Peptide basics

How to read an adverse-event database without drawing the wrong conclusion

This is a methods page. It explains what these records are, what they can support, and the specific mistakes that produce confident wrong answers. It makes no claim about whether any compound is safe or harmful, and it gives no advice about symptoms — that requires someone who can examine you.

What the record actually is

FDA's adverse event system — currently being folded into a broader Adverse Event Monitoring System — is a store of reports submitted by manufacturers, health professionals and members of the public. Manufacturers are required to forward what reaches them under 21 CFR 314.80; everyone else reports voluntarily through MedWatch, or not at all.¹¹¹²

Nothing in that pipeline verifies the event. FDA's own limitations note lists duplicate and incomplete reports, states that the existence of a report does not establish causation, states that the information has not been medically confirmed, and states that occurrence rates cannot be established from the reports. It adds that the data by themselves are not an indicator of a product's safety profile.¹

Mistake one: treating a count as a rate

A count answers how many reports mention something. A rate answers how often it happens to people who take the drug. Converting the first into the second requires knowing how many people took it and for how long, and the database contains no such field. It has no record of the people who took the product and reported nothing.¹²

The size of that missing group is not merely unknown, it is large. Hazell and Shakir's review of 37 studies from 12 countries found a median under-reporting rate of 94% (interquartile range 82–98%), and still 85% when restricted to studies of specific serious drug–event pairs. Because the shortfall varies by setting, seriousness and product, it cannot be corrected for.⁶

Mistake two: assuming the query found the drug

The searchable field holds the product name as written on the report. Semaglutide is sold in the United States as Ozempic, Wegovy and Rybelsus, and reports arrive under all of those strings plus the generic. In the extract dated 28 April 2026, a search for SEMAGLUTIDE returned 6,618 reports; a search for OZEMPIC alone returned 57,104. Same molecule, an order of magnitude apart, purely as a function of which word the query asked for.³⁴

Combining the four strings in one query returned 83,614 reports in that extract — which is the closest a simple query gets to the molecule, and still not a clean total, because the union may double-count reports naming more than one of the strings and misses reports that spelled the product some other way.⁵

• A generic-name query silently excludes the brand-name reports, which are usually the larger share.
• A brand-name query silently excludes every other brand of the same molecule.
• Compounded and counterfeit products are named inconsistently and may not match any expected string.
• Two queries with different totals are not two findings; they are the same record cut two ways.²³⁴⁵

Mistake three: reading a rising line as a rising risk

Report volume responds to attention. Pariente and colleagues tested four safety alerts in the French national database and found reporting of the alerted event jumped afterwards — including reports of events that had happened before the alert. For strokes with atypical antipsychotics the reporting odds ratio went from 0.10 before the alert to 1.10 after it, with one report before and 16 after. Nothing about the drugs changed in between.⁷

de Boissieu and colleagues quantified how much of a headline signal that bias can account for. For osteonecrosis of the jaw with bisphosphonates the raw reporting odds ratio was 3448 (95% CI 1413–8417); once reports that also named a known alternative risk factor were reassigned, the ratio under the maximum-bias assumption was 87 (95% CI 63–121). Both numbers came from the same 148 reports.⁹

Time since launch matters too, though not in the way it is usually stated. Hoffman and colleagues analysed 334,984 reports across 62 drugs approved between 2006 and 2010 and found that most did not follow the classic pattern of reporting peaking around the second year; the common shape was a rise over the first three quarters and a roughly flat count after that. Whichever curve applies, it is a curve of reporting behaviour.¹⁰

Mistake four: trusting the statistic to fix the data

Analysts often move from raw counts to disproportionality measures such as the proportional reporting ratio, which compares how often a term appears with a given drug against how often it appears in the rest of the database. This controls for overall database growth. It does not create a denominator, and it does not remove the biases above.⁸

Moore and colleagues worked through sertindole, an antipsychotic suspended in 1998 after the proportion of reports of fatal arrhythmia-suggestive reactions came out roughly ten times higher than for other atypical neuroleptics in the UK. When actual death rates were examined in prescription event monitoring and a large retrospective cohort, the reporting pattern showed signs of skew. A disproportionality result is a hypothesis about where to look next, and the looking is done elsewhere.⁸

A short checklist for any FAERS figure you meet

1. Is the exact query published, including which product names were searched and the extract date? If not, the number cannot be checked.
2. Is it presented as a count of reports, or has it been turned into a percentage or a rate? A rate is a red flag on its own.
3. Is it being compared with another drug's count? Two drugs have different user numbers, marketing histories and reporting patterns, none of which the comparison holds constant.
4. Does the term describe an event or a behaviour? Terms such as off label use and incorrect dose administered are not physiological outcomes.
5. Has anything changed in the news, in litigation advertising or in prescribing over the window shown? All three move report volume.¹⁷⁸¹⁰

Semaglutide is the worked example throughout this page because it is the case where the naming problem is largest: four product strings in the US market, and a tenfold gap between the generic query and the biggest brand query in the same extract.

Common questions