Carolinas Regional Explorer — Methodology Guide

Methods, definitions, and caveats for every indicator

Author

Kailas Venkitasubramanian

Published

October 21, 2026

Note

Living document. An accessible reference for the methodological choices behind the Carolinas Regional Explorer. Part I covers global methods shared across indicators; Part II is a per-indicator catalog. Update whenever methods change or indicators are added/modified. Engineering details are in the Data Pipeline Technical Documentation.

1 Scope & geography

The Explorer covers the 14-county Charlotte region: 11 North Carolina counties (Anson, Cabarrus, Catawba, Cleveland, Gaston, Iredell, Lincoln, Mecklenburg, Rowan, Stanly, Union) and 3 South Carolina counties (Chester, Lancaster, York). The publish geography is the 2020 Census tract (~752 tracts); indicators are also rolled up to county and region for context.

Every indicator is designed to update annually or biennially. Each carries its source and vintage, a 90% margin of error where applicable, and a reliability flag.

2 Part I — Global methodology

2.1 ACS estimates and comparability

Most indicators come directly from the American Community Survey (ACS) 5-year estimates, published at the tract level. Two properties drive how we handle them:

  • They are survey estimates with sampling error. We always carry the margin of error (Section 2.2).
  • 5-year vintages overlap. The 2020–2024 and 2019–2023 estimates share four years of sample, so adjacent years are not independent. We therefore treat the annual series as a rolling 5-year estimate and only compute change and statistical significance between non-overlapping periods:
Comparable period Endpoint year used
2010–2014 2014
2015–2019 2019
2020–2024 2024

The ACS indicators carry the full annual series 2014–2024 (each year is a rolling 5-year estimate). For a selected tract the app reports two non-overlapping comparable changes, the 5-year (latest−5 → latest) and 10-year (latest−10 → latest), each with the 90% significance test (Section 2.2). These periods are derived from the latest year, so they advance automatically: when 2025 is added they become 2020→2025 and 2015→2025. The non-overlapping endpoints are also derived in the pipeline ({latest, latest−5, latest−10}), so adding a year to region.acs.annual_years is the only change needed.

2.2 Margins of error & reliability

ACS publishes a 90% margin of error (MOE) with each estimate. We convert and propagate it consistently.

  • Standard error: \(SE = MOE / 1.645\).
  • Coefficient of variation: \(CV = SE / |estimate| \times 100\%\).
  • Reliability flag: ok (CV ≤ 15%), caution (15% < CV ≤ 30%), unreliable (CV > 30%).

Propagation (Census formulas). For derived measures:

  • Aggregated counts (sum of estimates): \(MOE_{sum} = \sqrt{\sum_i MOE_i^2}\), with one exception for zero estimates (below).

  • Zero estimates in a sum. ACS publishes a model-based MOE even when an estimate is zero. Adding the MOE of every empty category inflates the result, so when a sum includes zero estimates we count only the largest of their MOEs, once: \[MOE_{sum} = \sqrt{\sum_{i:\,e_i \neq 0} MOE_i^2 + \max_{j:\,e_j = 0} MOE_j^2}\] This is the Census Bureau’s rule (“MOE for zero estimates are based on a model. When approximating a sum, use only the largest zero estimate MOE, once”; Census Bureau ACS margin-of-error training, 2016, slide 26). It applies wherever CRXP adds ACS estimates: categories summed into a numerator or denominator (for example uninsured people across 18 age-by-sex cells), the region’s raw cells summed across counties, and pooled rest-of-region and custom-area estimates. For the region it is applied at two levels (each cell across counties, then the cells into the numerator), which is slightly more conservative than one flat sum. Tract harmonization (Section 2.3) allocates each variable separately and does not apply the rule. Adopted 2026-10-17; before that date every zero category’s MOE was counted.

  • Published direct MOEs for many-cell sums. The root-sum-of-squares formula treats summed cells as independent. Cells of one table are not independent (people move between age-by-sex cells, and the weighting controls their totals), and with many cells the approximation can be far off, as the Census Bureau warns. For three indicators the Census Bureau publishes the same percentage in a subject table with an MOE computed directly from the replicate weights, and CRXP uses that MOE:

    Indicator Cells summed Direct source Median tract MOE, 2024 (derived → direct)
    Health insurance coverage 18 (B27001) S2701, percent insured 17.7 → 5.2 points
    High school graduate or higher 9 (B15003) S1501, percent high school graduate or higher 10.2 → 4.4 points
    Limited-English-speaking households 4 (C16002) S1602, percent limited English-speaking households 3.6 → 3.6 points

    (Mecklenburg County tracts.) The value is still computed from the detailed-table counts. Every row that has a published percentage is checked against it (within 0.15 points, the rounding of published percents), and the build stops if more than 2% of a year’s rows disagree, which would mean the subject-table column measures something else; with 14 counties, a single disagreeing county stops it. The build also stops if the subject-table download fails for some counties, or if an indicator configured for direct MOEs ends up with none. A year whose subject table does not exist keeps the derived MOE, and builds without a Census API key use derived MOEs throughout. To carry the direct MOE through tract harmonization and pooling, it is converted into an equivalent numerator MOE, \(MOE_{num}^{*} = \sqrt{(MOE_{\%}/100 \cdot den)^2 + p^2\, MOE_{den}^2}\), which the proportion formula maps back to the published MOE exactly. Rows whose value differs from the published percentage keep the derived MOE. Five other summed indicators were compared with a published percentage (bachelor’s degree or higher, disability, child poverty, older adults, youth); their median derived MOE is within 20% of the direct one (0.83 to 1.11 times), so they keep the approximation. The other summed indicators have not yet been compared; the app impact report (Plan 0C) covers them. Adopted 2026-10-17.

