days 8–21, with an absolute risk difference 3.4 per 100,000 second doses in males 12–29—clinically relevant even if absolute risk remains low.

Be explicit about how confirmatory results drive decisions: label updates, RMP changes, targeted studies, or additional monitoring. Keep quality context tight—confirm that lots remained in shelf-life and within hygiene controls (PDE and MACO examples) so reviewers do not attribute patterns to manufacturing or cross-contamination. Where diagnostics define cases, include laboratory method performance (e.g., cardiac troponin LOD 1.2 ng/L; LOQ 3.8 ng/L) and chain-of-custody.

Case Study (Hypothetical): From Screen to Confirmed Signal in Six Weeks

Week 1–2: Screen. Passive reports show 18 myocarditis cases clustered in males 12–29 after dose 2; PRR 3.1 (χ² 9.8), EB05 2.4. Week 3: O/E. 1.2 M doses administered to males 12–29; expected in 7-day window ≈0.48; observed 6 adjudicated cases → O/E 12.5. Week 4–5: RCA boundary crossed. MaxSPRT triggers for Days 0–7; immediate clinical adjudication confirms Brighton Level 1–2 in most cases. Week 6: SCCS. IRR 4.6 (2.9–7.1) Days 0–7; IRR 1.8 (1.1–3.0) Days 8–21. Action. Update labeling and RMP, issue HCP guidance, and launch a registry. Quality cross-check. Lots were in specification; monitoring shows cold-chain in range; representative PDE and MACO controls unchanged—supporting a biological, not handling, explanation.

Documentation & Submission: Making ALCOA Obvious

Inspection readiness depends on traceability. Keep a crosswalk that links SOPs → data cuts → code → outputs → decisions. Archive: (1) spontaneous-report screen definitions and deduplication rules; (2) background-rate sources and O/E worksheets; (3) RCA simulation and configuration files; (4) SCCS/cohort protocols, code, and outputs; (5) adjudication minutes with case definitions; (6) quality context (shelf-life, cold-chain, representative PDE/MACO evidence). For the eCTD, place analytic reports in Module 5 and the integrated safety summary in Module 2.7.4/2.5, cross-referencing the RMP. Keep terminology consistent across SOPs, dashboards, and labeling to avoid inspector confusion.

Key Takeaways

Signals are hypotheses, not verdicts. Use a layered approach—disproportionality to sense, O/E to anchor, sequential monitoring to watch, and SCCS/cohorts to confirm. Surround analytics with clinical adjudication, causality assessment, and manufacturing/handling context (PDE, MACO, and assay LOD/LOQ where relevant). Document everything with ALCOA discipline. Done well, your signal detection system protects patients, preserves trust, and accelerates clear, defensible decisions.