The FAERS Database: Its Role in Drug Safety and Limitations

The FDA Adverse Event Reporting System, known as FAERS, is the United States government’s main repository for reports of bad reactions to marketed drugs and therapeutic biologics. It collects what are formally called individual case safety reports, submitted by healthcare professionals, patients, manufacturers, and lawyers, and stores them in a searchable database that now contains millions of records.¹ FAERS has helped trigger drug withdrawals, label changes, and safety warnings, but it also has structural weaknesses that make its data easy to misread, and a growing body of low-quality research based on the publicly available dashboard is raising alarm among pharmacovigilance experts.

What FAERS Actually Does

After a drug reaches the market, the controlled environment of a clinical trial is over. The drug is now being taken by a much larger and more diverse population, often alongside other medications, in people whose health conditions were exclusion criteria during trials. FAERS exists to catch safety problems that emerge during this messy real-world phase. When a patient or clinician suspects a drug caused an adverse event, a report can be filed with the FDA either directly or through the drug’s manufacturer. Manufacturers themselves are required by law to submit reports they receive; for everyone else, reporting is voluntary.

Each report in the database includes information about the suspected drug, the adverse event experienced, patient demographics when available, other medications the patient was taking, and the reported outcome. These reports are not verified clinical records. Nobody confirms the diagnosis, checks whether the patient actually took the drug as described, or rules out other explanations. The system is intentionally permissive: the FDA would rather receive a noisy flood of reports than miss a genuine safety signal because the reporting bar was too high.

How Safety Signals Are Found

Raw report counts alone do not tell you much. A widely prescribed drug will naturally accumulate more adverse event reports than a niche one, and common side effects like nausea will show up more often than rare ones simply because they are common. To separate real signals from background noise, researchers and FDA reviewers use a set of statistical methods known collectively as disproportionality analysis. The idea is straightforward: you check whether a particular drug-event combination shows up more often in the database than you would expect based on how frequently both the drug and the event appear separately.

Four methods dominate the published literature. The Reporting Odds Ratio and the Proportional Reporting Ratio are relatively simple calculations that flag drug-event combinations appearing at a higher-than-expected rate, with a signal considered positive when the lower bound of the confidence interval exceeds one and at least three reports exist.² The other two methods use Bayesian statistics to shrink estimates toward zero, which helps avoid false alarms when reports are few. These go by cumbersome names, but the logic is the same across all four: if a drug-event pair occurs far more often than expected, something may be going on.³⁴

A disproportionality signal is not proof that a drug causes harm. It is a flag that says “this deserves a closer look.” The FDA may then investigate with other data sources, review the individual case reports in detail, consult with advisory committees, or commission new studies. Some signals lead to label changes or black box warnings. Many turn out to be artifacts of reporting patterns.

When FAERS Has Caught Real Problems

The system’s track record includes some genuinely important catches. One well-documented example involves a class of diabetes drugs called SGLT2 inhibitors. The FDA identified 55 unique cases of Fournier gangrene, a severe and life-threatening infection of the genital area, in patients receiving these drugs between March 2013 and January 2019.⁵ The condition is rare enough that 55 cases in a specific drug class stood out sharply against the background rate. This finding led to a safety communication and label updates, warning both patients and prescribers about a risk that had not been visible during clinical trials.

Cases like this illustrate the system’s core value: rare events that affect one in ten thousand or one in a hundred thousand users are almost impossible to detect in pre-approval trials of a few thousand people. A passive surveillance system that covers millions of prescriptions can spot patterns that would otherwise stay hidden for years. The tradeoff is that the system sees everything through a fog of incomplete data and uncontrolled confounders.