  • Controlled estimates. Where ACS shows the MOE as ***** (the API’s -555555555), the estimate “has been controlled to be equal to a fixed value and so it has no sampling error. A standard error of zero should be used for these controlled estimates when completing calculations” (ACS Accuracy of the Data, 2019–2023 5-year). County total population and county Hispanic or Latino totals are controlled this way. CRXP uses an MOE of 0 for them (adopted 2026-10-17; before that they were treated as missing), so a county population share carries the MOE of its numerator alone. Controls apply to counties, not tracts, so CRXP reads an MOE of 0 as exact only at the county level: a county comparison in which both sides are 0 is exact and is reported without a test (no_sampling_error). A tract MOE of 0 comes from the proportion formula when every unit falls in the category (for example 13 of 13 homes occupied, 100% ± 0); it is not exact, so such tract comparisons and changes are listed as zero_moe and not tested. Every other annotation (**, ***, -666666666, and so on) still means the MOE is missing.

  • Proportion \(p = num/den\) where the numerator is a subset of the denominator: \[MOE_p = \frac{1}{den}\sqrt{MOE_{num}^2 - p^2\, MOE_{den}^2}\] If the radicand is negative, we fall back to the ratio formula (per Census guidance).

  • Ratio \(R = num/den\) where the numerator is not a subset of the denominator: \[MOE_R = \frac{1}{den}\sqrt{MOE_{num}^2 + R^2\, MOE_{den}^2}\]

Statistical significance of a difference (90%) between two estimates: \[\frac{|e_1 - e_2|}{\sqrt{SE_1^2 + SE_2^2}} > 1.645\] Used only between non-overlapping comparable periods (Section 2.1).

Tip

At the tract level, small-population rates often have high CVs (frequently caution/unreliable). This is inherent to ACS tract data — the app surfaces the MOE band and flag so users interpret accordingly, rather than hiding it.

CDC PLACES publishes a 95% confidence interval instead of a 90% MOE; we convert the CI half-width to a 90%-equivalent MOE (\(\times\,1.645/1.96\)) so reliability is comparable with ACS. PLACES values are never put through the significance test above (Section 2.4.4).

2.3 Geographic harmonization (2010 → 2020 tracts)

ACS 5-year estimates through 2019 use 2010 census tracts; 2020 onward use 2020 tracts. In this region 611 (2010) tracts became 752 (2020) tracts; 135 of the 2010 tracts were split or redrawn (fewer than 99% of their residents fall in a single 2020 tract). To make the time series comparable on a single 2020 geography we build a population-weighted crosswalk: each 2010 tract’s value is allocated to 2020 tracts in proportion to the 2020 Census block population that falls in each piece. Concretely, using the Census 2010↔︎2020 block comparability file, each 2020 block’s population is apportioned across the 2010 tracts it overlaps in proportion to the overlap land area (a dasymetric refinement — a block that straddles two old tracts contributes to both, rather than being assigned whole to the larger overlap); those apportioned populations are then summed to \[afact(t_{10}\!\to\! t_{20}) = \frac{\text{2020 block population apportioned to } (t_{10}\cap t_{20})}{\text{2020 block population apportioned to } t_{10}}\]

Allocation happens on the raw variables before indicator calc:

  • Counts: \(count_{20} = \sum_{t_{10}} count_{10}\cdot afact\).
  • Margins of error: \(MOE_{20} = \sqrt{\sum (afact\cdot MOE_{10})^2}\) (independence approximation).
  • Medians / averages / indices: population-weighted average of the contributing 2010-tract values.

flowchart LR
  T10[2010 tract A] -->|afact 0.6 by population| T20X[2020 tract X]
  T10 -->|afact 0.4| T20Y[2020 tract Y]
  T10B[2010 tract B] -->|afact 1.0| T20X

Why population (not area): area weighting assumes people are spread uniformly within a tract, which badly misallocates tracts that split into pieces of widely different density. For example, a small-but- dense 2020 tract that holds ~21% of an old tract’s people gets only ~0.4% under area weighting but its true ~21% under population weighting. Population weighting is used by default; an area-weighted crosswalk remains as a fallback. This harmonization raised comparable-period coverage from ~480 to ~749 of 752 tracts.

Known limitation — medians. Medians (income, age) are harmonized as a population-weighted average of the contributing tracts’ medians. This is a pragmatic approximation, not the true pooled median: a mean of medians is biased whenever the within-tract distributions differ in shape, and the bias has no guaranteed sign. It matters only for the ~135 changed 2010 tracts in pre-2020 years; the unbiased fix is to reconstruct medians from the afact-weighted binned ACS distributions (e.g. B19001 income brackets, B01001 age brackets) by interpolation, a planned refinement. Readers should treat pre/post-2020 median trends across changed tracts with corresponding caution.

