Open health data sources every analyst should know

Start with the question, not the dataset

Most analysis that goes wrong with public health data goes wrong in the first hour, when the analyst picks a dataset because it is familiar rather than because it answers the question. Every federal health source was built for a specific purpose. A household survey measures what people report about themselves. A hospital database counts stays, not people. A death file records what a certifier wrote. A surveillance summary counts cases that were tested, reported and confirmed. Ask any of them a question outside its design and it will still return a number.

So before opening anything, write down what you are counting (people, events or dollars), for what population, at what geography, and for what period. That sentence will eliminate most sources immediately and tell you which pitfalls to watch in the ones that remain.

What follows is a working tour of the sources we reach for most often, grouped by what they are made of, with the pitfalls that matter in practice.

The sources at a glance

The first table covers sources built from surveys, where the unit is a sampled person or household and the estimates depend on weights. The second covers sources built from records, where the unit is an event or a claim.

Survey source	What it measures	Geography	Access	Main pitfall
NHIS	Self-reported health status, conditions, access and use of care	National; detailed geography restricted	Free public-use files	Questionnaire redesign in 2019 breaks many trends
NHANES	Interviews plus physical exams and lab tests	National only	Free public-use files; some restricted	Several weight types; using the wrong one biases results
BRFSS	Health behaviors, chronic conditions, preventive care	State; some metro areas; model-based small areas via PLACES	Free public-use files	2011 method change limits comparison with earlier years
MEPS	Health spending, use of services, insurance, by person and event	National and Census region	Free public-use files; restricted geography via data center	Modest sample; rare conditions need pooled years
ACS	Population, insurance coverage, disability, income	Down to block group with 5-year estimates	Free	Margins of error are large for small areas and often ignored

Records source	What it measures	Geography	Access	Main pitfall
CDC WONDER	Deaths and births from death and birth certificates, plus other CDC datasets	National, state, county	Free online query under data use terms	Suppressed small counts; underlying versus multiple cause
HCUP	Hospital stays, emergency visits and ambulatory surgery, with charges	National samples; state databases by participating state	Purchased under a data use agreement with required training	Counts discharges, not patients; charges are not costs
CMS public use files	Medicare services, payments and prescribers; geographic variation	Provider, county, state	Free	Mostly fee-for-service only; small cells suppressed
NNDSS and ArboNET summaries	Reported cases of notifiable and mosquito-borne diseases	State and county	Free published tables	Counts reported cases; mild cases and testing gaps undercount
BLS	Employment, wages, compensation, prices, time use	National, state, metro, county by program	Free	Different programs define employment differently

CDC WONDER: deaths and the questions around them

CDC WONDER is a free query interface to death and birth certificate data and several other CDC datasets. For mortality, you can tabulate deaths and rates by cause, age, sex, race and ethnicity, place of residence and year, down to the county, with age-adjusted rates computed for you.

Two decisions shape every mortality query. First, underlying cause versus multiple cause: underlying cause is the condition that started the chain toward death, and multiple cause counts every condition on the certificate. Chronic conditions that contribute to death without starting the chain are badly undercounted by underlying cause alone. Second, suppression: sub-national counts below a small threshold are suppressed, and rates built on fewer than 20 deaths are flagged as unreliable. County queries for uncommon causes return mostly blanks, which is a signal to aggregate geography or pool years, not to estimate the blanks.

Provisional data arrive faster than final data and can change. For anything that will be cited, note which you used.

NHIS and NHANES: what people report and what an exam finds

The National Health Interview Survey is the federal government's main household survey of health. It covers conditions, functioning, health behaviors, insurance, access to care and use of services for the civilian population living in households. Its strength is breadth and continuity. Its limits are that it relies on self-report, excludes people in institutions, and in public files withholds geography below broad regions. The 2019 questionnaire redesign changed many questions, so trends that cross that year need care.

NHANES adds what self-report cannot: physical examinations and laboratory tests on a national sample. It is the source for measured blood pressure, blood glucose, body measurements and biomarkers, which means it can estimate undiagnosed conditions that NHIS cannot see. It produces national estimates only, and its sample is small enough that data are usually pooled across cycles. Its field operations were interrupted in 2020, which is why some data are released as a combined prepandemic file.

NHANES has interview weights, exam weights and subsample weights for particular lab tests. The correct weight is the one for the smallest subsample whose variables you use. Using the interview weight on lab variables is a common, silent error.

BRFSS and PLACES: the state view

The Behavioral Risk Factor Surveillance System is a telephone survey run by state health departments with CDC. It is the largest continuous health survey in the country and the standard source for state-level estimates of chronic conditions, risk behaviors and preventive care. Some metropolitan areas have their own estimates. States can add their own question modules, so not every question is asked everywhere every year.

BRFSS added cell phone respondents and changed its weighting method in 2011. Estimates from before and after are not directly comparable, and CDC advises against treating the change as a trend.

For counties and smaller areas, CDC's PLACES project publishes model-based estimates derived from BRFSS and Census data. These are useful for comparing places and targeting resources. They are modeled, not measured, so they will smooth over real local clusters and should never be presented as a local survey result.