What FAERS Cannot Tell You

The single most important limitation is also the most frequently misunderstood: FAERS cannot establish that a drug caused a reported event. A report in the database means someone believed a drug might have been involved. The patient may have had the condition already. Another medication may have been responsible. The timing may have been coincidental. Because reports are not verified, the database is better understood as a collection of suspicions than a collection of confirmed drug reactions.⁶

There is also no denominator. FAERS knows how many reports exist for a drug, but not how many people are taking it. If Drug A has 500 reports and Drug B has 50, that might mean Drug A is more dangerous, or it might mean Drug A has ten times as many users. Without prescription volume data, you cannot calculate a true incidence rate. Researchers who want to estimate actual risk have to bring in external data sources like insurance claims databases or electronic health records, and even then the exercise involves substantial uncertainty.⁶

The Duplicate Report Problem

Duplicate reports are a persistent headache. The same adverse event can be submitted by a patient, their doctor, and the drug manufacturer, each with slightly different details. An evaluation of duplicate reports in FAERS found that among confirmed duplicate sets, only about 16% had all seven key data elements coded identically. The narrative text was much more consistent, with a median similarity score of 0.87 between duplicates compared to 0.48 between unrelated reports.⁷ In other words, the structured fields in duplicate reports often disagree while the free-text descriptions closely match, which makes automated deduplication tricky. Failing to remove duplicates inflates the apparent frequency of a drug-event combination, which can generate false safety signals or exaggerate real ones.

Who Files Reports and Why It Matters

The mix of reporters shapes the data in ways that are not always obvious. Manufacturers submit reports they receive through their own pharmacovigilance programs, but these may be filtered, delayed, or reflect the specific concerns that prompted patients to contact the company. Healthcare professionals tend to report events they consider clinically significant. Consumers who report directly to the FDA bring a different perspective, and contrary to what you might assume, a study found that the completeness of adverse event reports from consumers was generally greater than that of reports from healthcare professionals.⁸

The content of reports also differs by reporter type. Consumer reports have a higher proportion of female patients and more frequently list disability as the outcome. Consumers also tend to report more concomitant medications than healthcare professionals do.⁹ These differences are not necessarily biases in the negative sense; consumers are reporting what happened to them, while clinicians are reporting what they diagnosed. But the practical consequence is that the database’s portrait of a drug’s safety profile shifts depending on who is doing the reporting, and any analysis that does not account for reporter type is working with a distorted picture.

The Weber Effect, or Lack Thereof

For decades, pharmacovigilance textbooks described what is called the Weber effect: a predictable spike in adverse event reporting during the first two years after a drug’s approval, followed by a steady decline regardless of whether the drug’s actual safety profile changed. The concern was that this pattern could confuse signal detection, making new drugs look temporarily more dangerous than they are. An analysis of 62 drugs approved between 2006 and 2010, covering nearly 335,000 primary suspect reports, found that this pattern no longer holds in modern FAERS data. Reporting volume increased over the first few quarters after launch, as expected, then stayed relatively constant rather than declining.¹⁰

A separate study focused specifically on oncology drugs confirmed the finding from a different angle: among seven cancer drugs examined, five distinct reporting patterns were observed, but none matched the classic Weber curve. Only one drug, cetuximab, showed anything resembling a second-year peak, and even that was not followed by continuous decline.¹¹ The takeaway is that the Weber effect is probably an outdated concern, at least for drugs approved in the modern era. Reporting dynamics have changed, likely because of electronic submission systems, increased awareness, and manufacturer reporting obligations that keep the flow of reports more consistent over time.

Pediatric and Pregnancy Surveillance

FAERS was built primarily around adult medicine, and its coverage of special populations reflects that. From 2010 to 2020, only about 3% of all reports in the database described pediatric patients. Among those pediatric reports, the pattern was striking: younger children had higher proportions of serious outcomes, with neonates reaching 96% serious, compared to roughly 53% for adolescents. The most common reported event for pediatric patients was off-label use, underscoring a chronic problem in children’s pharmacology: many drugs prescribed to children were never formally studied or approved for that age group.¹²

Pregnancy-related reporting faces similar constraints. Pregnant women are routinely excluded from clinical trials, so postmarketing surveillance becomes one of the few tools for understanding drug risks during pregnancy. A pharmacovigilance study of statins during pregnancy found 477 cases of pregnancy-related adverse events submitted by healthcare professionals. Most statins did not show a disproportionate signal for abortion or stillbirth, but specific drugs did flag for particular outcomes: lovastatin showed increased fetal complication risk, and pravastatin was associated with elevated signals for preterm birth and low birth weight.¹³ Findings like these are not proof of causation, but they point clinicians toward questions that deserve more rigorous study.