2.4 CDC PLACES time series

CDC PLACES uses multilevel regression and poststratification (MRP) on Behavioral Risk Factor Surveillance System (BRFSS) data to produce model-based tract-level health prevalence estimates. Each annual release is a fresh model fit — not a revision of prior estimates. See Section 2.2 for the CI-to-MOE conversion applied to all PLACES indicators.

2.4.1 Multi-release stitching

CRXP draws on six PLACES releases (2020–2025). Rather than using only the most recent snapshot, the pipeline stitches all six into a per-indicator time series using a newest-release-wins rule: where multiple releases report the same data year, the most recent release’s estimate is retained.

Per-measure cadence varies across the resulting series:

  • Most established measures (obesity, diabetes, smoking, mental distress, etc.) span data years 2018–2023.
  • Some biennial measures appear in odd data years only, e.g. high blood pressure: 2017, 2019, 2021, 2023.
  • Health-Related Social Needs (HRSN) measures (food insecurity, housing insecurity, transportation barriers, social-emotional support) cover 2022–2023 only.
  • Loneliness has 2023 only.

2.4.2 Tract-vintage harmonization (provenance-driven)

PLACES shifted its census-tract geography between releases: the 2020–2023 releases use 2010 census tracts; the 2024–2025 releases use 2020 census tracts. To unify the series on the same 2020-tract geography the app uses, the pipeline applies a provenance-driven approach: every fetched row is tagged with its source release’s tract vintage, and only 2010-vintage rows are harmonized; 2020-native rows pass through untouched. This logic is correct regardless of which data year a given measure happens to report in each release.

Because PLACES publishes rates, not raw counts, allocation uses a population-weighted pseudo-count:

  1. Reconstruct cases: \(cases_{10} = (prevalence_{\%} / 100) \times pop_{10}\), where \(pop\) is the total 2010-tract population (the adult-only denominator totalpop18plus is absent from pre-2024 PLACES releases, so total population is the harmonization weight; the pseudo-count is used only as a weight, never published as a denominator).
  2. Allocate via crosswalk: cases and population are distributed from each 2010 tract to its overlapping 2020 tracts using the committed population-weighted afact weights (the same crosswalk used for ACS — see Section 2.3).
  3. Recompute the rate: \(prevalence_{\%,20} = 100 \times \Sigma\,cases_{20} / \Sigma\,pop_{20}\).
  4. Combine MOEs in quadrature: \(MOE_{20} = \sqrt{\sum (afact \cdot MOE_{10})^2}\).

County boundaries are stable across tract vintages, so county data is not harmonized; county releases are stitched newest-release-wins directly. Because the 2020 county release uses a legacy name-keyed schema with no FIPS codes, the county/region series begins in 2019, one year later than the most-established tract measures begin.

2.4.3 Temporal comparability: PLACES estimates are not for trend analysis

The multi-release series above is assembled so users can read each year’s level in context — it is not a basis for measuring change over time, and the platform deliberately shows no change figure for these indicators. CDC is explicit on this point:

“The current modeling procedure does not support using the estimates to track changes at the local level over time.” — CDC PLACES, Using the Data FAQ (https://www.cdc.gov/places/faqs/using-data/index.html)

Two structural features of the model make year-to-year (or release-to-release) comparison invalid:

  1. Sub-county geographies (tract, place, ZCTA) hold the population distribution fixed. Estimates are poststratified to a fixed census population structure rather than re-based annually, so a single year’s estimate is not a true point-in-time snapshot that can be differenced against another year.
  2. Time is not a model variable. Even at the county level, where annual population estimates are used, time is not included in the model, so the series cannot support trend assessment at any geography.

Compounding this, apparent movement is frequently an artifact rather than real change: the BRFSS questionnaire and case definitions change between cycles (CDC specifically flags a rise in colorectal-cancer-screening prevalence caused by a survey-question change), each release re-fits the model with different BRFSS input years and ACS denominators, and the 2020 BRFSS is distorted by pandemic-era data-collection effects. There is therefore no pair of PLACES releases or data years that can be validly compared for change — not across releases (different models/inputs), not within a release (each measure carries a single year), and not at the county level (time is unmodelled).

How the platform honours this. PLACES indicators are flagged in the data contract with crossReleaseTrend = true. Every downstream view that would otherwise express change is suppressed for them, while the levels and their uncertainty bands are still shown:

Surface ACS / measured indicators CDC PLACES (crossReleaseTrend)
Trend-chart band note “90% MOE; ACS 5-year estimates are rolling — compare only non-overlapping periods” “90% model uncertainty interval; read each year’s level, not the change between years”
Comparable-period change markers (5-/10-year, significance-tested) shown hidden
Significance tests between areas (analysis contract) canonical families none; intervals only (Section 2.4.4)
Custom-area report Change column shown, over each indicator’s own period rendered as “—”
County-profile card delta (“since YYYY”) shown hidden (the level sparkline is retained)

