How City+ works

Where the numbers come from

City+ uses only sources with a consistent, multi-year, city-level series. Each card names its source and the year it runs through; no observation values are modeled or imputed by City+ (the one computed input, the population denominator behind the crime rates, is documented under recent readings). "Survey" means answers collected from people (with sampling error), "administrative" means agency records, and "modeled" means a statistical estimate published by the source itself.

Metric notes: sources, caveats, and recipes

Each entry below documents one metric: its source, any computation City+ performs, and its caveats. The "about this metric" link on each City+ card opens the corresponding entry. Where City+ computes a rate, the exact formula is shown; formulas combine same-year values from the same survey, and each underlying series can be opened from the card's chart link.

How peer cities are chosen

Comparison against a national average carries little information for local decision-making, so City+ always compares against a named peer set. The comparison frame is chosen per indicator card: every card carries "Compared against" chips (similar-size cities is the default everywhere), a chip switches that card alone, and the card's Compare table lists every member of the active set.

• Similar-size cities (the default): places within 80% to 125% of the city's population, the band widened in 10% steps until at least 8 peers qualify; capped at 30, nearest in population first (population distance, then dcid, so the set is reproducible).
• Regional peers: every same-state city of at least 100,000 people, plus other-state cities in the same census division of at least 100,000 or a quarter of the city's population (whichever is larger), nearest in size first. Five of the 20 seats are reserved for the division so a city-rich home state cannot crowd it out; when fewer than five division cities are near-size, the division's largest top the set up (the card's note says so).
• Aspirational peers: the 10 nearest cities above, between 1.25x and 4x the population; who the city is growing into. When fewer than 3 such cities exist (the country's largest places), the set falls back to every larger city, topped up with the largest nearest in size, and the note discloses the fallback.

Because each peer set contains different cities, the same value may rank differently under each. The active peer set is therefore displayed as part of every claim.

Level and pace of change

Every metric is described on two independent axes, reported separately because neither substitutes for the other:

• Level: where the value stands today. Shown as plain language against the middle of the peer set ("about three-quarters of the peer median") and ranked in the Today's rank column.
• Pace of change: how quickly the value is changing, expressed as a percentile among the peers. "Improving faster than 70% of these peers" means that this city's change over the trend window, measured relative to its own starting value, exceeds that of 70% of the peers in the set. The city is excluded from its own comparison pool, so a peer group of N yields percentiles over the remaining N−1. Below the median, the complementary form is shown ("slower than 85% of these peers" rather than "faster than 15%") so the statement reads in its natural direction. The Change rank column reports every city's percentile on the same basis.

The two axes answer different questions: housing costs may be low today yet rising quickly (favorable level, unfavorable pace), and crime may be low today yet worsening (favorable level, unfavorable pace). Each card and table reports both.

Pace is measured relative to each city's own starting value, not in raw units. A $50 rent increase on a $1,000 rent is faster growth than $60 on a $3,000 rent, and the ranking reflects that.

Both axes sit side by side in every comparison table; the shared layout is described under one table.

The three headline groups: Bright spots, What needs attention, Big shifts

The three groups at the top of every city page are a triage. Each metric is assigned to at most one group by a single significance-aware rule.

Within each group, cards are ordered by the size of the relative percent change (the colored figure under each card's value). Only changes the source can certify compete in that ordering, under the same tests as the tables: survey changes must clear the published margins, and incident-count changes must clear the Poisson chance test. An uncertified change shows in gray and its card sits at the end of the row, present for peer standing, not movement.

