How the ZipVitals Score is built

The ZipVitals Score recipe

The Score is a weighted sum of nine domain sub-scores — not a flat average of ~100 raw numbers. Each metric is put on the same scale, capped so no single outlier dominates, combined within its domain, then combined across domains.

1. 1Standardize. Every metric is converted to a z-score against the national ZCTA distribution, so a percent, a rate, and a distance become comparable.
2. 2Cap outliers. Each z is winsorized to [-3, +3] before weighting — right-skewed exposure metrics (toxic-release burden) otherwise throw z ≈ +40 spikes that would wreck the scale.
3. 3Build domain sub-scores. Metrics combine within a domain weighted by source, then equally within a source — so NaNDA's 66 metrics can't swamp ACS's 8 inside Social Need.
4. 4Weight across domains. The nine domain sub-scores combine by the published weight table below into a single composite z.
5. 5Rescale to 0–100. An empirically-fit linear rescale (not a fixed 50 + 10z) targets a readable spread; higher is healthier. The center and spread are fit on the full-coverage ZIPs only — see "what the scale is calibrated on" below.
6. 6Renormalize for coverage. A ZIP missing an entire domain has the remaining weights renormalized to sum to 1, and the confidence band widens.
7. 7Rank-eligibility floor. A ZIP with 4 or fewer of 9 domains present shows "limited data" and is not ranked (see the ranking rules below).

Domain weights — published in full

These are the current weights (locked 2026-07-08). Life expectancy is held out of the composite and used as an independent validator; the nine scoring domains renormalize to sum to 1 (each effective = nominal ÷ 0.95). The bar encodes the effective weight — relative magnitude only, never a running total.

Collinearity — the honest caveat on these weights

We also use crude, not age-adjusted PLACES prevalence — CDC publishes no age-adjusted estimates at ZCTA level (they are 100% null there) — so an older ZIP may score modestly worse on disease burden partly because of its age structure rather than its health.

Income and the Score

Median household income accounts for about 61% of the variation in the ZipVitals Score across the 21,774 ZIP codes that carry both it and a life-expectancy estimate. Wealthier ZIP codes really do score higher, and by a large margin. Income is also one of the Score’s own inputs, carrying about 0.4% of its weight directly, and several domains read economic conditions besides — see the collinearity note above.

On the same ZIP codes, income accounts for about 28% of the variation in life expectancy — the health outcome we hold out of the Score entirely. Our index tracks income more closely than mortality does, which is a real limit on it. We publish the whole comparison, including what it means for reading a single ZIP code’s Score, rather than summarising it here.

What income explains about a ZIP code’s health — the full disclosure

How we validated the Score

We make a deliberately modest claim. The Score clears a non-inferiority floor — it adds out-of-sample predictive signal beyond the CDC/ATSDR Social Vulnerability Index when predicting life expectancy — and every domain contributes incremental cross-validated signal. The domain-incremental result is the one we lean on. We read the headline “beats SVI” margin conservatively: life expectancy shares variance with the chronic-disease domain, so part of that margin reflects the overlap rather than independent signal. We also ran one pre-registered check for an age artifact in the ranking: if low scores were mostly age, the lowest-scoring ZIPs would skew old. They do not — median age in the bottom decile is 41.4 years against 43.2 nationally, and the correlation between a ZIP’s rank and its median age is −0.06. Read that narrowly. It says the national ranking is not an age ranking in disguise; it does not undo the crude-prevalence caveat just above, which is a per-ZIP artifact of the underlying CDC estimates that no check of ours can remove. This is a floor, not a claim of optimal weights — no ZCTA-scale gold standard exists, so our defense is transparency plus face validity.

Coverage & gaps

Not every ZIP is measured equally well, and some places our data simply stops. This is the part most ranking sites hide. First, how we label confidence; then every named limit.

Confidence bands

The confidence band reflects within-domain measure coverage, not just how many domains are present — so a ZIP whose domains are single-measure shells is correctly banded lower even if the domain “exists.” About 73% of the ~32,940 ZCTAs are “full.”

Full

All nine scoring domains present and well-populated. Eligible for national listicles.

