What a burden estimate is actually for
A burden-of-illness estimate answers one question for someone who has to allocate money: how much does this condition cost the country, and in what currency? Cases, deaths, dollars of care, and work that never happened. The number goes into a budget justification, a grant program's priority list, a council briefing or a congressional request, and it will be read by someone looking for the weakest link. Every component therefore has to point to a public source, and every assumption has to be written down where a skeptic can find it.
The good news is that the United States publishes almost everything needed to build one. National surveys measure who has a condition. Federal expenditure surveys and hospital databases measure what care costs. Death certificates, compiled nationally, count deaths. Labor statistics price the time people lose. None of these sources was designed for your condition, and each one carries a caveat that will move the total if you ignore it. The craft is in assembling them without letting their blind spots stack.
What follows is the sequence we use: pick the design, map each component to two sources, build the cost side carefully, stress the result, and write it up on one page.
Choose the design before you open a dataset
There are two basic designs, and they answer different questions. Mixing them is the most common structural error in published burden work.
Prevalence-based estimates count everyone living with the condition in a given year and total the costs incurred in that year. This is the right design for a budget question: what does this condition cost us annually right now? It is also the easier design to build from public data, because national surveys and expenditure files are organized by calendar year.
Incidence-based estimates follow the people newly diagnosed in a year and total the lifetime costs attributable to their condition, discounted to present value. This is the right design for a prevention question: what would we save by stopping one new case? It requires disease duration, survival curves and a discount rate, and most of those inputs come from the literature rather than from a federal file.
Then fix the perspective. A societal perspective counts medical costs, lost productivity and unpaid caregiving regardless of who pays. A payer perspective counts only what a health plan or program pays. A government budget perspective counts program outlays, including transfers such as disability benefits, which a societal perspective excludes because they move money between people rather than consuming resources. Write the perspective in the first line of the method section. Half of all arguments about a burden number turn out to be two people using different perspectives.
Every component, and the public source that answers it
A full cost-of-illness estimate has five or six components. The table maps each one to the public U.S. source that most often answers it, an independent second source to check it against, and the caveat that most often distorts it.
| Component | Primary public source | Independent check | Main caveat |
|---|---|---|---|
| Prevalence | NHIS (self-reported diagnosis) or BRFSS for state-level estimates | NHANES where a lab or exam measure exists; claims-based counts from CMS public files | Self-report counts diagnosed cases only; surveys exclude people living in institutions |
| Medical spending | MEPS household and event files, using an incremental-cost model | HCUP inpatient and emergency department databases plus CMS utilization files | MEPS sample is thin for rare conditions; HCUP records charges, not payments |
| Deaths | CDC WONDER multiple cause of death files | NHIS linked mortality files for risk relative to people without the condition | Chronic conditions are often left off death certificates; small counts are suppressed |
| Lost work (morbidity) | MEPS and NHIS questions on missed work days and work limitation | CPS employment rates by age, sex and disability status | Self-reported absence is noisy; presenteeism is rarely measured |
| Earnings value | BLS Occupational Employment and Wage Statistics | BLS Employer Costs for Employee Compensation for the benefits share | Average wages overstate losses for groups with lower employment |
| Unpaid caregiving | American Time Use Survey time spent on care | Condition-specific caregiver surveys in the literature | Hard to attribute care time to one condition |
Read the right-hand column before you read anything else. Each caveat tells you which direction that component is likely biased, and a total built from six components biased in the same direction is not a conservative estimate, whatever its author calls it.
Prevalence: the number everything else multiplies
Prevalence sits at the top of every chain. A 20 percent error in prevalence is a 20 percent error in medical costs, lost work and caregiving at the same time, so it deserves more scrutiny than any other input.
National household surveys ask whether a doctor or other health professional has ever told the respondent they have the condition. That is a count of diagnosed, self-reported cases among people who live in households and answered the survey. It misses people who are undiagnosed, people in nursing homes and other institutions, and people too ill to respond. For a condition with a lab marker, NHANES lets you measure the undiagnosed share directly, because it examines and tests a national sample. For conditions without a marker, the undiagnosed share has to come from the clinical literature, and it should be presented as a range.
Claims-based counts are the natural second source and the natural disagreement. A claims count includes only people who sought care, were coded with the condition, and are enrolled in the plan whose data you hold. It is a floor. When the survey estimate and the claims count differ, do not average them. Explain the gap. The explanation is often the most useful paragraph in the whole report, because it tells the reader who is being missed.
For state or county estimates, BRFSS supports state-level prevalence for the conditions it asks about, and CDC's PLACES project publishes model-based county and tract estimates derived from it. Model-based small-area estimates are useful for targeting and should be labeled as modeled, because they borrow strength from demographics and will smooth over genuine local differences.
Medical spending: attributable cost, not total cost
The most frequent inflation in cost-of-illness work comes from totaling everything spent on people who have the condition. People with one chronic condition usually have others. If you sum their total spending, you have counted their heart disease, their diabetes and their knee replacement as costs of the condition you were asked about.
