What CDC and FDA datasets actually say about health claims

What to take away

• A number is only usable once you can name its dataset, years, geography, and population in one sentence.
• Counts track population size; rates track risk. A rising count is not a rising rate, and an age-adjusted rate is a third thing again.
• A VAERS report is an unverified filing, not a confirmed injury, and the raw count cannot be divided by doses to produce a risk.
• An FDA label bounds the approved use, dose, and population; a claim outside those bounds needs its own evidence.
• The FTC requires advertisers to hold competent and reliable scientific evidence before a health claim runs, not after.

What these datasets can and cannot tell you

Federal health datasets are surveillance instruments. They count what someone reported, coded, or measured, in a defined place and window. That makes them strong on patterns across large groups and silent on one person's case.

CDC WONDER returns counts and rates from national collections, including mortality and natality files. You set the years, the geography, the cause-of-death codes, and the demographic breakdown. Change any one and the number changes.

CDC FastStats is the short version: national totals and rates, from asthma to vaccination. Useful for orientation, thin on the subgroup detail a viral claim usually needs.

On the drug side, FDA drug approval information covers what a product is approved to treat, at what dose, in which patients. The label is a legal and clinical document, not a marketing summary.

None of these can tell you why one person got sick, whether one product caused one outcome, or what happens next. Those are causal and individual questions. Surveillance data answer descriptive ones.

The test: can you state the dataset, the years, the geography, and the population in one sentence? If not, you have a screenshot, not a finding. The science health news verification guide applies that same discipline to any study or dataset story.

Reading WONDER and FastStats without misreading rates

A rate is a fraction. The numerator is events; the denominator is the population at risk during the period. Most viral misreadings treat a numerator as the whole story.

Take the common one. A post says deaths from a condition rose 40 percent in a state, citing WONDER. Possibly true. But if the state's population grew and aged over the same decade, the crude rate may have risen far less, and the age-adjusted rate may be flat.

That is the misread to keep in your pocket: a 40 percent rise in a count, presented as a 40 percent rise in risk. Counts track population size. Rates track risk. Age-adjusted rates let you compare groups with different age structures.

A checklist for any WONDER or FastStats figure:

• Which dataset and which years are in the query?
• Is the number a count, a crude rate, or an age-adjusted rate?
• What is the denominator, and does it match the population described?
• Are the geography and the time window stated, or assumed?
• Is the comparison group named, or is the number standing alone?
• Do provisional and final data differ for this period?
• Does the source page carry a caveat about small counts or suppression?

Small counts deserve their own warning. When a county records a handful of deaths, year-to-year swings look dramatic and mean little. Agencies suppress or flag those cells for exactly this reason.

Provisional data also shift. Death certificates get amended, cause codes get revised, and a figure pulled in March may not match the final file. A claim built on a provisional number should say so.

FastStats gives you the national anchor. If a post claims a national trend, check FastStats first, then go to WONDER for the subgroup. When the two disagree, the subgroup definition is usually why. For where surveillance sits relative to trials and case reports, see health research evidence types.

VAERS reports: what a raw count means

VAERS is the Vaccine Adverse Event Reporting System, co-managed by CDC and FDA. Anyone can file a report. That openness is the point: it is an early warning system, not a verdict.

A report means someone believed an event followed a vaccination and chose to file. It does not mean the vaccine caused the event. Reports arrive from clinicians, patients, manufacturers, and lawyers, with varying documentation.

The most common viral misuse is arithmetic: take the raw report count, divide by doses administered, announce a risk. That calculation is invalid. Reports are unverified, duplicated, and biased toward events people already suspect.

Underreporting is real and uneven. A mild event may never be filed. A serious event near a vaccination date is more likely to be. The raw count captures neither pattern.

CDC and FDA clinicians review reports for patterns that warrant further study. A signal can prompt investigation. It cannot, by itself, establish that a product caused an injury. The same logic covers the FDA's adverse event reporting for drugs and devices: the report is a hypothesis generator.

When a post says a system logged a certain number of reports, the follow-up questions are: how many were serious, how many were verified, what did the review conclude, and what do controlled studies show?

What an FDA label claim covers

An FDA-approved label is the reference document for what a drug may claim. It states the approved indication, the population studied, the dose, the route, and the known risks.

Read it by separating four things: what the drug is approved to treat, what the trials measured, what the warnings say, and what remains unknown. A viral post often collapses all four into one sentence.

Suppose a post says a drug cuts heart attack risk by a third. The label might show that result in a specific group, over a specific follow-up, against a named comparator, on a composite endpoint. The headline drops every qualifier.

Composite endpoints are a frequent trap. If the endpoint combines death, hospitalization, and a lab change, a reduction may be driven by the softest component. The label's trial section usually breaks this out.