MEPS and HCUP: utilization and spending

The Medical Expenditure Panel Survey, run by AHRQ, follows households for two years and records every medical event, what it cost and who paid, with household reports supplemented by data collected from their providers. It is the core public source for health spending by person, condition, service type and payer, and the basis for most incremental cost estimates. Public files carry only broad geography, and the sample is modest, so rare conditions require several pooled years.

The Healthcare Cost and Utilization Project, also AHRQ's, assembles hospital discharge records from participating states. Its national products sample inpatient stays and emergency department visits nationally; its state databases cover all discharges in each participating state. Access is by purchase under a data use agreement that requires completing a short training course and restricts how small cells may be reported, and HCUPnet offers free summary queries. HCUP records charges, which are list prices, so cost analysis needs the cost-to-charge ratio files HCUP provides. The national inpatient sample counts discharges, not people, and cannot follow a patient from one stay to the next; readmission work uses the separate readmissions database.

CMS public use files

CMS publishes free files on Medicare services and payments: utilization and payment by provider and service, prescribing by provider, and geographic variation in spending and use by state and county. They are excellent for questions about Medicare supply, practice patterns and regional variation.

Most of these files describe traditional fee-for-service Medicare. Beneficiaries enrolled in Medicare Advantage are largely absent, and that share varies by region, so a county comparison can reflect enrollment mix as much as care. Cells with very small counts are suppressed under CMS's cell size policy. Person-level Medicare and Medicaid data exist, but they are research identifiable files requested through CMS's research data process, with a data use agreement, a fee and months of lead time.

Notifiable disease and arboviral surveillance

Reporting of notifiable diseases is required by state law and shared voluntarily with CDC through the National Notifiable Diseases Surveillance System, which publishes provisional weekly tables and finalized annual tables. ArboNET, the national arboviral surveillance system, collects reports of human cases of mosquito-borne and tick-borne viral diseases such as West Nile virus, along with data on infected mosquitoes and animals, and CDC publishes summaries by state and county.

These are counts of cases that someone tested for, reported and classified under a case definition. They are not incidence. Mild illness that never reaches a clinician is missed, testing practice varies by place and year, and case definitions change. For arboviral disease, severe neuroinvasive cases are reported far more completely than milder ones, which is why analysts often treat neuroinvasive counts as the more stable series. Reporting lists also differ by state, so a condition absent from one state's numbers may simply not be reportable there.

ACS and BLS: denominators and the economy

The American Community Survey supplies the denominators most health analysis needs: population by age, sex, race, income, insurance coverage, disability status and much more. One-year estimates are published for areas of 65,000 people or more; five-year estimates cover every area down to the block group. Every ACS estimate comes with a margin of error, and for small areas that margin can be larger than the difference you are trying to describe. Census advises against comparing overlapping five-year periods, because they share most of their data.

The Bureau of Labor Statistics covers the economic side. Occupational Employment and Wage Statistics give wages by occupation and area. The Quarterly Census of Employment and Wages counts jobs and pay by industry down to the county. The Current Population Survey, run with Census, measures employment and includes disability questions. Employer Costs for Employee Compensation gives the benefits share of total compensation, the Consumer Price Index includes a medical care component for adjusting health dollars, and the American Time Use Survey measures time spent on unpaid care. Each program defines employment and earnings its own way, so use one consistently within an analysis.

The pitfalls that trip analysts

These are the errors we find most often when reviewing analysis built on public health data. Each one produces a plausible number that is wrong. The first is the most common. Survey weights are not a refinement you add at the end. An unweighted tabulation describes the people who were sampled, not the population.

Combining sources without double counting

Real questions usually need more than one source, and combining them is where the most damaging errors happen. Three rules prevent most of them.

Keep units straight. HCUP counts discharges, MEPS counts people and events, CMS files count services or beneficiaries, and surveillance counts reported cases. A person admitted three times is three discharges. Before adding or dividing, convert every source to the same unit, or keep them in separate lines.

Find the overlap before you add. Sources often describe overlapping slices of the same activity. Medicare inpatient stays appear in both CMS files and HCUP. MEPS spending already includes hospital care, so adding HCUP hospital costs on top counts admissions twice. A case reported by a county is the same case in the state count and the national table. Use one source per slice, and use the others to check it.

Match numerators to denominators. A rate is only as consistent as its parts. Use the denominator built for the numerator's population and geography, the same year, and the same definitions. Mixing a death count by residence with a population from a different vintage, or a Medicare-only numerator with an all-ages denominator, produces a rate that describes nobody.

Write a short source map for every combined analysis: each component, the single source it comes from, the unit, the year, and which other source checks it. If two components point to sources that overlap, the map shows it before the total does.

Before you publish a number

• The question is written down: unit, population, geography and period.
• Each source's design matches the question, and its exclusions are stated.
• Survey weights and design variables are applied, with the correct weight for the subsample used.
• Suppressed and unreliable cells are reported as such, not estimated.
• Every coding or method change in the period is checked for comparability.
• Combined sources are mapped by unit and overlap, with one source per component.
• Every pull is pinned to a file, vintage, query and date.

Frequently asked questions