A projection under the modeled risk, not a diagnosis

Where the harm lives

Our lead-risk map predicts, neighborhood by neighborhood, where American children are poisoned. Read it forward in time and it becomes a forecast: a short list of the places this country, by doing nothing, keeps poisoning, and the bill that comes due a decade later.

An adjunct to the lead-risk white paper. The full method, every assumption, and its honest limits are linked at the bottom. DetectLead / Fluoro-Spec Inc., 2026.

Projected childhood lead exposure over the next decade. Brightness scales with projected newly-exposed children per county. The bright places are not having a bad year; they are on this list every year. Hover a county for its forecast.
23
counties hold a quarter of the nation's projected future childhood lead harm. 106 hold half.
216,000
children newly exposed in the worst 10% of neighborhoods over the next decade (range 161k–271k).
$6–10B
in lost lifetime earnings there, counted before any medical, special-education, or justice cost.
~$0.5B
to screen every one of those homes once. We are on track to pay the bill instead.

The map is a forecast

We built a map that predicts, neighborhood by neighborhood, where American children are lead-poisoned. It agrees with the measured blood-lead data in every state we could check. Most people read it as a snapshot of where the problem is.

Read it forward in time and it stops being a snapshot. It becomes a forecast. A ranked list, worst to first, of the neighborhoods where this country, if it keeps doing nothing, will keep poisoning children.

That is a stranger and more damning object than a risk map. A risk map describes a condition. A forecast describes a choice, because a forecast can be acted on, and the decision not to act is still a decision.

The harm has an address

Here is the part that should be intolerable: this is not a national fog of slightly elevated risk that everyone shares a little of. It is concentrated, hard, onto a short list of specific places, and it stays there.

Twenty-three counties hold a quarter of the entire country's future childhood lead harm. One hundred and six counties, under four percent of the counties in America, hold half of it. Drop to the neighborhood level and it tightens further. The worst ten percent of neighborhoods carry thirty-seven percent of the harm while holding under a fifth of the young children, so children born into them face, on average, about twice the national-average risk before their first step.

The harm clusters tighter than the children do. That is the whole thing in one sentence. If lead exposure were only a matter of how many kids live somewhere, the map of harm would match the map of children. It does not. It pulls into a few thousand neighborhoods and a few dozen counties and concentrates, because the thing that drives it, old deteriorating paint stacked on top of poverty, is itself concentrated there and does not move.

And it is the same places every year. The county that is on this list in 2026 is on it in 2027 and in 2031, because nothing about the housing changed, only the children did. The concentration is not a snapshot of bad luck spread thin across the country. It is a standing feature of a few hundred neighborhoods that the rest of the country has quietly learned to route around.

Lead poisoning in these places is not an event. It is a process.

The housing does not change. The paint laid down before 1978 is still on the windowsills and the door frames, still turning to dust every time a sash slides. What changes is the children. A new cohort is born into the same housing every year, crawls the same floors, teethes on the same sills, and a predictable fraction of them absorbs lead and loses a measurable piece of who they would have been.

So the worst neighborhoods on the map are not having a bad year. They are on a trajectory. The same houses are on track to expose next year's babies at close to the rate they exposed last year's, and the year after that, and the map can see it coming the way you can see a slow leak filling a basement.

What the trajectory costs, in numbers

Take the worst ten percent of neighborhoods on the map and run them forward one decade, at the elevated blood-lead rates actually observed in those neighborhoods' tested children, against the real count of young children the Census says lives in them. The forecast for those places alone, over the next ten years:

  • More than 200,000 children newly exposed to harmful levels of lead. The honest range runs from about 160,000 to 270,000. The uncertainty is in the size of the number, not the direction of it.
  • Close to half a million IQ points lost across them. Permanent. There is no treatment that returns them.
  • At least $6 billion in lost lifetime earnings, and on central assumptions closer to $10 billion. Either way it is earnings only, counted before a single dollar of special education, medical management, or criminal-justice cost that the same science attributes to childhood lead.

Widen the lens to the worst twenty-five percent of neighborhoods and the decade's toll is around 350,000 children and somewhere between $10 and $17 billion in lost earnings alone.