2.4.4 No significance tests for PLACES estimates

CRXP does not run significance tests on PLACES estimates, between areas or between years. Comparisons show each area’s estimate with its interval and stop there; the analysis contract publishes no significance families for PLACES indicators (skipped: model_based_estimates) and lists significance_test among their disallowed uses. Adopted 2026-10-17. The reasons:

  1. The test assumes independent sampling errors, and PLACES errors are not independent. The difference test (Section 2.2) treats the two estimates as independent samples. PLACES estimates come from one multilevel model fit to BRFSS data (Zhang X, Holt JB, Lu H, et al., “Multilevel regression and poststratification for small-area estimation of population health outcomes: a case study of chronic obstructive pulmonary disease prevalence using the Behavioral Risk Factor Surveillance System”, American Journal of Epidemiology 179(8):1025–1033, 2014): every tract and county shares the same fixed effects, and areas in the same county or state share random effects. Their errors are correlated, usually positively, so the independent-error formula misstates the standard error of a difference by an amount that CDC’s published intervals do not let us recover.
  2. The intervals describe the model, not a sample of the area. CDC’s 95% intervals come from Monte Carlo simulation of the fitted model. Most of an area’s estimate comes from its demographic make-up (age, sex, race and ethnicity, poverty) passed through the model, not from survey respondents in that area. Two tracts with similar populations get similar estimates whatever their actual health, so a “significant difference” would mostly restate a difference in demographic composition.
  3. CDC frames comparison by intervals, not tests. CDC’s guidance is that areas “can be compared on individual measures”, that “it is not adequate to simply compare point estimates”, and that “it may not be appropriate to draw conclusions when comparing estimates that have very broad confidence intervals”. The same guidance says the data “should not be used for ranking the overall health” of any area, does not support tracking local change over time, and should not be used to evaluate local programs (https://www.cdc.gov/places/faqs/using-data/index.html).

What users get instead is both intervals side by side. The analysis contract publishes each PLACES value with its 90%-equivalent MOE (converted from CDC’s 95% interval, Section 2.2), and the CRXP agent engine compares two areas by those 90% intervals: where they do not overlap it says the model places one area higher or lower; where they overlap it says the model cannot separate the two areas, never “no difference”. Both are labelled as model-based estimates. Rank intervals and distinctive-feature lists exclude PLACES as well.

This is a direct application of the platform’s core principle — never let a downstream view assert more certainty than the source supports (Section 2.2). PLACES model-based estimates are genuinely valuable for showing where health burden concentrates; they are not a longitudinal instrument, and the app is built to keep users from reading them as one.

2.5 Small-area estimation (current status)

We use a tiered approach:

  • Direct tract estimates (current): ACS published tract tables, the great majority of indicators.
  • Model-based small-area estimates (current): CDC PLACES health measures use multilevel regression and poststratification (MRP) on BRFSS data, published per tract with confidence intervals.

2.6 Spatial statistics

  • Z-scores: per indicator/year, standardized to mean 0, sd 1 across tracts.
  • Class breaks: quantile breaks computed once and held consistent across years so map color means the same value in every year.
  • Spatial weights: k-nearest-neighbor (k = 8), row-standardized, from each tract’s representative point — a point guaranteed to lie inside the polygon (unlike a bounding-box midpoint, which for L-shaped or coastal tracts can fall outside and distort the neighbor graph).
  • Local Moran’s I / LISA: local spatial autocorrelation classified into High-High, Low-Low, High-Low, Low-High, or not significant. Significance is a conditional-permutation pseudo p-value: tract i’s value is held fixed while its hypothetical neighbours are drawn without replacement from the other tracts’ scores (9,999 permutations). The test runs on rank-based inverse-normal scores, \(\Phi^{-1}(r_i/(n+1))\), rather than raw values: many indicators are strongly right-skewed, and a few tracts 7–9 SD above the mean widen the permutation null until almost no clustered tract survives correction (poverty rate, 2024: 7 significant tracts on raw values versus 153 on rank-normal scores). Quadrants therefore mean high or low rank within the region. The permutation count matters too: its p-value floor \(1/(P+1)\) must sit well below \(\alpha/m\), and at 999 permutations BH could reject nothing unless at least 15 tracts hit the floor. Because ~752 tracts are tested at once, the pseudo p-values are then FDR-controlled (Benjamini–Hochberg, α = 0.05) before a tract is called significant — without this correction, dozens of tracts would light up by chance alone. The app consumes these precomputed quadrants; it does not recompute LISA client-side.

2.7 Raster-derived indicators — zonal statistics

Environment indicators come from the USGS National Land Cover Database (NLCD), a 30-metre satellite land-cover product, accessed auth-free as region-clipped GeoTIFFs from the MRLC Web Coverage Service (EPSG:5070). Tract values are area-weighted zonal statistics computed with exactextract, which uses the exact fraction of each pixel falling inside a tract (not just pixel centroids):

  • Class share (forest, farmland, wetlands, developed): the coverage-weighted fraction of a tract’s non-nodata area whose land-cover class is in the indicator’s class set, e.g. forest = NLCD classes 41 (deciduous) + 42 (evergreen) + 43 (mixed).
  • Continuous mean (impervious surface): the coverage-weighted mean of the percent-impervious raster.

