Spontaneous Reporting Biases

In plain language

Spontaneous safety reports are valuable warnings, but they are not a count of everyone who had the side effect. Some events are never reported, some are reported several times, and reporting can surge after a label warning, lawsuit, media story, or new-product launch. Analysts should use these reports to find and investigate signals, not to calculate incidence or rank drugs by raw counts.

Spontaneous reporting biases

are the predictable distortions created when adverse-event databases are built from suspected reports rather than from a defined population under complete follow-up. FAERS, VigiBase, EudraVigilance, and company safety databases are indispensable for early warning, rare serious case review, and disproportionality screening, but they are not census systems. The numerator is incomplete, selective, duplicated in places, and influenced by clinical awareness, regulation, media, litigation, product age, reporter type, and manufacturer follow-up.

Core conceptual distinction

Reporting bias is not one flaw; it is a family of mechanisms that affect different decisions differently. Under-reporting means most actual adverse reactions are never submitted. Selective reporting means serious, unusual, new, or well-publicised events are more likely to be submitted than mild, expected, or common events. Notoriety or stimulated reporting means reports increase after safety alerts, label changes, litigation, social media, publications, or regulatory attention. Weber/time-on-market effects mean reporting intensity can peak early after launch and decline as familiarity grows. Duplicate and follow-up reporting means one underlying case can appear several times. Masking and competition bias mean a very strong product-event or class-event pattern can distort the database margins used by disproportionality methods and hide or exaggerate other pairs.

Pros, cons, and trade-offs

- vs ignoring reports entirely: Biases do not make spontaneous reports useless. They make them hypothesis-generating rather than incidence-estimating. Use them for early warning, rare narratives, and signal triage; move to claims, EHR, registry, or designed follow-up when risk magnitude is needed. - vs disproportionality statistics: PRR, ROR, IC, and EBGM partially standardize within the reporting database, but they do not remove stimulated reporting, missing denominators, confounding by indication, event masking, or duplicate case submission. A high ROR can be a reporting artefact; a low ROR can be masking. - vs denominator-based RWE: Claims/EHR cohorts bring person-time and comparator control, but can miss clinical detail and rare unexpected events that ICSRs capture quickly. Spontaneous reports are the scout; denominator-based RWE is the risk quantifier. - vs case-level causality assessment: A single well-documented report can be compelling despite population-level reporting bias. The bias problem is mainly about aggregate interpretation, raw counts, and reporting ratios, not about dismissing every individual case.

When to use

Use this concept whenever interpreting raw spontaneous report counts, reporting rates, disproportionality results, signal timelines, pre/post safety communication trends, cross-country comparisons, new-launch surveillance, or safety claims based on FAERS, VigiBase, EudraVigilance, JADER, VAERS, or company ICSR databases. It should be part of every signal validation packet and every methods section for spontaneous-reporting analyses.

When NOT to use - and when it is actively misleading

- Do not use "reporting bias exists" as a blanket reason to ignore a serious plausible signal. Bias is a reason to structure the review, not to dismiss evidence. - Do not calculate incidence, prevalence, absolute risk, or comparative safety from spontaneous report counts without an external denominator and explicit assumptions. - Do not compare raw counts across products, countries, or calendar periods as safety rankings. Utilisation, launch age, local reporting rules, stimulated reporting, and duplicate handling can dominate the count. - Do not interpret a decline in reports after a risk minimisation measure as effectiveness unless reporting intensity, exposure, denominator, and case ascertainment are addressed. - Do not assume disproportionality controls confounding. Indication, co-medication, disease severity, product channeling, and differential outcome recognition can still drive the reporting association.

Data-source operational depth

- FAERS/FDA public extracts: Deduplicate by case/version logic before analysis; normalize products to ingredients; distinguish primary suspect, secondary suspect, interacting, and concomitant roles; stratify by reporter type and serious/non-serious status. Public extracts omit full narratives, limiting assessment of alternative causes. - EudraVigilance and VigiBase: Cross-country reporting culture, regulatory requirements, duplicate detection, product coding, and access rules affect analyses. Do not pool regions blindly when a safety alert or local reporting obligation affected only some countries. - Company safety databases: Manufacturer follow-up can improve case detail but reporting is influenced by product maturity, patient support programmes, call centres, solicited sources, literature surveillance, and affiliate practices. - Solicited, registry, literature, and patient-support reports: These are not purely spontaneous. They may have active follow-up, known denominators for a programme, or publication selection. Keep report source as a stratifier and do not mix with spontaneous reports without justification. - Claims/EHR linkage or follow-up: Use denominator-based data to test whether a reporting signal corresponds to a real incidence or comparative risk signal. Claims/EHR have their own biases, but they solve the missing-denominator problem.

Worked example

Drug A launches with a strong educational campaign about liver injury. In the first six months, FAERS receives 120 liver injury reports for Drug A and only 40 for older Drug B. A naive analyst concludes Drug A is three times riskier. A pharmacovigilance analyst checks exposure, launch timing, case versions, seriousness, label status, media, and reporting source. Drug A has three times the number of new users, a label warning that asks clinicians to report liver cases, and 25 duplicate follow-up versions. After deduplication and stratification, Drug A still has several serious cases with compatible timing and positive dechallenge, so the signal is not dismissed. But the aggregate count is no longer interpreted as comparative risk. The correct conclusion is: spontaneous reporting supports clinical signal evaluation and denominator-based follow-up, not a direct claim that Drug A's incidence is three times Drug B's.

Worked example

Scenario

A safety team compares spontaneous reports for liver injury after Drug A's launch with an older comparator Drug B. The raw count appears alarming, but reporting artefacts and utilisation differences change interpretation.

Dataset

Simplified spontaneous-report bias review.

Steps

• Deduplicate initial and follow-up reports before interpreting counts.

• Check exposure or utilisation context, even if it is external to the reporting database.

• Stratify by report source, reporter type, seriousness, region, and time since launch.

• Review whether safety alerts, label changes, media, litigation, or education stimulated reporting.

• Treat remaining clinically strong cases as signal evidence, but do not estimate incidence from the report count.

Result

Drug A's raw count is not a comparative risk estimate. The evidence supports signal evaluation and a denominator-based follow-up study.