These are not the historical totals. The historical bill for American childhood lead exposure is already measured in the hundreds of millions of IQ points. This is the part we have not paid yet. The worst tenth of the country's neighborhoods is on track to lose another half-million IQ points of its children's futures in the next ten years, and we can name the states where most of it happens: New York, California, Pennsylvania, Illinois, Ohio, Massachusetts, Michigan.

Why "doomed" is the right word, and why it is a scandal

A thing is a tragedy when no one could have stopped it. This is not that. Doomed is a word about the trajectory, not about any single child. No child's future is sealed by a map, and any one of them can be the exception. What is all but sealed, if nothing changes, is the pattern: these same neighborhoods keep producing newly exposed children, year after year. Four facts sit on top of each other here, and each one alone would be damning:

  1. It is foreseeable. We have the map. We can rank the neighborhoods in advance.
  2. It is preventable. Finding the hazard before a child meets it costs a few dollars a home. The science on prevention is settled and lopsided, returning seventeen to two hundred and twenty dollars for every dollar spent.
  3. It is permanent. A poisoned child does not recover the lost points. The damage is taken once and kept for life.
  4. It is deferred. The cost does not land on whoever could have prevented it. It lands fifteen and twenty years later, on a school system, a hospital, a court, an employer, and most of all on the child, who never knew there was a version of themselves with the lead left out.

Put those four together and you get a machine that runs forever. The people with the budget to prevent the harm are not the people who pay for it, and by the time the bill arrives the prevention window has been closed for two decades. So nothing moves. The same neighborhoods produce a fresh cohort of exposed children every year, the map predicts exactly which ones, and the trajectory just continues because no one with the power to break it is the one holding the receipt.

The cleanest way to see the scandal is to put the two prices next to each other. Screening every home in the worst ten percent of neighborhoods, one time, costs on the order of half a billion dollars. The lost earnings from not doing it, in those same neighborhoods over the same decade, run to at least six billion and likely ten. We are choosing the expensive option. In advance. With the map in our hand.

The honest boundary

This is a projection, not a diagnosis. The map predicts where risk concentrates from public housing and poverty data. It does not test any one child or certify any one home, and the numbers above are a scenario built on a chain of reasonable assumptions, each of which carries error: how risk maps to exposure, how exposure maps to IQ, how IQ maps to earnings. Every figure here should be read with a band around it, not a decimal point. We will show that band, not hide it.

Several of the assumptions cut against each other, which is the honest reason to trust the order of magnitude even while distrusting the last digit. The exposure rates come from children who were actually tested, who skew toward higher risk, which pushes the count up, and holding today's child counts and exposure rates flat for ten years, when both are slowly falling, pushes it up further. Cutting the other way: the dollar value counts only lost earnings and ignores the medical, educational, and justice costs, and discounting future earnings more aggressively than the standard rate would roughly halve the dollar figure. And the map is known to run conservative in predominantly Black neighborhoods, where it under-states risk, so the true burden in the places already carrying the most of it is likely higher than this forecast, not lower.

None of that softens the conclusion. It sharpens it. The uncertainty is in the size of the number, never in the sign.

The turn

A burden map is not a darker version of the risk map. It is the risk map with the consequences left in.

We usually strip the consequences out. We say "ninetieth-percentile risk" because it sounds measured and objective, and a percentile asks nothing of anyone. But the consequences are the entire reason the map is worth building. Leaving them in is not editorializing. It is the honest thing, because a neighborhood's children losing a quantifiable share of their futures, on a schedule we can read in advance, is the actual content of the word "risk."

So we are going to put the consequences back. Not as a guess dressed up as a measurement, but as a clearly-labeled forecast of what the next decade costs the places we have already decided, by doing nothing, to keep poisoning. The map told us where. This says what happens there next, and what it is worth to change it.

Methodology and honest limitsEvery input, its distribution and source, and the direction each known bias pushes. Click to open.

What this is, and is not

It is a projection of where childhood lead exposure concentrates over the next decade, and what that exposure costs, built by joining three things: the validated national risk index, the elevated-blood-lead rates actually observed in states that publish them, and the real count of young children the Census records in each neighborhood.

