FDA Boxed Warnings: Evidence Patterns Behind Labeling Action
For regulatory affairs and pharmacovigilance teams, the boxed warning is FDA's most consequential labeling action — one that can reshape prescribing, trigger REMS requirements, and signal class-wide scrutiny. The useful question is less a complete product list than the evidentiary threshold that moves the agency to act.
The analysis below organizes representative boxed-warning cases by the type of evidence that drove the labeling change — post-marketing trials, meta-analyses, animal carcinogenicity data, and spontaneous adverse-event reports — and identifies the product, the risk, and the data the agency relied on.
Want to ask Rhizome your own regulatory questions? Try it for free.
FDA boxed warnings: evidence patterns that trigger labeling action
The boxed warning (colloquially the "black box") is the strongest safety statement FDA can require on a prescription drug or biologic label. It sits at the top of the prescribing information and signals a risk serious enough that prescribers must weigh it before every prescription. The instructive question for regulatory affairs teams is not simply which products carry one, but what class of evidence moved FDA to require it. The answer varies widely: some warnings rest on a single large randomized safety trial, others on pooled meta-analyses, animal carcinogenicity data, spontaneous post-market reports, or a class-wide signal accumulated over decades.
The examples below are organized by the type of evidence that triggered the warning, because that is the pattern worth internalizing. Each entry names the product, the warning content, and the specific data the labeling itself cites as the basis.
Post-marketing randomized safety trials
The clearest modern template is a mandated post-approval safety study that reads out against an active comparator.
Tofacitinib (Xeljanz / Xeljanz XR), a JAK inhibitor. The boxed warning covers five distinct risks: serious infections leading to hospitalization or death; higher all-cause mortality including sudden cardiovascular death; malignancies including lymphoma and lung cancer; major adverse cardiovascular events (MACE, defined as cardiovascular death, non-fatal MI, and non-fatal stroke); and thrombosis (pulmonary embolism, deep vein thrombosis, and arterial thrombosis) 34. The trigger was a large randomized post-marketing safety study in rheumatoid arthritis patients aged 50 and older with at least one cardiovascular risk factor, identified in the label as "RA Safety Study 1," which compared tofacitinib 5 mg or 10 mg twice daily against TNF blockers 1578. In that study, tofacitinib showed higher all-cause mortality including sudden cardiovascular death 1, higher rates of malignancy excluding non-melanoma skin cancer 78, higher MACE in the at-risk population 78, and an increased incidence of thrombotic events versus TNF blockers 56. This is the paradigm of a warning built on head-to-head post-market trial data rather than spontaneous reports.
Pooled analyses and meta-analyses of controlled trials
When no single trial is decisive, FDA has required boxed warnings on the strength of pooled or meta-analytic signals across many studies.
Antidepressants (fluoxetine and the SSRI/SNRI class), suicidality. The warning states that patients of all ages started on antidepressants may experience worsening depression or emergence of suicidal ideation and behavior, and that the risk is increased in children, adolescents, and young adults aged 18 to 24 4344. The evidentiary basis cited in the label is explicitly pooled: 24 short-term placebo-controlled trials of 9 antidepressants in more than 4,400 pediatric patients, plus 295 short-term trials of 11 antidepressants in more than 77,000 adults 4344. The label quantifies the risk difference per 1,000 patients treated: 14 additional cases under age 18, 5 additional cases in ages 18 to 24, roughly 1 fewer in ages 25 to 64, and 6 fewer in patients 65 and older 4344. The gradient by age, drawn entirely from pooled trial data, is what shaped both the warning and its age-specific framing.
Rosiglitazone (Avandia), myocardial ischemia. The boxed warning addresses two cardiovascular risks. For congestive heart failure, it warns the drug can cause or worsen CHF and is contraindicated to initiate in NYHA class III or IV heart failure 72. For myocardial ischemia, the label cites a meta-analysis of 42 clinical trials (mean duration about 6 months, 14,237 patients) that found an increased risk of ischemic events such as angina and MI, with an odds ratio of 1.4 (95% CI 1.1 to 1.8) versus pooled comparators 6972. Notably, the label also concedes the picture is unresolved: three longer prospective trials (ADOPT, DREAM, and RECORD) neither confirmed nor excluded the risk, so the overall evidence is described as inconclusive 6872. The heart-failure component draws additionally on a 52-week placebo-controlled echocardiographic study in NYHA class I/II patients 67 and five 26-week trials of rosiglitazone added to insulin 70. Avandia illustrates that FDA will act on a meta-analytic signal even when confirmatory trials are equivocal.
Class randomized-trial data combined with epidemiology
Some warnings synthesize controlled-trial results with real-world observational data, and are then applied across an entire drug class.
NSAIDs (celecoxib and the class), cardiovascular and gastrointestinal risk. The boxed warning has two arms. The cardiovascular arm states that NSAIDs increase the risk of serious thrombotic events, including fatal MI and stroke, with risk beginning as early as the first weeks of treatment and rising at higher doses 111113115130. The gastrointestinal arm warns of serious GI inflammation, bleeding, ulceration, and perforation, which can be fatal and can occur without warning symptoms; the label notes serious upper-GI events in roughly 1% of patients treated for 3 to 6 months and 2% to 4% at one year 111115130. The cited evidence blends trial and epidemiologic sources: the APC (Adenoma Prevention with Celecoxib) trial showed roughly a threefold increase in the composite of CV death, MI, or stroke versus placebo 111113115; the PRECISION trial established non-inferiority of celecoxib to naproxen and ibuprofen on the composite CV endpoint 111113115; CABG surgery trials of a COX-2 selective NSAID found increased MI and stroke 111113115; and Danish National Registry data showed post-MI NSAID use raised reinfarction and death, with a first-year death rate of 20 per 100 person-years on NSAIDs versus 12 without 111113115.