NLCD is a wall-to-wall classification, so these indicators carry no sampling margin of error (they have classification accuracy instead); the app shows no error band or reliability flag for them. The Annual NLCD product supplies land cover for most years from 2001 onward, so these indicators carry a multi-year series (the year slider shows the available years) rather than only decadal snapshots.

Light pollution uses the EOG VIIRS Nighttime Lights (VNL V2) annual composites (median-masked): the area-weighted mean radiance (nW/cm²/sr) per tract, for 2014 / 2019 / 2024. EOG downloads are account-gated (the global GeoTIFFs are fetched once, manually or via the EOG OAuth API, then clipped to the region and cached); like NLCD it is a measurement, not a survey, so it carries no MOE. Tree canopy uses the NLCD Tree Canopy Cover (TCC, USFS) continuous percent-canopy product (per-year GeoTIFFs read directly from the supplied zip-of-zips via GDAL /vsizip/, clipped once and cached): area-weighted mean canopy % per tract, annual, 2015–2023. Canopy % is broader than the forest land-cover class (it includes trees in developed/suburban areas), so its values run higher.

2.8 Bivariate analysis

Two indicators are each split into terciles (rank thirds of the raw values; equivalent to terciles of their standardized values) and combined on a 3×3 grid, coloring each tract by its joint class. The legend is interactive: hovering a cell highlights the matching tracts. Useful for seeing where two conditions co-occur (e.g. high poverty × low insurance).

Alongside the map, a correlation scatter plots the two indicators’ standardized (z-score) values (one dot per tract, colored by its 3×3 class), with marginal distributions, a least-squares regression line, and both the Pearson and Spearman (rank) correlations. Spearman is shown alongside Pearson because most indicators are skewed and the choropleth itself colors by rank. Both are presented as descriptive only: tracts are spatially autocorrelated, so the effective sample size is far below the tract count and a naïve significance test would be badly anti-conservative; the scatter therefore reports no p-value. Hovering a dot highlights the tract on the map (and vice-versa) and shows its neighborhood name.

2.9 County and region estimates

County and region values are not computed by averaging tracts — that is biased for rates (it over-weights small tracts) and invalid for medians and indices (a county median cannot be recovered from tract medians). Instead:

  • County = the authoritative published county estimate from each source. County boundaries are stable across 2010/2020, so the engine fetches each county directly from ACS county tables (with the smaller county MOE), the CDC PLACES county model (swc5-untb), or county zonal statistics for satellite measures (land-area-weighted tract values, which equal the county zonal value), and runs the same indicator calculator used for tracts. This gives the correct value and a proper county MOE for every kind.
  • Region = pooled by indicator kind: counts are summed; rates/proportions use the pooled numerator ÷ pooled denominator (with the Census proportion/ratio MOE on the summed components, the MOE variables combined in quadrature); medians and indices use a population-weighted mean of the county values (a documented approximation, as an exact regional median needs microdata).

This is implemented in aggregate.build_county_region, persisted to a county_indicators warehouse table (geoid = county FIPS, plus a REGION row), and consumed by export_crxp.build_aggregates → aggregates.json (countyAvg/regionAvg with countyMoe/regionMoe). The app labels these lines County and Region.

User-supplied tract files (tract_csv). A file brought in through the generic tract source has no published county estimate, so its county and region values are derived from tracts: counts are summed, with MOEs combined in quadrature, and every other kind is a population-weighted mean of tract values. A count total is published only when every tract in the area has a value; with suppressed or missing tracts it is left empty rather than understated. For rates and proportions this matches the pooled rate only when the tract denominators are proportional to total population; for medians and indices it is the same documented approximation used for regional medians. The app’s metadata labels these values as derived. If the file is on 2010 tracts, counts are allocated to 2020 tracts by allocation factor and rates/proportions by population-weighted pseudo-counts (the PLACES method); medians and indices must be supplied on 2020 tracts.

Note. The one remaining simple unweighted mean is the ad-hoc average of a user-selected set of tracts in a custom report (series.js seriesForSet) — a selection summary, not a published geography.

2.10 Analysis contract methods

These methods support the CRXP agent engine and are published in the analysis contract. They do not yet change the explorer.

Rest-of-region comparisons. A county is compared with the rest of the region, not with a region that contains it, because a whole that includes the area is not an independent estimate. For poolable ACS indicators the rest of the region is pooled from the other counties’ numerators and denominators; for medians, indices, and model-based estimates it is the population-weighted mean of the other counties’ values and is flagged as approximate. For poolable indicators, tracts are compared with the rest of the region the same way; for other kinds a tract is compared with the whole region, which includes the tract, and the result is flagged as approximate. Counts (such as total population) get no comparison with the rest of the region, because a part is always smaller than the summed rest; they get change tests only. Controlled county totals have an MOE of 0 (Section 2.2), so county population shares are tested on their numerator’s MOE; where both sides are controlled, the difference is exact and is listed with the reason no_sampling_error instead of a test. Where ACS publishes no county MOE for any other reason, the comparison is listed with the reason moe_not_published and not tested. Model-based PLACES estimates get no significance tests and no ranks (Section 2.4.4).