• Survey metrics: the claim decides. A real improvement (change beyond the published margin, in the better direction) is a Bright spot; a real worsening or an early down-turn signal goes to What needs attention. When the change is within the margin, level standing breaks the tie: a card clearly ahead of peers is placed in Bright spots, one clearly behind in What needs attention (above the 60th or below the 40th polarity-aligned percentile of today's values, the same thresholds as the standing badge). The chip on a survey triage card IS the card's badge, so the strip and the card can never disagree.
• Administrative metrics (crime, unemployment, housing markets): two doors, either one admits the card. The peer-relative door uses the pace-of-change percentile, the freshest reliable signal (the 12-month percentile where a monthly series exists): well ahead of peers is a Bright spot, well behind is What needs attention. The own-change door admits a card on its own merits: a 12-month change that is statistically certified (the same Poisson chance test as the peer tables), in a definite direction, and at least 5% relative qualifies even when the city sits mid-pack among its peers, because a real improvement does not stop being one when the whole cohort is improving too. A verified improver lands in Bright spots even from a weak level, with a chip stating both truths.

Each headline card carries two quiet chips just above its Learn more link: the metric's own claim (hover for the one-sentence finding behind it) and its standing across all three peer groups (hover for the per-group readings; leading/trailing means above the 60th or below the 40th polarity-aligned percentile, and any trailing group decides the label). The chips stay muted; on a headline card the colored element is the relative change under the value, green or red only when the change is certified. The claim chip speaks in the metric's own direction of movement: a falling crime rate reads "Dropping fast", never "gaining"; a rising income reads "Gaining fast"; a crime rate moving up reads "Rising, the wrong way". The chip's verb follows the trend badge's own direction call, so it never claims a metric is dropping when it is merely worsening slower than its peers. The card's trend badge ("Verified improver", "Worsening") answers a different question (see what the badges mean) and stays on the full indicator card the click lands on.

National leaders and strong performers (the stars)

A star on a card marks standing against the whole country, in two tiers: a gold star (★) names a national leader, at or above the 90th percentile of the city's national size cohort on that indicator in the metric's better direction; a silver star (★) names a strong performer, at or above the 80th. The comparison base is deliberately national rather than the card's peer lens: "2nd of 15 selected peers" is not a national claim, so the stars are computed across every roster city in the same population tier that reports the indicator (the hover names the exact cohort and count), and no star is awarded from a cohort smaller than 30 cities. The largest two tiers are pooled into one large-cities cohort (250k and up) so the biggest cities are honored against a meaningful base. For most metrics the stars reflect the current value, an enduring standing. Crime metrics are the exception: City+ does not rank cities by crime levels (see crime data cautions), so their stars reflect the 12-month change: the fastest-improving slice of the national cohort at the same two thresholds, and never a worsening one. Stars mark standing or momentum against the nation; they are not effectiveness verdicts, and the Verified improver credential remains the earned-trajectory honor. Every star is listed on the star board.

Big shifts is a separate strip for metrics with no declared better direction, such as prices and population. These are never scored as good or bad. A metric appears on a certified rise or fall (beyond the survey margin, for survey series), or when today's value sits at least 25 percentile points from the peer median; the strip orders by relative percent change like the others, gap-only cards last, so a large change is surfaced without being characterized as progress or decline.

How trends are measured: two estimators, chosen by the data

Trends cover each source's most recent 5-year span of published values, and every sentence names the actual years, because "the last 5 years" would imply recency to today that a lagging source cannot deliver. (A longer 10-year view appears where the source reaches back that far.) Which estimator summarizes that window depends on what kind of series the source publishes, and every metric card carries a small "trend:" tag in its footer saying which one is in use:

• Non-overlapping survey periods (ACS metrics). Census ACS numbers are 5-year pooled estimates: adjacent releases share four-fifths of their sample, so year-over-year movement in the chart is largely the same people re-averaged and carries no claims. Change is therefore measured between the newest release and the most recent release sharing zero collection years (for example 2015-2019 vs 2020-2024) and tested against the survey's own published margins of error at its 90% confidence level: beyond the margin is a real change, within the margin is reported as "changed less than the survey can measure". When history is one release short of a clean pair, the two periods share one survey year and every number that uses them carries an asterisk. Dollar metrics are reported in each survey's own published dollars, disclosed as not inflation-adjusted: the commons serves the source's numbers and the disclosure; choosing a deflator is the reader's call. Two qualified shape stories are allowed: the previous non-overlapping span reversing direction (both spans tested), and an early signal when 3 or more consecutive releases each moved the average the same direction and their cumulative movement reaches 1.0x the newest release's margin - consecutive releases swap different years in and out, so a run is real evidence of direction, though never of size. The full year-by-year line stays on every chart as context.
• Robust fit (administrative sources). Crime data (RTCI) and unemployment (BLS) are real single-year (or monthly) administrative counts, where one anomalous year, a reporting glitch, the pandemic crater, genuinely occurs. Drawing a line between two endpoints there would let one bad year fabricate a trend. These metrics use the Theil-Sen estimate: take every pair of years in the window, compute the yearly change between each pair, and use the middle (median) of all those changes. A single spike year barely moves it. Because a fitted slope can disagree with the raw endpoints, two safeguards apply. First, the 2020-distortion flag: when a single year silently reverses the fitted direction, the card says so. Second, turn detection: when the window's path visibly reverses direction partway through (a fall to a bottom followed by a sustained rise, or the mirror image), a single slope would compress that V into one misleading number, sometimes with a sign the chart's endpoints contradict. A reversal counts as visible when the interior extreme sits at least 0.35x the window's value range away from the nearer endpoint (charts scale to that range, so this is the bend a reader can see), clears 2.0x the series' own noise around the two-segment path, and is followed by at least two points. When a turn is detected, the sentence abandons the slope entirely and describes the shape: the endpoint-to-endpoint change with both values, the turn year and its value, and the direction since, with the recent leg judged against the metric's better direction.

Sub-city views follow the same rule as their card. Survey tract maps color by each tract's change between the same two survey periods as the card, judged against the tract's own published margins; tracts whose change is within the margin keep their color (spatial pattern survives noise better than any single tract) but are starred and say so on hover. Demographic subgroups use the same two periods; their margins are wide and unpublished at that grain, so their ordering is soft and the panel says so. Administrative sub-city views keep the robust fit.

If a source contains a known break (a methodology change, a missing survey year), the window is truncated so that it never spans the break: a shorter window that respects the break is preferred to a longer one fitted across it.

The three checks behind every improvement claim

City+ never labels a city an improver solely because a number moved; a single-period change can reflect chance, noise, or a reporting artifact. Before a trend is described as "sustained improvement, worth investigating", it must pass all three of the following checks:

• Bigger than noise. Survey metrics: the change between the two survey periods must clear the combined published margin of error at the survey's own 90% confidence level. Administrative metrics publish no margins, so the error bar is estimated from the series itself; this is the noise proxy named in the sources table. Concretely: take the window's year-over-year changes, subtract the fitted slope from each so the trend is not counted as noise, and estimate the per-year noise scale as 1.4826 × the median absolute deviation of those detrended steps (the MAD-to-sigma factor; robust, so one outlier year cannot inflate it). Multiplied by √n for the n yearly steps in the window, that is the floor the fitted change must exceed by 1.5×.
• Sustained, not a one-year pop (administrative metrics only). At least 70% of the meaningful year-to-year moves must point the same direction as the overall trend, counting only moves at least as large as the per-year noise floor above (sub-noise years neither help nor hurt). Survey periods are pooled averages, so this check has no meaning there; the margin test carries the load.
• Big enough to matter. Statistically real but tiny is still tiny. The change must be at least 25% of the interquartile range of the peers' current values (computed when at least 4 peers report), or failing that at least 5% of the city's own starting value.

A trend that moves the right way but fails a check is labeled honestly on the card and treated as a raw signal, not a finding. Every constant here is a named tunable in the thresholds table.

What the badges mean

The badge on each card answers one question: can this city's own trend be trusted? It judges the trend itself, not the pace percentile; a city can be changing faster than 94% of its peers while its own trend still fails the checks. Survey and administrative metrics carry different badge vocabularies because different tests back them.