The defensible approach is incremental cost. Using MEPS, compare annual spending for people with the condition against people without it, holding age, sex, insurance, region and other conditions constant. A two-part model is standard: one part for the probability of any spending and a second for the amount among those who spend. The difference between predicted spending with and without the condition is the attributable cost per person. MEPS also allows a cruder event-level approach, summing only the medical events households linked to the condition, which tends to undercount because people under-attribute.
MEPS has two limits worth naming. It covers the civilian population living in households, so nursing home and long-term institutional care fall outside it. And its sample is modest, so a rare condition may have too few respondents in a single year. Pooling several years helps, at the cost of turning your estimate into a period average, which you then state plainly.
HCUP is the check on the hospital side. The National Inpatient Sample and the Nationwide Emergency Department Sample give national counts of stays and visits with the condition as a principal or secondary diagnosis. They record hospital charges, which are list prices, not what anyone paid. Convert them with the cost-to-charge ratio files HCUP provides, and say that you did. HCUP data are purchased under a data use agreement, so budget the time for that before the work plan assumes you have them.
Charges are not costs
A hospital charge is a list price. Payments are what payers actually paid, and costs are what the hospital spent to deliver the care. They can differ by a large multiple. Totaling raw HCUP charges into a cost-of-illness estimate is the single fastest way to produce a number that a health economist will dismiss on sight. Convert charges to costs, or use payment data, and label which one you used in every table.
Deaths and years of life lost
CDC WONDER provides death counts compiled from every death certificate filed in the United States, queryable by cause, age, sex, race, place and year. Two choices drive the result.
First, underlying cause versus multiple cause. The underlying cause is the single condition that started the chain of events leading to death. Multiple cause counts every condition listed anywhere on the certificate. For a condition that kills directly, underlying cause is right. For a chronic condition that contributes to deaths attributed to something else, underlying cause will undercount badly, and multiple cause will overcount by crediting every death where it was merely present. Report both and explain which one your total uses.
Second, small numbers. WONDER suppresses sub-national counts below a threshold to protect privacy and flags rates based on fewer than 20 deaths as unreliable. A county-level mortality estimate for an uncommon condition will be mostly suppressed cells. Aggregate up to a state or region, or pool years, rather than imputing values into suppressed cells.
Deaths become years of life lost by subtracting age at death from remaining life expectancy at that age, which NCHS publishes in its life tables. If the estimate will value those years in dollars, use the method and values your agency's guidance specifies and cite them, because valuation of life years is where reviewers expect to see a reference rather than a judgment call.
Productivity and the costs nobody bills
Indirect costs are often larger than medical costs for conditions that strike working-age adults, and they are where estimates diverge most between authors.
The standard human capital method values lost time at market wages. Take the work days lost to absence, reduced hours and labor force exit, and multiply by a compensation rate. BLS Occupational Employment and Wage Statistics supply wages by occupation and area; BLS Employer Costs for Employee Compensation supplies the share of compensation paid as benefits, which should be added because the employer loses the whole cost of the worker, not just the paycheck. The friction cost method, which counts only the period until a worker is replaced, gives much smaller numbers and is worth reporting as a sensitivity case.
Employment loss is the larger and harder piece. The defensible approach compares employment rates for people with the condition against similar people without it, using survey data, and attributes the difference. Applying the general population's employment rate to everyone with the condition overstates losses for groups who would have had lower employment anyway, which is why age and sex adjustment matters here more than anywhere else.
Unpaid caregiving is real resource use and is usually omitted. The American Time Use Survey measures time spent caring for household and non-household members. Attributing that time to one condition is difficult, so treat caregiving as a separate line with its own range rather than folding it into the headline.
A worked structure, with illustrative numbers
The table below shows the arithmetic of a prevalence-based, societal-perspective estimate for a hypothetical condition. Every input is illustrative and invented for the example. None describes a real disease. The point is the structure: each line names its input, its source type and its calculation, so a reviewer can change any single assumption and see the total move.
| Line | Illustrative input | Calculation | Illustrative result |
|---|---|---|---|
| Adult population | 250 million adults (illustrative) | Denominator for the estimate year | 250,000,000 |
| Prevalence | 0.8% of adults (illustrative survey estimate) | 250 million × 0.008 | 2.0 million people |
| Attributable medical cost | $4,000 per person per year (illustrative incremental model) | 2.0 million × $4,000 | $8.0 billion |
| Labor force exit | 60% would otherwise be employed; 10% of those leave work (both illustrative) | 2.0 million × 0.6 × 0.1 × $70,000 compensation | $8.4 billion |
| Absence among those still working | 6 extra days a year at $280 a day (illustrative) | 1.08 million × 6 × $280 | $1.8 billion |
| Premature deaths | 3,000 deaths, 15 years lost each (illustrative) | Reported in life years, valued separately | 45,000 life years |
The illustrative total for the three dollar lines is $18.2 billion, before caregiving and before any value assigned to the 45,000 life years. Notice how the structure exposes the drivers. Halve the labor force exit assumption and the total falls by $4.2 billion, more than any other single change. That tells you where to spend your validation effort: on the employment comparison, not on refining the medical cost to the nearest hundred dollars.