Canonical significance families. Each indicator, year, and comparison type is tested once per release, and p-values are adjusted with Benjamini–Hochberg across the family (q ≤ 0.10 is reported as significant). Every consumer reads the same results, so whether a difference is significant never depends on how a question is asked. The Census Bureau’s own testing tools apply no multiple-comparison adjustment; this is a CRXP convention.

Constant dollars. Changes in dollar indicators are computed after converting the earlier estimate and its MOE to the later year’s dollars with the BLS R-CPI-U-RS annual average, as Census recommends for comparing ACS dollar values across years.

County rank intervals. Rather than a single rank, each county gets a rank interval: the median and 5th–95th percentile of its rank across 10,000 draws from normal distributions centered on each estimate (SE = MOE / 1.645). This follows Marshall and Spiegelhalter’s recommendation to report the uncertainty of league tables instead of point ranks.

Interval-first reliability (reliability_v2). A CRXP convention, adopted 2026-10-17. It is published in the analysis contract and used by the CRXP agent engine; the explorer keeps the CV labels of Section 2.2 until the app parity work. Each value gets one label from one of two rules, chosen by the kind of measure:

Kind of measure Rule Informative Use with caution Not informative
Proportions, rates, shares, and any zero estimate Share of the region’s tracts (same indicator and year) whose values fall inside the value’s 90% interval up to 0.40 0.40 to 0.70 over 0.70
Counts, medians, averages, indices Coefficient of variation up to 15% 15% to 30% over 30%

Non-sampled measures (satellite, LODES) are not applicable; values without a published MOE are no MOE. County and region values are judged against the same tract distribution as the tracts.

What the span rule measures. It asks whether the interval is narrow compared with how much tracts in the region differ: whether the estimate can say where the tract stands. It is not a test of absolute precision. That is why it replaces the CV for proportions: the CV divides by the estimate, so a tight interval around a low rate (poverty of 3% ± 2 points) is labelled unreliable although the rate is known to be low. The span rule has the opposite edge case, and it is intended. Where almost every tract is high or almost every tract is near zero (occupied homes, routine checkups, limited-English households), the regional range is compressed, so even an interval of a few points spans most of it and many tracts are not informative. Those estimates are precise in absolute terms but cannot rank or place the tract, and the label says so. Health insurance and high school diplomas looked like this case before the 2026-10-17 MOE fixes (535 and 366 tracts not informative), but mostly because their MOEs were overstated; with the Census direct MOEs the counts are 71 and 26. Users should read not informative as “cannot tell this tract apart from most of the region”, not as “wrong”.

Cutoffs are in config/region.yml (reliability_v2); the label distributions before and after, by indicator, are in docs/reports/2026-10-reliability-v2-calibration.md.

2.11 Density and land area

Density indicators divide a count by 2020 tract land area in square miles (from the Census relationship files, water excluded). Because both estimate and MOE are divided by the same constant, CV and reliability are unchanged.

2.12 Data sources & sustainability

Source Used for Cadence Sustainability
ACS 5-year (Census API / Census Reporter) demographics, economy, education, housing, transportation, insurance, disability annual (rolling) active
CDC PLACES adult health & wellbeing (16 measures surfaced; 40 staged) multi-release series (2020–2025 releases; 2017–2023 data years) active
USGS NLCD (via MRLC Web Coverage Service, WCS) environment (forest, farmland, wetlands, developed, impervious) ~biennial series (2001–2021) active
EOG VIIRS Nighttime Lights (VNL V2) environment (light pollution) annual (2012–) active (account-gated download)

Some historically useful tract sources are static / discontinued and, if added later, must be labeled with their last vintage: Eviction Lab (≤2018), USDA Food Access (2019), HUD LAI (2019), USALEEP life expectancy (2010–2015). (The ACS subject table S1401 jams its public/private split at the tract level, but the detailed table B14002 publishes it cleanly, so Private school enrollment (id 13) is a current, fully reproducible indicator with a margin of error — see its row above.)

3 Part II — Indicator catalog

Definitions below are the canonical reference. Direction “—” means neither higher nor lower is inherently better. All ACS indicators are 2020-tract, harmonized (Section 2.3), with MOE (Section 2.2); CDC indicators are model-based (Section 2.5).

65 indicators across 8 of 10 themes (Safety and Arts & Culture not yet populated).

3.0.1 Character

ID Indicator Source Table/Measure Kind Direction Definition
1 Older adults (65+) ACS B01001 proportion — Share of residents aged 65 and older
2 Youth (under 18) ACS B01001 proportion — Share of residents under age 18
3 Veterans ACS B21001 proportion — Share of civilian adults who are military veterans
4 Asian residents ACS B03002 proportion — Residents who are Asian (non-Hispanic)
5 Black residents ACS B03002 proportion — Residents who are Black or African American (non-Hispanic)
6 Hispanic or Latino residents ACS B03002 proportion — Residents who are Hispanic or Latino
7 White residents ACS B03002 proportion — Residents who are White (non-Hispanic)
8 Other races ACS B03002 proportion — Residents of other or multiple races (non-Hispanic)
21 Median age ACS B01002 median — Median age of residents, in years
34 Limited-English-speaking households ACS C16002 proportion more need Households where no one 14+ speaks English “very well” (corrected 2026-10-17; see note below)
50 Total population ACS B01003 count — Total resident population
51 Population density ACS B01003 ÷ land area density — Residents per square mile of land area (Section 2.11)
53 Households with children ACS B11005 proportion — Households with one or more people under 18
54 Households with seniors ACS B11007 proportion — Households with one or more people 65+
63 Foreign-born residents ACS B05012 proportion — Residents born outside the United States
Warning