How FAERS Overlaps with Global Systems

FAERS is not the only adverse event database in the world. The European Union maintains EudraVigilance, and the World Health Organization operates VigiBase, which aggregates reports from more than 130 countries. An investigation comparing safety signals across these three systems found substantial overlap but not perfect agreement. At the broadest categorization level, the overlap between the European system and FAERS or VigiBase was about 98%. At more granular levels, the overlap dropped to around 85%.¹⁴ Differences in marketing authorizations and regional prescribing patterns partly explain why a signal might appear in one database but not another. A drug that is widely used in Europe but has limited penetration in the US may generate a signal in EudraVigilance long before FAERS has enough reports to flag the same problem. For regulators, this means no single database tells the whole story.

The Growing Problem of Misinterpretation

In recent years, the FDA has made FAERS data publicly available through a user-friendly online dashboard. The intention was transparency. The unintended consequence has been an explosion of low-quality research. Thousands of published papers now use the dashboard to run simple disproportionality analyses, often without adequate understanding of the data’s limitations. These papers produce statistical associations that get presented as “safety signals,” creating alarm that may not be scientifically grounded and that can influence how doctors prescribe and how patients take their medications.¹⁵

A case study using VMAT2 inhibitors, a class of drugs used for movement disorders, illustrated some of the pitfalls. Researchers found that factors in data quality and differences in manufacturer pharmacovigilance programs can dramatically skew the results of analyses based on the public dashboard. Safety data drawn from the dashboard needs to be interpreted in the context of what is already known about each drug’s safety profile, not taken at face value.¹⁶ The problem is not the data itself but the gap between what the data can support and what people conclude from it. A disproportionality signal from FAERS is a starting point for investigation, not an endpoint.

Do Black Box Warnings Change Reporting Patterns?

When the FDA issues a black box warning, the most serious type of safety communication, the question arises whether that action changes how adverse events get reported. A study examining reporting patterns before and after black box warning updates found that about half of new or major updates were followed by a meaningful increase in the proportion of relevant adverse event reports. For minor updates, only about a quarter showed such an increase.¹⁷ This suggests that major warnings do raise awareness and prompt more focused reporting, but the effect is far from universal. Minor label tweaks often go unnoticed by the reporting community, which means the database may not reliably capture the downstream effect of incremental regulatory actions.

Machine Learning and Where the System Is Headed

The sheer volume of reports in FAERS, millions and growing, has pushed researchers toward computational approaches. Machine learning models are being developed to help assess causality in individual reports, a task that has traditionally relied on manual expert review. One line of research explored using natural language processing on the free-text narratives within reports, combined with information from external data sources, to train supervised learning models that predict how reports should be classified.¹⁸ The duplicate report problem described earlier is another target for automated tools, since the narrative text is often the most reliable way to match reports that have discordant structured fields.

Beyond improving how FAERS data gets processed, the broader trajectory of pharmacovigilance is moving toward integrating multiple data streams. Electronic health records, insurance claims databases, social media monitoring, and wearable device data are all being explored as complements to traditional spontaneous reporting.¹⁹ Electronic health records, for instance, provide the denominator that FAERS lacks: you know how many patients received a drug, not just how many reported a problem. Social media surveillance picks up patient-reported experiences that never make it into formal reporting channels. None of these sources alone is sufficient, but linking them could address many of the weaknesses that have defined spontaneous reporting systems for decades.

Estimating Economic Costs from FAERS Data

Some researchers have attempted to go beyond counting events and estimate the financial burden associated with drug-related adverse events captured in FAERS. One approach mapped adverse event codes from FAERS reports to medical cost data from government healthcare surveys, creating a drug safety rating system based on the estimated costs of the serious outcomes reported for each drug.²⁰ The idea is appealing: if you could attach a dollar figure to a drug’s postmarketing safety profile, insurers and health systems would have another tool for comparing treatment options. But the same limitations that make FAERS unreliable for incidence calculations apply here too. Without knowing how many patients took a drug, cost-per-user estimates are speculative. Still, these efforts represent an attempt to extract more practical value from a database that, for all its flaws, remains one of the largest collections of real-world drug safety data anywhere.