Randomized controlled trials establish whether a drug works and is reasonably safe under controlled, monitored conditions with a defined population. They are not well-suited to catching rare adverse events, understanding how a drug performs across the full diversity of real-world patients, or tracking what actually happens once a medication scales to millions of users outside a trial's structure. That's the specific gap real-world evidence registries and pharmacovigilance databases are built to fill — and for GLP-1 medications, several distinct data sources are now generating a genuinely large post-marketing evidence base.
What the major data sources actually are
- FAERS (FDA Adverse Event Reporting System) — a spontaneous reporting database where clinicians, patients, and manufacturers submit adverse event reports; used for disproportionality analysis (comparing how often an event is reported for one drug versus others) rather than measuring true incidence
- TriNetX — a federated network of de-identified electronic health records across many health systems, used for retrospective cohort and case-control studies with actual denominators (total patients exposed), unlike FAERS
- The NIH All of Us Research Program — a large, diverse longitudinal cohort integrating EHR data with other data types including wearable device data, enabling studies that connect prescribing to a broad range of downstream diagnoses
- Commercial claims and EHR datasets (Truveta, Epic Cosmos, Blue Health Intelligence, Prime Therapeutics) — used heavily for adherence, persistence, and health economics research
Safety signals that emerged post-marketing
A February 2026 narrative review synthesizing FDA and EMA regulatory databases alongside pharmacovigilance reports identified rare but serious post-marketing safety signals including acute pancreatitis, intestinal obstruction, thyroid C-cell hyperplasia, acute kidney injury, and transient worsening of diabetic retinopathy.[1] More recent reports from 2023–2025 identified additional signals — perioperative aspiration risk, neuropsychiatric events, and mild tachycardia — that prompted updated product labeling.[1]
A specific example: what a real registry study looks like, and its limits
A real-world pharmacovigilance analysis comparing gynecological hemorrhagic events between tirzepatide and semaglutide, using FAERS data from 2022–2025, found no disproportionate signal indicating tirzepatide carries greater risk than semaglutide for these events.[2] But the study's own methodology section flagged a specific limitation worth understanding broadly: 94.6% of tirzepatide adverse event reports came directly from consumers, versus 53.4% for semaglutide — a substantial reporting-source disparity that complicates direct comparison, since consumer-submitted reports and clinician-submitted reports often differ in detail and follow-up completeness.
Why this methodological detail matters generally
FAERS-based disproportionality analyses are useful hypothesis-generating tools, not incidence measurements — there's no reliable denominator of how many total patients were exposed to each drug, and reporting behavior itself can vary by drug, time period, media attention, and reporter type. A "no signal detected" finding in FAERS means the data didn't show a disproportionate reporting pattern; it does not mean a risk has been ruled out with the same confidence a randomized trial would provide.
An unexpected finding from a broader phenome-wide study
A study using the All of Us Research Program (n=18,746 adults with type 2 diabetes) compared a wide range of downstream diagnoses following GLP-1RA prescription against SGLT2 inhibitor and DPP-4 inhibitor prescriptions.[3] While prior research had reported reduced substance use and psychotic disorder diagnoses among GLP-1RA users, this analysis found an increased risk of dysthymic disorder diagnosis relative to SGLT2 inhibitor users (intention-to-treat hazard ratio 2.30, 95% CI 1.53–3.46) — a specific, quantified finding that illustrates why phenome-wide registry approaches can surface signals neither trials nor narrower safety studies were designed to detect.[3]
The long-term effectiveness of GLP-1s outside controlled trials remains unclear.
Why adherence data matters as much as safety data
Registry-based persistence data has consistently found real-world adherence substantially below what trial completion rates would suggest — a 2021 commercial claims study of over 16 million people found only 36% one-year adherence for Wegovy initiators and 47% for Ozempic; a 2025 follow-on study found just 14.3% remained on therapy at two years.[4] This matters clinically because most of the dramatic weight-loss figures reported in trials reflect sustained use over the full trial period — real-world effectiveness at a population level depends heavily on how many patients actually remain on treatment long enough to reach those outcomes.
Registry studies generally cannot establish causation with the same confidence as randomized trials. Patients who discontinue a medication, or who are prescribed one drug over another, differ systematically from those who don't — confounding that propensity matching and other statistical techniques can reduce but not eliminate entirely.
Compare GLP-1 telehealth providers
Source GLP-1 tracks pricing, licensing, and provider structure across the market.