Correction (2026-10-17): limited-English-speaking households (id 34). Before this date the numerator summed C16002_003, _006, _009, and _012, which are the totals of households speaking Spanish, other Indo-European, Asian and Pacific Island, and other languages, so the indicator measured households that speak a language other than English (about a quarter of households in many tracts), not limited-English households. The numerator is now the limited-English cells C16002_004, _007, _010, and _013, and the values match the Census Bureau’s published S1602 percentage (typically 1–4% of households). The explorer has shown the corrected values since 2026-10-20.

3.0.2 Economy

ID Indicator Source Table/Measure Kind Direction Definition
9 Employment rate ACS B23025 proportion better Employed ÷ population age 16+ (employment-population ratio)
10 Households with internet access ACS B28002 proportion better Households with a broadband internet subscription
20 Median household income ACS B19013 median better Median annual household income
30 Poverty rate ACS B17001 proportion more need People with income below the federal poverty level
31 Unemployment rate ACS B23025 proportion more need Unemployed ÷ civilian labor force
33 Households receiving SNAP ACS B22010 proportion more need Households that received SNAP in the past year
57 Income inequality (Gini) ACS B19083 index more need Gini index of income inequality (0 = equal, 1 = unequal)
59 Management & professional occupations ACS C24010 proportion — Employed residents in management, business, science & arts occupations (Section 3.1.1)
64 Child poverty (families w/ children) ACS B17010 proportion more need Families with related children <18 below poverty ÷ all families with related children (Section 3.1.3)

3.0.3 Education

ID Indicator Source Table/Measure Kind Direction Definition
11 Bachelor’s degree or higher ACS B15003 proportion better Adults 25+ with a bachelor’s degree or higher
12 High school diploma or higher ACS B15003 proportion better Adults 25+ with at least a high school diploma
58 School enrollment (age 3+) ACS B14001 proportion — Population age 3+ enrolled in school
13 Private school enrollment ACS B14002 proportion — Students enrolled in private school ÷ all students enrolled (private + public, all levels) (Section 3.1.6)

3.0.4 Health

ID Indicator Source Table/Measure Kind Direction Definition
40 No health insurance (adults 18–64) CDC PLACES ACCESS2 model-based more need Adults 18–64 without health insurance
41 Obesity (adults) CDC PLACES OBESITY model-based more need Adults with obesity (BMI ≥ 30)
42 Diabetes (adults) CDC PLACES DIABETES model-based more need Adults ever told they have diabetes
43 Frequent mental distress (adults) CDC PLACES MHLTH model-based more need Adults with 14+ days of poor mental health in the past month
44 Depression (adults) CDC PLACES DEPRESSION model-based more need Adults ever diagnosed with depression
45 High blood pressure (adults) CDC PLACES BPHIGH model-based more need Adults ever told they have high blood pressure
46 Current asthma (adults) CDC PLACES CASTHMA model-based more need Adults who currently have asthma
47 Current smoking (adults) CDC PLACES CSMOKING model-based more need Adults who currently smoke cigarettes
48 No leisure-time physical activity CDC PLACES LPA model-based more need Adults reporting no leisure-time physical activity
49 Fair or poor health (adults) CDC PLACES GHLTH model-based more need Adults reporting fair or poor general health
65 Routine checkup (adults) CDC PLACES CHECKUP model-based better Adults with a routine checkup in the past year
66 Food insecurity (adults) CDC PLACES FOODINSECU model-based more need Adults experiencing food insecurity in the past year
67 Housing insecurity (adults) CDC PLACES HOUSINSECU model-based more need Adults experiencing housing insecurity in the past year
68 Transportation barriers (adults) CDC PLACES LACKTRPT model-based more need Adults lacking reliable transportation in the past year
69 Lack of social & emotional support CDC PLACES EMOTIONSPT model-based more need Adults who lack needed social and emotional support
77 Loneliness (adults) CDC PLACES LONELINESS model-based more need Adults who report feeling lonely
61 Health insurance coverage ACS B27001 proportion better Residents (all ages) with health insurance (Section 3.1.2)
62 Disability (any) ACS B18101 proportion — Civilian noninstitutionalized residents with a disability

3.0.5 Engagement

ID Indicator Source Table/Measure Kind Direction Definition
81 Residential stability ACS B07001 proportion — Residents who lived in the same home one year ago (a proxy for rootedness, not direct civic participation)

3.0.6 Housing

ID Indicator Source Table/Measure Kind Direction Definition
14 Owner-occupied homes ACS B25003 proportion better Occupied homes that are owner-occupied
15 Occupied homes ACS B25002 proportion — Housing units that are occupied
16 Vacant homes ACS B25002 proportion more need Housing units that are vacant
32 Rent-burdened households ACS B25070 proportion more need Renter households paying 30%+ of income on gross rent (of those with a computable rent-to-income ratio)
52 Average household size ACS B25010 average — Average people per occupied household
55 Owner cost burden ACS B25091 proportion more need Owner households paying 30%+ of income on housing costs (of those with a computable cost-to-income ratio)
56 Homes built since 2010 ACS B25034 proportion — Housing units built in 2010 or later (Section 3.1.4)

