FDA Animal & Veterinary Adverse Events: The Federal Record of Veterinary Drug Safety

A dog seizes hours after chewing a flea-and-tick preventive; a cat goes into kidney failure after a long-acting injection; a horse colics after a routine dewormer; a herd of cattle fails to respond to an antibiotic that should have worked. When a veterinarian, a manufacturer, or an animal's owner writes down what happened and sends it to the FDA, the report joins the federal record of veterinary drug safety—roughly 25,000 adverse-event reports, one row per report, the passive-surveillance stream the Center for Veterinary Medicine mines for the safety signals an animal drug's trials never saw.

This article covers what the veterinary adverse-events dataset is and how the FDA's Center for Veterinary Medicine frames it; the way it parallels—and differs from—the human FAERS system; the passive-surveillance, spontaneous-reporting model that produces the data, and who is required to report versus who reports voluntarily; the anatomy of a report, from the animal and the product to the reactions and the outcome; the species-and-reaction taxonomy that structures any analysis, including the VeDDRA coding dictionary and the special category of lack-of-effectiveness reports; how the accumulating signal drives label changes, warnings, and the occasional withdrawal; how the table joins to the broader animal-drug record and to its human FAERS counterpart; a Python workflow that pulls reports from the openFDA animal-and-veterinary API and ranks products by report volume; and the caveats—underreporting, reporting bias, and the cardinal rule that a single report is not proof of causation—that every analyst must internalize before drawing conclusions.

What the dataset is

The veterinary adverse-events dataset is the FDA's record of the side effects, product failures, and unexpected outcomes reported when animals are treated with FDA-regulated veterinary drugs. It spans the full range of treated animals—companion animals such as dogs, cats, and horses, and food-producing animals such as cattle, swine, and poultry—and the full range of regulated products, from flea-and-tick preventives and dewormers to antibiotics, vaccines, anesthetics, and chronic-disease medications. Each report is a structured account of a single adverse experience: which animal, which product, what reaction, and what outcome. The accumulating collection of these reports is what the FDA calls its adverse drug experience (ADE) record, historically distributed as the Center for Veterinary Medicine's “FOI Drug Experience Reports” and now surfaced through the openFDA animal-and-veterinary endpoint.

In our database this record is stored as the table fda_animal_vet, with the grain of one row per adverse-event report: a single product implicated in many reports contributes many rows, and a single report describing one animal with several reactions is still one row with a nested list of reactions. Roughly 25,000 reports are held. The columns capture the four pillars of every report—the report's identity and metadata, the animal, the product or products, and the reactions and outcome:

unique_aer_id_number     -- report identifier (one per adverse-event report)
original_receive_date    -- date the FDA first received the report
type_of_information      -- safety/adverse reaction vs lack of expected effect
primary_reporter         -- veterinarian, owner/caregiver, or other
serious_ae               -- whether the event was classified as serious
-- the animal that was treated:
animal.species           -- Dog, Cat, Horse, Cattle, Swine, ...
animal.gender            -- male, female, neutered, spayed, unknown
animal.age               -- reported age and unit
animal.weight            -- reported weight and unit
-- the product(s) administered:
drug.brand_name          -- marketed product name
drug.active_ingredients  -- active substance(s)
drug.dosage_form         -- chewable, topical, injectable, oral, ...
drug.route               -- route of administration
-- the reported outcome:
reaction.veddra_term_name -- clinical sign coded to the VeDDRA dictionary
outcome.medical_status   -- recovered, ongoing, euthanized, died, unknown

The unique_aer_id_number is the load-bearing column—the persistent identifier for a single adverse-event report, the key by which a report is deduplicated, updated, and referenced. The animal.species field anchors nearly every analysis, because veterinary safety is fundamentally species-specific: a dose safe in a dog can be fatal in a cat, ivermectin tolerated by most dogs is dangerous in certain herding breeds, and a drug cleared for cattle may never have been studied in horses. The drug.brand_name and drug.active_ingredients fields identify the implicated product, and a single report can name more than one drug, which complicates the attribution of a reaction. The reaction.veddra_term_name field carries the clinical sign—vomiting, lethargy, seizure, pruritus, death—coded to the standardized Veterinary Dictionary for Drug Regulatory Activities (VeDDRA), and the type_of_information field draws the critical line between a true adverse reaction and a lack of expected effectiveness: the report that the dewormer did not deworm, the antibiotic that did not clear the infection, the preventive that failed to prevent. Together these columns turn a row from a bare complaint into a diagnostic of which products, in which species, produce which problems.