High / Moderate

Most domains present; some partial. Ranked, with the band shown.

Limited

Four or fewer domains — score not computed, no rank, excluded from lists (see the ranking rules).

Where our data stops

Every one of these is a real, named limit, so you can judge exactly how far to trust a given number.

How we describe our numbers

Not every number on this site was arrived at the same way, and the difference changes what you can conclude from it. We use three labels, and we apply them consistently.

Observed

Something was actually measured or counted in this ZIP — most often a household survey like the Census ACS. The honest limit: A survey reflects the people who answered it. At a geography this small the margin of error is real, and non-response is not random.

Model-based

A statistical estimate. A publisher takes a measurement made at a larger scale and predicts a local value from it, usually using census demographics as the bridge. The honest limit: The model reads demographics, not outcomes — two ZIPs with similar demographic profiles get near-identical estimates whether or not their real rates differ. Never read a change between releases as a change on the ground.

Administrative

A record that exists because someone was required to file it — a regulatory violation, a facility on a federal register, a Medicare-participating hospital. The honest limit: It sees only what the reporting requirement captures. The gaps are systematic, not random: a facility below a reporting threshold, or a well nobody is required to test, simply is not in the data.

The ZipVitals Score itself blends all three, so it inherits the widest uncertainty of any input it contains. It is an editorial product built from published data — not a measurement.

Modeled & derived measures — what each number actually is

Several of the numbers we publish are model estimates or rankings rather than things counted in your ZIP, and a few are measurements of something narrower than their name suggests. Read as bare figures they would mislead, so each is stated plainly here — including where the underlying data is resolved more coarsely than a ZIP and simply applied to it.

Rollups, proximity & data vintages

Spatial rollups are population- or area-weighted

Every source that isn’t ZCTA-native is rolled up to ZCTA against a published Census boundary file, weighted by population or land area as appropriate — never a nearest-match shortcut.

Proximity is measured to the ZCTA polygon edge

Distances (to a hospital, a toxic-release facility) are measured to the ZCTA boundary, not its centroid, as a straight-line distance rather than a driving route — so a facility just outside a large rural ZIP reads near zero.

Food access is a derived signal, not USDA’s designation

Our access legs are in-ZCTA store density proxies, not USDA’s distance-to-supermarket measure. Missing-leg ZIPs read "not determinable," never "not a food desert."

Vintages and licenses are disclosed here, per source

Data vintages differ by source — PLACES is the 2025 release (2023 BRFSS), ACS is the 2020–2024 5-year, USALEEP carries a 2010–2015 mortality vintage (disclosed as stale), AirToxScreen is 2020, FARS is 2015–2024. Six of eleven NaNDA datasets are CC BY-NC 4.0 and carry a 2021 NETS vintage — both the non-commercial license and the 2021 vintage are disclosed where those metrics surface. Per-source release years live on this page (not in per-ZIP prose).

Place names: GeoNames (CC BY 4.0) · US Census Bureau — CC BY 4.0 license. Census place files are US-government public domain.

How ZIP codes roll up to metro areas and states

When we compare places larger than a ZIP — a metro area, a state — we are stacking up its ZIP codes. Two rules govern how, and both have consequences you can see on the page.

Only ZIP codes that actually sit close together are compared

Where we measure how much health varies between neighboring ZIP codes, a pair counts only when the two are within 5 miles of each other. Lists of "nearest neighbors" are self-scaling: in a dense city the twelve nearest ZIP codes share boundaries, while in open country they can be tens of miles apart. Without the distance rule the measure would quietly report how spread out a place is and label it health. Both ZIP codes in a pair must also have complete data for every health domain, which is why some states are listed as held back rather than ranked.

Each ZIP code belongs to exactly one metro area

A ZIP code is assigned to a single metro area — the one holding the largest share of its population — so a ZIP code straddling a metro boundary contributes all of itself to its assigned metro and nothing to the other. That is a real simplification: 4,648 of 24,471 metro assignments (about one in five) are split across more than one metro area, 1,197 of them sitting below a 75% share and the lowest around 35%. We record the actual share for every ZIP code rather than smoothing the split away, and a ZIP code in no metro area is left out entirely rather than filed into a rural stand-in.