3.0.7 Transportation

ID Indicator Source Table/Measure Kind Direction Definition
17 Households without a vehicle ACS B08201 proportion more need Households with no vehicle available
18 Commute over 20 minutes ACS B08303 proportion more need Workers with a commute of 20+ minutes
19 Drove alone to work ACS B08301 proportion — Workers who commute by driving alone
60 Mean travel time to work ACS B08013 ÷ B08303 rate more need Average one-way commute time, minutes (Section 3.1.5)

3.0.8 Environment

Raster-derived (NLCD), area-weighted zonal statistics (Section 2.7); no sampling MOE. Land-cover/impervious indicators carry the ~biennial NLCD series (2001, 2004, 2006, 2008, 2011, 2013, 2016, 2019, 2021); tree canopy is annual (2015–2023) from NLCD Tree Canopy Cover.

ID Indicator Source Classes/stat Direction Definition
70 Forest cover NLCD land cover classes 41,42,43 — Share of land that is forest (deciduous/evergreen/mixed)
71 Cropland & farmland NLCD land cover classes 81,82 — Share of land in agriculture (pasture/hay + cultivated crops)
72 Wetlands NLCD land cover classes 90,95 — Share of land that is wetland (woody + emergent herbaceous)
73 Developed land NLCD land cover classes 21–24 more need Share of land developed (open space → high intensity)
74 Impervious surface NLCD impervious mean % more need Average share of surface that is impervious
75 Light pollution EOG VIIRS VNL V2 mean radiance more need Mean nighttime-light radiance (nW/cm²/sr); 2014/2019/2024 (three vintages surfaced from an annual source)
76 Tree canopy NLCD Tree Canopy Cover mean % better Mean continuous tree-canopy cover; annual 2015–2023

3.1 Indicator notes (special cases)

3.1.1 Management & professional occupations (“white-collar”)

Defined as the Census “Management, business, science, and arts” occupational major group (C24010, male + female) ÷ employed civilians 16+. This major group is the standard proxy for professional/managerial (“white-collar”) work in labor-market and social-stratification research — it captures the professional-managerial class and is distinct from sales/office, service, and production/transportation (blue-collar) groups. A broader “white-collar” definition would additionally include sales-and-office occupations; we report the professional-managerial core for interpretability and consistency with the literature.

3.1.2 Health insurance: two complementary measures

  • No health insurance, adults 18–64 (id 40, CDC PLACES): model-based adult uninsured prevalence.
  • Health insurance coverage, all ages (id 61, ACS B27001): direct ACS measure across all ages.

They differ in population (adults vs all ages), method (model-based vs direct), and direction (uninsured vs insured). Use id 61 for an all-ages coverage rate; id 40 for the adult uninsured story.

3.1.3 Child poverty (id 64)

Numerator: families with related children <18 with income below poverty (married-couple + male- and female-householder families). Denominator: all families with related children <18 (the summed six-line denominator across below- and at/above-poverty branches of B17010). This is a family-based child-poverty measure (families with children in poverty), complementary to the overall poverty rate (id 30).

3.1.4 Homes built since 2010 (id 56)

From the B25034 “year structure built” distribution, we report the share built 2010 or later as a growth/new-construction signal, especially informative for this region’s fast-growing exurbs. Alternatives (median year built; share built before 1980 for housing-age/condition) can be added if needed.

B25034 redefines its category boundaries by vintage, so a fixed set of variable indices does not mean the same thing every year. From the 2015 ACS5 onward the two newest categories together span exactly “built 2010 or later,” but in the 2014 ACS5 those same indices are “Built 2010 or later” plus “Built 2000 to 2009”; summing them would sweep in two extra decades and inflate 2014. The indicator therefore uses a per-vintage numerator override (num_vars_by_year in the config): for 2014 the numerator is the single “2010 or later” category; for 2015–2024 it is the two newest categories. This keeps the series measuring the same quantity in every year and is the one indicator where a fixed variable list would silently change meaning across vintages.

3.1.5 Mean travel time to work (id 60)

Aggregate travel time (B08013) ÷ workers who commute (B08303), giving mean one-way minutes. This is a rate (numerator not a subset of the denominator), so the ratio-MOE formula applies. It complements the categorical “commute over 20 minutes” (id 18).

3.1.6 Private school (id 13) — recreated as a current indicator

Originally a legacy orphan (committed app values only, no config, no margin of error) because the ACS subject table S1401 jams its public/private split at the tract level. It has been rebuilt from the detailed table B14002 (private-school enrollees across all levels ÷ all enrolled students), so it now carries the full 2014–2024 series, a margin of error and reliability flags, and an authoritative county estimate like every other indicator.

3.1.7 Median harmonization caveat

For the approximation applied to medians, averages, and indices on pre-2020 split/merged tracts, see Section 2.3 (“Known limitation — medians”).