The Center for Veterinary Medicine and the regulatory frame

Veterinary drugs in the United States are regulated by the FDA's Center for Veterinary Medicine (CVM), the veterinary counterpart to the agency's human drug center. The authority traces to the Federal Food, Drug, and Cosmetic Act, which makes an animal drug a “new animal drug” that may not be marketed until the FDA has approved it as safe and effective for its intended use. A pharmaceutical company seeking to sell an animal drug submits a New Animal Drug Application (NADA)—or, for a generic copy of an already-approved product, an Abbreviated New Animal Drug Application (ANADA)— and CVM reviews the supporting data before clearing the product for market. Approval is specific to a species, an indication, a dose, and a route; the same molecule approved for one species and one use is unapproved for any other.

Pre-market trials, however, are limited—conducted in a controlled population of a few hundred or a few thousand healthy animals over a bounded period. They cannot detect rare reactions, reactions confined to a particular breed or age, interactions with other drugs or with concurrent disease, or problems that emerge only after years of use in millions of real-world animals of every conformation and health status. The veterinary adverse-event system is the program's answer to that gap: a post-market pharmacovigilance apparatus that watches the drug in the field after approval, collects the reactions the trials were too small or too short to see, and feeds the signal back to CVM's post-marketing surveillance staff. The same authority that lets CVM approve a drug lets it require label changes, mandate stronger warnings, demand additional studies, or—in the gravest cases— withdraw the approval when the post-market signal shows the drug to be unsafe as used. The adverse-event record is the evidentiary substrate for all of those actions.

The FAERS parallel and how it differs

The veterinary adverse-events dataset is the direct counterpart to the FDA Adverse Event Reporting System (FAERS), the human-drug safety database. The architecture is the same: a spontaneous, largely voluntary stream of reports describing a patient, a drug, a reaction, and an outcome, accumulated centrally and mined for signals that the controlled pre-market trials could not produce. Anyone analyzing veterinary safety data who already knows FAERS will recognize the shape immediately—the passive-surveillance model, the denominator-free counts, the duplicate-report problem, and the iron rule that the data generates hypotheses rather than proving causation all carry over wholesale.

But the differences are instructive and they run deep. The most obvious is the patient: where FAERS has a single species, the veterinary record spans dozens, and species is not a covariate but the dominant axis of the entire dataset, because toxicity, metabolism, and safe dosing differ radically between a cat and a cow. The second is the reporter: FAERS is dominated by patients, physicians, and pharmacists, while veterinary reports come from veterinarians, manufacturers, and—a category with no clean human analogue— animal owners describing a pet at home, who may or may not have a clinical diagnosis. The third is the existence of food-producing animals, which introduces concerns FAERS never confronts: not just the welfare of the treated animal but drug residues in meat, milk, and eggs entering the human food supply, a human-health dimension layered on top of the animal-health one. The fourth is the formal inclusion of lack of effectiveness as a reportable adverse experience—a product that simply does not work is treated as a safety-relevant event in the veterinary system in a way that is less central on the human side. These differences mean the veterinary record is not merely “FAERS for animals”; it is a structurally richer surveillance problem.

The passive-surveillance model and who reports

The defining feature of the dataset—and the source of both its value and its limitations—is that it is a passive-surveillance, spontaneous-reporting system. Nobody systematically samples the population of treated animals to estimate how often a reaction occurs. Instead, reports arrive spontaneously when someone notices a problem and chooses, or is required, to report it. The data is therefore a stream of signals—invaluable for spotting patterns and emerging problems—but it is not a census, and it carries no denominator: the count of reports for a product reflects not only how often the product causes harm but how widely it is used, how recently it launched, how much publicity its risks have received, and how diligently its reactions are reported.

Who reports, and under what obligation, shapes the data profoundly. Reporting is mandatory for manufacturers: a sponsor of an approved animal drug is legally required to report adverse experiences it learns about to CVM, on defined timelines, with serious and unexpected events expedited. This means the large majority of reports in the system are routed through manufacturers—an owner or veterinarian frequently contacts the company first (often via a number on the label), and the company then files the report with the FDA. Reporting is voluntary for veterinarians and for animal owners, who can report directly to the FDA but are not compelled to. The consequence is a reporting funnel with several leaks: a reaction is noticed only if it is recognized as drug-related, reported only if the observer bothers, and routed accurately only if the manufacturer's pharmacovigilance pipeline captures it faithfully. Because manufacturers are the mandatory channel, the products of attentive, well-resourced sponsors with prominent customer-service lines may be over-represented relative to products from smaller sponsors—a reporting artifact, not a safety difference.