It is not a diagnosis of any child or any home, not a measured case count, and not validated to the standard of the risk paper. The risk map is validated science (Spearman 0.48 to 0.77 against measured blood lead in nine states). This is a scenario layered on top of it, and it inherits all of the risk map's limits plus new ones from each additional step.

Inputs

Input Source Notes
Tract risk percentile r The validated model (tract_risk.json, 83,388 tracts) 0.58 * z(housing-age risk) + 0.42 * z(poverty), ranked to a national percentile
Young-child counts ACS 2018-2022 5-year, table B01001 (B01001_003E + B01001_027E, ages 0-4) Real per-tract counts; national total 19.0M. This is the under-5 (five single-year ages) sum, on the high side of the ~18.4M 2020 decennial because the 5-year file pools pre-decline years. It is not the larger 0-5 six-age figure (~22M) that some tables report.
Observed elevated-BLL rate MI, OH, WI tract-level validation data (ob field) Share of tested children at or above each state's reporting threshold. Twelve Wisconsin tracts carry small negative values (cell-suppression artifacts); these are clipped to zero. Thresholds and periods differ by state (see biases); pooling them is a known limitation.

The projection, per tract

annual_cohort = under5 / 5 # children entering the 0-4 window each year elevated_fraction = calib(r) * selection # expected share of a cohort that becomes elevated newly_exposed/yr = annual_cohort * elevated_fraction newly_exposed_10yr = newly_exposed/yr * 10 IQ_lost = newly_exposed_10yr * iq_per_case earnings_lost = IQ_lost * dollar_per_iq

calib(r) is the observed elevated-BLL rate as a function of national risk percentile, estimated by a local-window average (+/- 0.03 in percentile space, minimum 80 neighbors) over the pooled MI/OH/WI validation tracts. It rises from about 1 percent in the lowest-risk neighborhoods to roughly 12 percent in the highest. Calibrated separately, the three validation states agree closely (at the 95th risk percentile: Michigan 8.4, Ohio 10.0, Wisconsin 10.0 percent; and they track each other across the range), so pooling them does not distort the curve despite differences in reporting threshold and period.

Parameters and their distributions

Every uncertain input is a distribution, not a point. The systematic parameters (selection, dose, value) are drawn once per Monte Carlo iteration and shared across all tracts, because their uncertainty is national and does not average away across neighborhoods. This produces wider, more honest intervals than treating them as independent per tract.

Parameter Distribution Basis
selection Triangular(0.50, 0.80, 1.10) Converts "elevated among tested" to "elevated among all children." Tested children skew higher-risk, which argues for a factor below 1; legacy reporting thresholds (mostly >=5 ug/dL) understate against the current 3.5 ug/dL reference value, which argues for a factor above 1. Peaked at 0.80, not a flat box. Modeled as one national factor, which under-corrects in the very highest-risk tracts where testing is most selective, so it runs conservative-upward there.
iq_per_case Triangular(1.0, 2.0, 4.0) IQ points lost by an elevated child vs an unexposed one. Lanphear 2005 pooled analysis: a rise from 2.4 to 10 ug/dL is a 3.9-point loss (95% CI 2.4 to 5.3), steepest at low exposure. Applied per case without scaling by how far above threshold the child sits, a deliberate simplification given the curve's supralinear low-dose slope.
dollar_per_iq Uniform($14,500, $28,000) Present-value lifetime earnings lost per IQ point. Grosse 2002 base case ~$14,500 (2000 USD); Gould 2009 ~$17,815 (2006 USD), ~$28,000 inflation-adjusted to present. Earnings only, discounted at ~3 percent (the literature standard); at a 7 percent rate the figure roughly halves, reported separately below.
horizon 10 years (fixed) Ten birth cohorts.
Conservative floor newly_exposed * $22,000 The risk paper's own rounded, earnings-only per-case figure, computed on the central calibration with selection = 1. Reported as an "at least" line.

Monte Carlo: 3,000 draws. Within each draw the validation tracts are bootstrap-resampled (so the calibration curve carries its own sampling error), and the systematic parameters are drawn. Reported as median with 5th-to-95th-percentile (90 percent) interval.