Double counting and mixed years: the two quiet inflations
Most indefensible totals are not wrong in any single line. They are wrong in how the lines were added. These are the overlaps we check every time.
Medical spending counted twice. MEPS spending already includes hospital stays. Adding HCUP inpatient costs on top of a MEPS total counts every admission twice. Use HCUP to check the inpatient share of the MEPS figure, not as an extra line.
Lost work counted twice. A person who leaves the labor force and then dies early can appear in both the morbidity line and a mortality productivity line for the same years. Assign each person-year to exactly one bucket.
Transfers counted as costs. Disability benefits and other transfer payments are not societal costs; the lost earnings are. Adding benefit payments on top of lost earnings counts the same loss twice. In a government budget perspective the outlays belong, and the lost earnings do not.
Comorbid spending. Total spending of people with the condition includes care for everything else they have. Use incremental cost.
Survey years that do not match. Prevalence from one year, spending from another, wages from a third and population from a fourth produce a number that describes no year at all. Align every component to a single reference year where possible. Where a source lags, inflate dollar values to the reference year with a named health price index and state the index. Watch for breaks: NHIS redesigned its questionnaire in 2019, the United States moved from ICD-9-CM to ICD-10-CM coding in October 2015, and utilization in 2020 and 2021 was distorted by the pandemic. A trend line that crosses any of these needs a note, and sometimes needs to be split.
Sensitivity ranges that mean something
A single point estimate invites false precision. A range built by adding ten percent to every input invites dismissal. The useful middle is a short set of sensitivity cases tied to the real uncertainties you documented.
Start with one-way sensitivity on the three inputs that move the total most, which the worked structure will have shown you. Use the bounds each source actually supports: the survey's confidence interval for prevalence, the alternative model specification for incremental cost, the friction method for productivity. Then run a combined low case and high case that set every driver to its plausible bound at once. If the reader has time for only one sentence, it should be the range, not the midpoint.
For an estimate that will be scrutinized hard, a probabilistic sensitivity analysis that samples each input from a distribution is worth the effort. It also forces you to state a distribution for every input, which is a useful discipline even if nobody reads the histogram.
Document every assumption
A burden estimate is a chain of borrowed numbers, and the reader is entitled to see every link. The estimate will be challenged months later by someone who was not in the room. Whether it survives depends entirely on whether the assumptions were written down at the time. This is the minimum documentation we keep for every burden model.
- Design and perspective stated in one sentence each.
- Reference year, and the price index used to bring every dollar to it.
- For each component: source, file name, release, years used and the exact variable or query.
- Case definition: the codes, survey questions or lab thresholds that define having the condition.
- Every value that came from the literature, with the citation and why that study was chosen.
- Every overlap considered and how it was resolved, including the ones that turned out not to matter.
- Suppressed or unreliable cells and how they were handled, without imputation into suppressed values.
- The sensitivity cases, their bounds, and the source of each bound.
Keep the analysis in code, not in a spreadsheet with pasted values. A script that pulls each source, applies each assumption and produces the summary table can be rerun when a new survey year arrives, and it is its own audit trail.
The one-page summary a decision-maker needs
The full method may run to thirty pages. The person deciding will read one. That page should hold the headline range with the midpoint, the share of the total in each component, the three assumptions that move the number most and what happens if each is wrong, the population and conditions excluded, and the perspective. Add one sentence on what the estimate is not: it is not a forecast, not a measure of what an intervention would save, and not comparable to estimates built on a different perspective.
Include a simple stacked bar of the components. Decision-makers grasp immediately that a condition whose burden is mostly lost work calls for different policy than one whose burden is mostly hospital care. That one observation is often worth more to the reader than the total itself.
Frequently asked questions
Yes, for most conditions common enough to appear in national surveys. Prevalence, spending, deaths and wages all have public U.S. sources. The weak points are rare conditions, where survey samples are thin, and conditions with a large undiagnosed share, where the undiagnosed fraction has to come from the clinical literature and be presented as a range.
Use prevalence-based design when the question is what the condition costs each year now, which is most budget and priority-setting questions. Use incidence-based design when the question is what preventing a new case would save, which requires lifetime costs, survival and discounting.
Because people with one chronic condition usually have others, and their total spending includes care for all of them. An incremental-cost model compares similar people with and without the condition and attributes only the difference.
Do not impute into suppressed cells. Aggregate to a larger geography, pool several years, or report the estimate at the level the data support, and say which you did. Suppression thresholds exist to protect privacy, and filling them in can defeat that purpose.