The species-and-reaction taxonomy

Any serious use of the dataset begins with two cross-cutting taxonomies: the species of the treated animal and the coded reactions it experienced. The species dimension is first among equals. Companion animals dominate the report counts—dogs and cats above all, with horses a distant third—both because they are treated with an enormous variety of products and because their owners form deep attachments and notice and report problems. Food-producing species—cattle, swine, poultry—appear in smaller numbers and skew toward different products (antibiotics, antiparasitics, production drugs) and different concerns (lack of effectiveness, residue questions). Because safe dosing is species-dependent, stratifying every metric by species is not optional refinement but a prerequisite for the analysis to mean anything: a reaction rate pooled across dogs and cats can hide the fact that a drug is well tolerated in one and dangerous in the other.

The reaction dimension is standardized through the Veterinary Dictionary for Drug Regulatory Activities (VeDDRA), the veterinary analogue of the MedDRA terminology used for human drugs. VeDDRA gives every reported clinical sign a controlled term—Vomiting, Diarrhoea, Lethargy, Pruritus, Seizure, Ataxia, Anorexia, Death—arranged in a hierarchy from specific signs up to broader system-organ groupings. The coding matters because free-text descriptions from owners (“my dog was shaking”) cannot be aggregated, whereas VeDDRA terms can be counted, ranked, and compared across products and species. The most frequent terms across the dataset tend to be the common, often mild reactions— vomiting, lethargy, lack of appetite, injection-site or application-site signs—but the analytically important ones are the serious, specific signs (seizures, neurological signs, hepatic or renal injury, death) whose disproportionate appearance against a particular product is the kind of signal that prompts regulatory attention.

A category that deserves separate treatment is the lack-of-effectiveness report, flagged by the type_of_information field. In the veterinary system, “the product did not work” is a reportable adverse experience, not merely a customer complaint. A heartworm preventive given correctly that is followed by heartworm infection, an antiparasitic that fails to clear the parasites, an antibiotic that does not resolve the disease—these matter because a failure of efficacy in a food animal can mean continued disease and economic loss, a failure in a companion animal can mean a preventable illness, and a cluster of effectiveness failures against a single product can signal emerging drug resistance, a manufacturing defect, or a labeling problem. Effectiveness reports must be analyzed separately from safety reactions, because conflating “the drug hurt my animal” with “the drug did nothing” produces meaningless aggregates.

From accumulating signal to label change

The purpose of collecting these reports is not the archive but the action they enable. CVM performs ongoing signal detection on the accumulating data: watching for a particular product to draw a disproportionate share of reports of a particular serious reaction in a particular species, for a sudden rise in reports against a previously quiet product, or for a pattern—a breed, an age, a co-administered drug—that points to a previously unrecognized risk. A signal is not a conclusion; it is a flag that triggers closer review, in which CVM weighs the biological plausibility, the consistency of the reports, the temporal relationship, and any dose-response pattern.

When the review confirms a real risk, the agency has a graduated set of responses, and the adverse-event record is what justifies each. The most common outcome is a label change: adding a newly recognized adverse reaction to the product's labeling, strengthening a precaution, adding a contraindication for a particular breed or condition, or revising the dosing guidance. A more emphatic step is a safety warning or communication to veterinarians and the public—the kind of advisory the FDA has issued about, for example, the potential for neurologic adverse events in certain classes of flea-and-tick products, which prompted label updates across an entire drug class. CVM may also require the sponsor to conduct additional studies. The rarest and gravest outcome is withdrawal, when the post-market signal shows the risk to be unacceptable as the product is used. In every case the through-line is the same: the spontaneous reports, individually inconclusive, accumulate into the evidence base that moves a product from approved-as-is to relabeled, restricted, or removed.

Joining to the broader animal-drug record

The adverse-events table is most powerful as one facet of a connected animal-drug record rather than in isolation, and several joins extend it. The most natural is to the approved-product record—the Animal Drugs @ FDA (Green Book) listing of approved NADAs and ANADAs. Joining reports to the approved-product record by product name and active ingredient supplies the regulatory context a bare report lacks: which species and indications the drug is actually approved for, who the sponsor is, and when it was approved. That context is what lets an analyst distinguish an on-label reaction (the drug used as approved) from a reaction following extra-label use (the drug used in a species, dose, or indication outside its approval), a distinction central to interpreting veterinary reports because off-label use is both legal and common in veterinary practice.