A state’s standing on one health domain is an average of its ZIP codes’ standings

Each state page shows where that state sits on each of the nine health domains. The figure is the average of its ZIP codes’ standings on that domain — how each ZIP code compares with ZIP codes across the country — weighted by how many residents live in each one, so it describes the state’s people rather than its map. The comparison set is the ZIP codes measured across all nine domains, not every US ZIP code, and 50 marks the middle of that set. A ZIP code counts toward a domain only if enough of that domain is actually measured there; a whole domain is held back and labelled "not enough data" unless the ZIP codes carrying it are where a majority of the state’s residents live. Two states are affected today: Pennsylvania and Kentucky, where the CDC does not publish most of its survey-modeled ZIP-level measures, so several domains are shown as held back rather than averaged from the handful of ZIP codes that have them.

Ranking & listicle rules

Low-coverage ZIPs are not ranked

A ZIP with four or fewer of nine domains renders "score not computed — limited data," gets no national, state, or metro rank, and is excluded from listicles. It still shows on its own page.

Three ranking scopes: the country, the state, and the metro area

A scored ZIP code is ranked three times against three different sets — every scored ZIP code in the country, every scored ZIP code in its state, and every scored ZIP code in its own metro area. The three are computed the same way and broken the same way when two ZIP codes tie, so they never disagree about which of two ZIP codes is healthier. They are not comparable to one another: "#5 of 12" in one metro area and "#5 of 853" in another are placements in two sets that were never merged, so we never rank one against the other or declare a winner across them.

Metro ranks need at least ten scored ZIP codes in the metro area

A metro area is used only when it holds at least ten scored ZIP codes; below that, no ZIP code in it gets a metro rank at all — omitted entirely rather than shown thin. Metro areas here are the federal government’s Metropolitan Statistical Areas, and each ZIP code belongs to exactly one (see how ZIP codes roll up, above). Micropolitan areas are a different federal geography and are not treated as metro areas. A ZIP code in no metro area simply has no metro rank, which is not a gap in the data — there is no metro to rank it in. At the current release this floor withholds a metro rank from 188 of 17,894 metro ZIP codes, about 1 percent.

National lists draw only from full-coverage ZIPs

The highest- and lowest-scoring ZIP codes nationally compare apples to apples — thin-coverage states (PA, KY) appear in their own state lists and on their per-ZIP pages, but not the single national extreme list.

Titles are descriptive, never evaluative

No "best" or "worst." A list is titled by the attribute it ranks, carrying the verb that attribute has earned — "measured" only where the number was actually measured or counted (a Census survey), "modeled" where it is a statistical estimate (CDC PLACES, EPA), "reported" where it is a filed record. So the chronic-disease list is titled "ZIPs with the highest modeled chronic-disease burden," never "highest measured." Both halves are enforced by a build-time title validator.

Framing & comparisons

"Similar ZIP" means a similar health profile

Similarity is the cosine similarity between two ZIPs’ domain sub-scores, not a demographic or income match — taken over the domains both ZIPs have data for, and only where that overlap covers at least six of the nine. That minimum-overlap floor means some ZIPs surface no similar-health neighbor at all — and we say so rather than forcing a weak match.

Geographic neighbors can cross state lines

A border ZIP’s nearby-and-similar block may show an out-of-state ZIP, because adjacency is nearest-by-edge-distance with no state filter.

The voice is descriptive, not a verdict

Bottom-ranked ZIPs are described by named, source-cited attributes — never as a place to "avoid." See the framing statement at the top of this page.

Combined exposures

Some ZIP pages carry a short “combined exposures” note when two separate environmental exposures are each elevated in the same area. It is context, not a new measurement — and never a claim that one exposure causes the other.

Two exposures, established in one place

The note appears only where this area independently ranks among the highest 10% of US ZIP codes for each of the two exposures — each is a fact about the ZCTA itself, established separately, then shown together. Neither number is derived from the other. Both are usually modeled estimates rather than readings taken here, and each is labeled as such where it appears.