Opioid analgesics (oxycodone and combinations). The class boxed warning bundles addiction, abuse, and misuse leading to overdose and death; life-threatening respiratory depression; neonatal opioid withdrawal syndrome after prolonged use in pregnancy; profound sedation, respiratory depression, coma, and death when combined with benzodiazepines, other CNS depressants, or alcohol; and altered oxycodone exposure with CYP3A4 inhibitors or inducers 131132152. These warnings reflect the established pharmacology of the class layered onto the population-level harm signal of the opioid crisis, and are now standard across opioid labeling.
Pivotal-trial adverse-event rates for novel modalities
For first-in-class products, the boxed warning often codifies the toxicity rates seen in the registrational trials themselves.
Tisagenlecleucel (Kymriah), a CD19 CAR-T cell therapy. The boxed warning covers cytokine release syndrome (CRS), including fatal or life-threatening reactions, and neurologic toxicities, which can be severe or life-threatening and can occur concurrently with CRS 167. The basis is the pivotal trial experience recorded in the label: in pediatric and young-adult relapsed/refractory B-cell ALL, CRS occurred in 77% of patients (Grade 3 in 48%) and neurologic toxicities in 71% (Grade 3 in 22%) 160; adult DLBCL and follicular lymphoma cohorts showed similarly high CRS and neurotoxicity rates 159161. For cell and gene therapies, boxed warnings routinely quantify these on-target toxicities directly from the approval dataset.
Post-marketing spontaneous reports
A boxed warning can also be driven by accumulated adverse-event reports after approval, without a controlled trial designed to test the signal.
Montelukast (Singulair), neuropsychiatric events. The warning states that serious neuropsychiatric events have been reported, that the mechanisms are not well understood, and that the benefits may not outweigh the risks in some patients, particularly where symptoms are mild and alternatives exist 101107. The label is explicit that the warning rests on post-marketing reports, some with clinical details consistent with a drug-induced effect 101104105108. The events named span agitation, aggression, anxiety, depression, hallucinations, insomnia, irritability, memory impairment, obsessive-compulsive symptoms, sleep-walking, tremor, and suicidal thoughts and behavior including completed suicide 101103106109110. Montelukast is a useful reminder that a robust spontaneous-report signal, absent a dedicated trial, can support the strongest labeling action.
Nonclinical (animal) carcinogenicity data
Occasionally the pivotal evidence is preclinical, with the warning framed around uncertain human relevance.
Semaglutide (Ozempic), thyroid C-cell tumors. The boxed warning states that semaglutide causes thyroid C-cell tumors in rodents and that it is unknown whether it causes such tumors, including medullary thyroid carcinoma, in humans 747681. The basis is animal carcinogenicity: a dose- and duration-dependent increase in thyroid C-cell tumors in mice and rats at clinically relevant exposures 7677. The 2-year mouse study showed a statistically significant rise in C-cell adenomas, and the 2-year rat study showed significant increases in both C-cell adenomas and, in males, carcinomas 86879298. The label states plainly that the human relevance could not be determined from clinical or nonclinical data 86879298. This is the template for the GLP-1 receptor agonist class: a warning grounded in rodent data with explicitly unresolved human relevance.
Class effects and cumulative-dose relationships
Some warnings capture a hazard that grows with duration or total exposure, applied across every product sharing the mechanism.
Fluoroquinolones (levofloxacin and the class). The warning flags disabling and potentially irreversible serious adverse reactions that can occur together: tendinitis and tendon rupture, peripheral neuropathy, and central nervous system effects, plus exacerbation of muscle weakness in myasthenia gravis, for which the class should be avoided 34373841. The "disabling and potentially irreversible" framing signals a class-wide safety posture accumulated across products and years of use.
Metoclopramide, tardive dyskinesia. The boxed warning states the drug can cause tardive dyskinesia, a serious and often irreversible movement disorder with no known treatment, and directs that treatment not exceed 12 weeks 190. The label ties risk directly to exposure: both the risk of developing tardive dyskinesia and the likelihood it becomes irreversible increase with duration of treatment and total cumulative dose 190. Here the evidence is the observed dose-and-duration relationship itself, which drives both the warning and the explicit 12-week limit.
What the pattern tells regulatory teams
Across these products, a few generalizations hold. First, FDA does not require a single evidentiary standard: a mandated post-market RCT (tofacitinib), a meta-analysis (rosiglitazone), pooled trial data (antidepressants), pivotal-trial toxicity rates (CAR-T), spontaneous reports (montelukast), animal carcinogenicity (semaglutide), and observed dose-response relationships (metoclopramide) have all sufficed. Second, the warning is frequently applied at the class level once a mechanism-linked signal is established (NSAIDs, opioids, fluoroquinolones, GLP-1 agonists). Third, the labeling itself is the primary record of the triggering evidence, and it will state candidly when human relevance or causality remains unresolved (rosiglitazone's "inconclusive" ischemia data; semaglutide's undetermined human relevance). For a submission or lifecycle team, the practical lesson is that any of these evidence types, individually, can be enough to move a risk into the box, and that the strength of the underlying data does not have to be definitive for FDA to act.