A second, conceptual join is to the human FAERS record, this dataset's direct counterpart. Many active ingredients straddle the two worlds—the same molecule may be marketed for both people and animals, and accidental human exposure to veterinary products (a child or a pet owner ingesting a flea preventive, for example) appears in one system while the animal reaction appears in the other. Comparing the reaction profile of a shared active ingredient across the human and veterinary records can corroborate a class effect or reveal a species-specific vulnerability. A third extension reaches toward the food-safety apparatus for food-producing animals, where drug-residue concerns connect the veterinary record to the broader regime governing contaminants in the human food supply. Through these joins the adverse-event report stops being an anonymous complaint and becomes a node in the full life cycle of a regulated veterinary product, from approval through field use to safety action.

Analytical uses

A national, product-resolved, species-stratified record of veterinary adverse events supports a distinctive set of analyses, provided each respects the passive-surveillance caveats developed below.

Products by report volume, stratified by species is the most immediate use: ranking which products and active ingredients draw the most reports in dogs, in cats, in horses, in cattle. The ranking is a starting point, never a conclusion, because report volume tracks usage as much as risk—a widely sold preventive will top any raw count—but it identifies where the reported burden concentrates and where to look harder. Reaction profiles by product exploit the VeDDRA coding to characterize what kind of problem a product produces—whether the reports against it skew toward mild gastrointestinal signs, toward serious neurological events, or toward injection-site reactions—which is far more informative than a count alone.

Disproportionality signal detection is the rigorous extension: computing, for each product-reaction pair, whether a reaction is reported more often for that product than would be expected given how often it is reported across all products—the family of measures (reporting odds ratios, proportional reporting ratios) that pharmacovigilance uses to surface signals from denominator-free spontaneous data. Effectiveness-failure analysis isolates the lack-of-effectiveness reports to watch for clusters that might indicate emerging resistance or a quality problem. And temporal and seriousness trendsuse the receive date and the serious-event flag to track whether reports against a product are rising, whether a label change or warning was followed by a change in reporting, and how the share of serious outcomes—including death and euthanasia—moves over time.

Python workflow: reports from the openFDA animal-and-veterinary API

The script below pulls veterinary adverse-event data from the openFDA animal-and-veterinary endpoint and computes three of the core metrics: the breakdown of reports by animal species, the most common reported reactions coded to VeDDRA terms, and a ranking of the products with the most reports for a chosen species. It uses the openFDA count parameter for the fast server-side tallies and pages full reports only where it needs the per-product detail. No API key is required for light use; an openFDA key raises the rate limits for larger pulls. Because a single report can name multiple products and multiple reactions, the script deduplicates report identifiers when counting so that a multi-product report is not triple-counted against each product it names.

import requests, pandas as pd
from collections import Counter

# openFDA Animal & Veterinary Adverse Event API
# Endpoint: https://api.fda.gov/animalandveterinary/event.json
# Source: FDA Center for Veterinary Medicine (CVM) adverse-event files.
# No API key required for <= 240 requests/min and 1,000/day; register at
# https://open.fda.gov/apis/authentication/ for a higher ceiling.
#
# This script:
#   1. Tallies reports by animal species (server-side count)
#   2. Tallies the most common reported reactions (VeDDRA terms)
#   3. Pages a slice of full reports and ranks products by report volume
BASE = "https://api.fda.gov/animalandveterinary/event.json"


def count_field(field, search=None):
    # openFDA count parameter returns a fast server-side tally.
    params = {"count": field, "limit": 1000}
    if search:
        params["search"] = search
    r = requests.get(BASE, params=params, timeout=60)
    r.raise_for_status()
    return {row["term"]: row["count"] for row in r.json().get("results", [])}


def species_breakdown():
    # animal.species holds the treated species: Dog, Cat, Horse, Cattle, ...
    return count_field("animal.species")


def top_reactions():
    # reaction.veddra_term_name: clinical signs coded to the VeDDRA dictionary
    # (e.g. Vomiting, Lethargy, Seizure, Death, Lack of efficacy).
    return count_field("reaction.veddra_term_name.exact")


def fetch_reports(species, limit=1000):
    # Pull full reports for one species and flatten product brand names.
    out, skip = [], 0
    while skip < limit:
        params = {"search": f"animal.species:{species}",
                  "limit": min(100, limit - skip), "skip": skip}
        r = requests.get(BASE, params=params, timeout=60)
        if r.status_code == 404:
            break
        r.raise_for_status()
        results = r.json().get("results", [])
        if not results:
            break
        for rec in results:
            for prod in rec.get("drug", []):
                name = prod.get("brand_name") or "(unspecified)"
                out.append({"report_id": rec.get("unique_aer_id_number"),
                            "product": name})
        skip += len(results)
    return pd.DataFrame(out)