A national ranking, headed by its coverage

The ranking behind the note is national: an area qualifies by landing in the highest 10% of US ZIP codes for each exposure, decided across the whole country before any state is considered. We say so in those words rather than calling it "among the highest in your state," which would describe a within-state ranking we never computed. The list we publish is national too, and it is never published bare: these pairs concentrate in specific regions (the granite belt, the Gulf Coast, the farm belt) and we cannot model every ZIP code, so the coverage and concentration figures are shown above the list rather than beneath it. Each state also has its own page with that state's full set.

Federal reference levels, typed correctly

Where an exposure crosses a fixed federal reference, we name the exact number and the instrument type. Arsenic is compared to the 10 µg/L EPA maximum contaminant level (a limit) for public water systems, and reported as a modeled likelihood for private wells, which are federally unregulated. Radon is compared to the 4 pCi/L EPA action level — an action level, not a limit — and shown as modeled radon. We only state that a value crosses a line where the value actually crosses it; otherwise the note stays descriptive ("ranks among the highest 10% of US ZIP codes").

Descriptive, not a health verdict

A combined-exposure note pairs two environmental exposures, never an exposure with a health outcome. It describes conditions in a place; it does not diagnose risk to any resident.

Care access and care deserts

Some ZIP pages, and our state-by-state care-desert lists, identify areas carrying two things at once: an elevated documented burden of health need, and real barriers to reaching care. It is an overlap of two separate readings of public data — never a claim that one causes the other, and never a severity score.

Two counts of six, and nothing is weighted

The need side reads six measures: how many people lack insurance, live in poverty, or face food or housing insecurity, the share reporting fair-or-poor health, and the local burden of chronic conditions — that last one counted once across diabetes, heart disease and high blood pressure together, so a single deprivation cannot be counted three times. The access side reads six barriers: distance to the nearest emergency room, distance to the nearest community health center, how thin the local supply of physicians and of pharmacies is, how many households have no vehicle, and how limited home internet is for telehealth. Each side yields a plain count from 0 to 6. Nothing is weighted and nothing is scaled — a count is just a count.

A measure counts only in the worst 10% of US ZIP codes

Each of those twelve measures counts toward its side only when the ZIP sits in the most-affected 10% of US ZIP codes on that measure — decided across the whole country, before any state is considered. It is the same threshold for every ZIP in America, so the counts mean the same thing everywhere.

The two counts are multiplied, so one side alone is never enough

The two counts are combined by multiplying them and taking the square root. That form collapses to zero unless both sides are elevated: a ZIP with heavy need and easy access, or thin access and low need, does not read as a care desert. The resulting number is internal — we never publish it for a ZIP. What we publish is the tier, and the tiers are counts: acute means at least three measures on each side, severe means at least four on each. Severe means the overlap runs across more of what we measure — never that a place is "worse."

The ranking is national, and so is the list

A ZIP qualifies by where it sits against the whole country, and we say so in those words rather than calling it "among the most affected in your state," which would describe a within-state comparison we never computed. The list we publish is national too, and it is never published bare: need and access compound unevenly by region, so the coverage and concentration figures are shown above the list rather than beneath it. That unevenness is worth naming and not merely implying. The list reaches most states rather than a handful, which is why we publish it nationally at all — but most of the overlap sits in the West and the South, and the severe tier is narrower still: it concentrates in the West and in rural areas, where the access barriers we read are largely distances, and parts of the Northeast carry none of it. Each state also has its own page with that state's full set.

Descriptive, and strictly not causal

These are two things established for the same place. We do not claim need causes the access gap or the reverse, we do not say anyone is underserved because of their health, we do not fuse the two into a single risk number, and we do not rank areas by how bad the overlap is. It describes conditions in a place; it says nothing about any resident of it.

Where it is not computed — and why absence is never "fine"

A measure that simply is not observable for a ZIP is treated as unknown, not as "does not count" — the difference matters, because the second would quietly read a thin-coverage ZIP as having no deficit. Where too many measures are unobservable on either side, the overlap is shown as not computed — limited data, and that ZIP appears on no list. Nor is a ZIP listed that is fully measured and simply does not carry the overlap. Absence from a care-desert list is never a claim that a place is fine.