Known biases, and the direction each pushes

This is the part a hostile reader should go to first.

  • Testing-selection (pushes estimates UP). Observed rates are among children who were actually tested, who are targeted for being higher-risk. Among all children the rate is lower. The selection factor's lower range (down to 0.50) is the correction; the headline already sits below the raw observed rates.
  • Legacy thresholds (pushes estimates DOWN). The MI/OH validation data are from 2005-2018, when the reporting threshold was typically 5 ug/dL or higher. At the current 3.5 ug/dL reference value, more children count as elevated, so the observed rates understate today's definition.
  • Earnings-only valuation (pushes estimates DOWN). The dollar figures count lost lifetime earnings and nothing else. They exclude special education, medical management, lost parental productivity, and criminal-justice costs that the same literature attributes to lead. True societal cost is higher.
  • Conservative in Black neighborhoods (pushes estimates DOWN where it matters most). The risk model under-predicts measured risk in predominantly Black neighborhoods (documented in the residual analysis). The true burden in the places already carrying the most of it is likely higher than this forecast, not lower. The map is not used, ever, to target by race; this is reported only as a direction-of-error.
  • No spatial-autocorrelation adjustment (widens the true interval). The risk model's validation correlations are not corrected for spatial autocorrelation, so its precision is somewhat overstated. This carries through here. It affects confidence in the calibration, not the central magnitude.
  • Static population, held flat for ten years (pushes estimates UP). The projection multiplies today's under-5 count by ten cohorts. U.S. births and the under-5 population have fallen since about 2015, so holding 2022 counts flat overstates the cumulative count.
  • Static exposure rates, held flat for ten years (pushes estimates UP). Blood-lead levels have declined for decades and keep falling under abatement, water-line replacement, and tighter dust rules. Holding the observed calibration flat ignores that decline.
  • Discount rate (pushes the dollar figure DOWN if raised). Earnings use the ~3 percent literature-standard discount rate. At a strict 7 percent the dollar figures roughly halve; the count and IQ figures are unaffected.
  • Ecological inference (a ceiling on what any of this can claim). Risk, exposure, and cost are all neighborhood-level. None of it describes an individual child. A high-risk neighborhood is a statement about housing stock and poverty, not about any one address.

These pull in both directions. Selection and the held-flat-for-a-decade assumptions push up; the legacy thresholds, the earnings-only valuation, the conservative race residual, and a higher discount rate push down. That two-sided pull is the honest reason to trust the order of magnitude while distrusting the last digit.

Headline results (10-year, national)

Segment Newly exposed children (median, 90% CI) IQ points lost Lifetime earnings lost, central (90% CI) Floor, at least
Worst 10% of neighborhoods 216,000 (161k-271k) ~484,000 $10.1B ($5.4-18.0B) $6.0B
Worst 25% of neighborhoods 357,000 (266k-448k) ~800,000 $16.6B ($9.0-29.7B) $9.9B
Entire United States 736,000 (547k-923k) ~1.65M $34.4B ($18.6-61.5B) $20.4B

All dollar figures use the ~3 percent literature-standard discount rate. At a strict 7 percent rate they roughly halve (worst-10% central ~$5.0B, floor ~$3.0B; entire-U.S. central ~$17.2B).

Concentration (real child counts, central calibration): the worst 10 percent of neighborhoods carry 37 percent of the projected harm; 23 counties hold a quarter of it; 106 counties (3.4 percent of all U.S. counties) hold half. Harm Gini 0.48 versus a young-child Gini of 0.37, meaning the harm clusters tighter than the children do.

Reproducibility

  • firm_pull.py pulls real ACS under-5 counts (keyed, by state) and merges to tract_risk2.json.
  • burden_solid.py runs the Monte Carlo and writes burden_results.json.
  • burden_concentration.py computes the concentration statistics and per-county harm (county_harm.json).
  • make_harm_map.py renders the map from county_harm.json.

All inputs are public (ACS, the state validation data, the published risk index). No personal or individual-level data is used at any step.

The map finds the risk. A test confirms it.

This forecast says where, and what it costs. It does not diagnose a child or a home. A measurement does that.

Read the white paper Open the national map