def rank_products(species, top=15):
    df = fetch_reports(species)
    if df.empty:
        print(f"No reports returned for {species}.")
        return
    by_product = df.groupby("product")["report_id"].nunique()
    by_product = by_product.sort_values(ascending=False)
    total = df["report_id"].nunique()
    print(f"{species}: {total:,} reports across {df['product'].nunique():,} products")
    for name, n in by_product.head(top).items():
        print(f"  {n:6,}  {name}")


if __name__ == "__main__":
    print("Reports by species:")
    for sp, n in sorted(species_breakdown().items(), key=lambda kv: -kv[1])[:10]:
        print(f"  {n:7,}  {sp}")
    print("\nMost common reported reactions:")
    for term, n in sorted(top_reactions().items(), key=lambda kv: -kv[1])[:10]:
        print(f"  {n:7,}  {term}")
    print()
    rank_products("Dog")

Two practical notes apply. First, the product ranking is a count of reports, not a measure of risk, and it must be read that way: the products at the top of the list are typically the most heavily used products in their category, and a fair comparison would normalize report counts by an estimate of usage or doses sold—which the dataset does not contain. Disproportionality measures, which compare a product's reaction mix against the background mix rather than against raw usage, are the better tool when the question is whether a product is unusually risky rather than merely popular. Second, for national-scale or longitudinal work—computing reporting odds ratios across the whole corpus, or tracking a product before and after a label change—the FDA's downloadable CVM adverse-event data files are more efficient than paginating the API thousands of times and carry the authoritative, version-stamped field definitions for the release.

Limitations and analytical caveats

The veterinary adverse-events dataset is the most comprehensive public record of animal-drug safety signals in the United States, but it is a passive-surveillance system and carries the structural limitations of one—limitations an analyst must internalize before drawing any conclusion.

A single report is not proof of causation. This is the cardinal rule, identical to FAERS. A report records a temporal association—a reaction observed after a product was given—not an established causal link. The animal may have had an underlying disease, been given several drugs at once, reacted to something unrelated, or experienced a coincidence. The reporter's belief that the product caused the harm is not a clinical determination. The data is for generating hypotheses and detecting patterns across many reports; no individual report should be read as a verdict that a product caused an outcome, and presenting it that way badly misrepresents what the dataset can support.

Underreporting is pervasive and uneven. Only a fraction of adverse experiences are ever reported. A reaction must be noticed, recognized as possibly drug-related, and then actually reported, and that funnel leaks at every stage— and it leaks unequally. Mild and self-limiting reactions are under-reported relative to dramatic ones; reactions in food animals on a working farm are reported far less than reactions in a beloved pet; problems with familiar old products draw fewer reports than the same problems with a heavily publicized new one. The counts in the dataset therefore measure the combined effect of true incidence, usage, salience, and reporting behavior, and cannot be read as incidence on their own.

There is no denominator. The dataset records the numerator—reported events—but not the denominator: how many animals received the product. Without knowing how many doses of a product were sold, a count of reports cannot be converted into a rate, and a product with more reports may simply be a product with more users. Every comparison between products is contaminated by differences in usage unless it is explicitly normalized, which is precisely why disproportionality methods—comparing a product's internal reaction mix rather than its raw volume—exist.

Reporting bias and stimulated reporting distort trends.Publicity changes reporting. An FDA safety communication, a media story, or a class-action advertisement about a product can trigger a wave of reports—many of them retrospective, many of them about events that occurred long before—producing a spike in the data that reflects awareness rather than a real change in the product's safety. The manufacturer-mandatory channel adds its own bias, since the products of attentive, well-staffed sponsors with prominent reporting lines may be captured more completely than those of smaller sponsors. Treating a rise in reports as evidence that a product became more dangerous, without accounting for what was happening in the world around it, is a classic misreading.

Held with these caveats in mind, the fda_animal_vet table is a uniquely valuable resource: a product-resolved, species-stratified, date-stamped record of the safety and effectiveness problems that surface only after a veterinary drug leaves the controlled world of its trials and enters the messy reality of millions of treated animals—the post-market half of the bargain that lets a regulator approve a drug on limited evidence and then watch, report by report, for the harms the evidence could not foresee.