Off-label use is legal and sometimes standard care, but a claim about an off-label use is not backed by the label. It needs its own evidence, and the FTC standard still applies to advertising.

Devices run through a separate pathway. FDA device information covers cleared and approved devices, their indications, and the safety communications where recalls and warnings live.

One habit: open the label, search the indication and the trial results, copy the exact population. If the viral claim's population is wider, the claim is wider than the evidence.

Reports versus causation

Causation requires more than sequence. Two things happening in order, or at once, is a correlation until a study design rules out the alternatives.

Spontaneous reports cannot rule out those alternatives. They lack a control group, a defined denominator, and consistent verification. That is not a flaw in the reporters; it is the design.

Establishing causation typically needs controlled comparisons: randomized trials where feasible, or observational studies that adjust for confounders and replicate. Even then, conclusions are probabilistic.

So "reports prove harm" is a category error. Reports can justify studying a question. They cannot answer it.

The reverse fails too. "No reports prove safety" is not an argument. Absence of reports may reflect a small population, a short window, or a system nobody uses.

When you audit a story built on reports, check whether the reporter contacted the agency, read the label, and looked for controlled studies. The audit clinical study story method covers that sequence.

The FTC standard behind an advertising claim

The FTC governs advertising for health products, including supplements, foods, devices, and services. Its health products guidance sets the evidence bar advertisers must clear before a claim runs.

For most health claims the standard is competent and reliable scientific evidence: well-controlled human studies, by qualified researchers, whose results support the specific claim.

The FTC weighs the whole body of evidence, not one study. It looks at design, population, endpoint, and whether independent replication exists. A single small trial rarely carries a broad claim.

Two patterns draw enforcement. A claim that a product treats or prevents a disease without adequate evidence. A testimonial or endorsement presented as typical when it is not.

Qualifiers matter. Adding "may" or "supports" does not rescue a claim the evidence cannot back. The FTC evaluates the net impression a reasonable consumer takes away.

This is where advertising and label claims diverge. A label claim is bounded by an approval. An advertising claim is bounded by the evidence the advertiser holds. For a writer, the guidance doubles as a checklist: what evidence exists, who ran it, and does it match the claim's scope?

Repeatable ways raw numbers go wrong

Most bad health numbers fail in a few familiar ways. Spotting the pattern is faster than re-deriving the statistics.

Base rate neglect is the most common. A large relative increase on a tiny base is a small absolute change. Leading with the percentage and burying the base misleads without stating anything false.

Denominator drift is next. A rate per 100,000 people and a rate per 100,000 tests are different quantities, and headlines swap them.

Confusing correlation with cause shows up in coverage of observational studies. So does treating a preprint as peer-reviewed, or a press release as a study.

Geographic aggregation hides local variation. A national average can hide counties moving in opposite directions. This matters for local coverage in five states:

• California
• Texas
• Florida
• Georgia
• New York

Public records work has its own version. Local governments publish budgets, inspection results, and health permits in inconsistent formats, and a county figure is not a state figure.

Then there is the retraction problem. When a paper is withdrawn, coverage built on it often stays online unchanged. A retracted health study correction explains how to handle that cleanup.

When to ask a researcher first

Ask before publishing whenever the claim turns on a rate, a subgroup, or a cause. Those are the three places a careful reader and a careless one diverge most.

Ask when the number comes from a query you cannot reproduce. If you cannot rebuild the WONDER query or find the FastStats table, you can only repeat the figure, not verify it.

Ask when the claim involves a rare event, a small county, or a short window. These are the cells agencies suppress or flag, and the ones screenshots omit.

Ask when the claim is causal and the source is a report count. A researcher who works with the system can say what the reports do and do not support in minutes.

Ask when a product claim lacks a cited study. The FTC standard requires evidence in hand, so its absence is itself a finding worth reporting.

A short email works: state the number, the dataset, the years, and the claim you plan to make. Ask what the number cannot support. Most researchers answer that readily. For the whole sequence, dataset to publication, the science health verification checklist runs the steps in order.

Common questions

What is CDC WONDER used for?

It queries national public health data collections. These include mortality and natality files. It returns counts and rates by year, place, cause, and demographic group. You set the parameters, so the same topic can return different numbers.

Does a VAERS report mean a vaccine caused the injury?

No. VAERS accepts unverified reports from anyone. A report signals that someone believed an event followed vaccination and can prompt review, but it does not establish cause.

Can I divide VAERS reports by doses given to get a risk?

No. Reports are unverified, duplicated, and subject to reporting bias, so they are not a numerator for risk. Controlled studies are the right source for that figure.

What does an FDA label actually cover?

It covers the approved indication, the studied population, the dose, and the known risks. A claim outside those bounds is not supported by the label and needs separate evidence.