• Survey metrics, read straight off the published margins: Real improvement (beyond the margin and big enough to matter), Improving (small) (beyond the margin but small next to the peer spread), Worsening (beyond the margin, the wrong way), No measurable change (within the margin), and Turning up / down (early signal) (within the margin, but the last 3 releases each moved the average the same way: a reliable direction signal whose size is not yet measurable). Estimates whose implied standard error (margin ÷ 1.645) exceeds 0.3 of the value at either endpoint carry a low-reliability caveat.
• Administrative metrics, judged by the three checks: Verified improver passed all three; Improving, unverified is moving the right way with at least one check failed; Worsening is moving against the better direction. These badges carry a suffix when the peer comparison changes the story ("leading peers", "slower than peers").
• Rising / Falling: a metric with no declared better direction; the badge only reports movement. Level only: a modeled estimate with no usable trend; no movement claims at all.

Recent readings vs trends

Crime rates and housing indexes go further: because each trailing-12-month or index value is already a full annual reading, their yearly series anchor to the freshest month still covered by at least 90% of reporting cities (the same month for everyone, so peers stay aligned), rather than to December. Their trends therefore run to within a few months of today, and each card's source line names the exact month it runs through.

How the crime rates are computed. Each month's rate is the sum of incident counts over the 12 months ending at that month, divided by the place's population, per 100,000 residents. The source publishes counts; the per-capita division is performed by DataWorks Commons using the Census Bureau's annual population estimates, interpolated monthly between adjacent estimates (which are July 1 figures) so the denominator changes smoothly rather than stepping at each new year (the standard convention for monthly demographic rates). A growing city's rate therefore reflects both its incident trend and its population growth. For the most recent months, where the Census has not yet published an estimate, the newest available estimate is carried forward, and those months' rates are restated when the next estimate publishes (the revision is typically within one to two percent). The handful of reporting agencies that do not correspond to a Census place use the population figure the source publishes for the agency. The violent and property crime totals are computed from their component offenses under the FBI's definitions (violent: homicide, rape, robbery, aggravated assault; property: burglary, larceny-theft, motor vehicle theft), and a month's total is published only when every component is reported. Cities whose agencies report some but not all components carry no total for that month; their individual offense series still appear. This is why a few large cities show component crime series but no violent or property crime rate.

Why totals are recomputed rather than read from the source. The source file carries its own violent and property crime columns, but those columns hold the sum of whichever components each agency reports: for an agency that reports only homicides, the published "violent crime" figure is its homicide count, roughly one percent of a true violent crime total. This was verified against every agency-month in the source: wherever components are missing, the published total equals the sum of the components present. A partial sum labeled as a total would understate crime most for the cities reporting least, so the totals shown here are always rebuilt from components under the all-components rule above.

Distinguishing a true zero from non-reporting. The source writes a literal 0, not a blank, for some offenses an agency does not report at all, and 0 is also a legitimate value: a small city can genuinely record no homicides in a year. The two are separated statistically, per offense, with no city-specific exceptions. For each offense, the lowest-reporting genuine agencies set a floor: when even the 1st percentile of reporting agencies averages well above zero for that offense (at least 10% of the median agency's level), an identically-zero series is treated as non-reporting and withheld, because no real city plausibly sits at zero. Property offenses clear that floor, so an all-zero property series is read as non-reporting; homicide does not, so cities with genuine zero-homicide years keep their series. Separately, a series that runs 12 or more consecutive zeros after a nonzero history is truncated at the point reporting stopped, the signature of an agency leaving the sample rather than of crime ceasing. These screens are inferences from the shape of the data, not certainties, and the thresholds involved (the 10% floor, the 12-month run) are documented here precisely because they are judgment calls.

Count toplines. Where a metric declares a count topline (currently homicides), the card's headline number, sparkline, and most-recent-reading line carry the trailing-12-month count itself, the figure officials and residents actually use, with the per-100,000 rate stated directly beside it. Peer tables and every cross-city comparison use the rate, because counts are not comparable across cities of different sizes. The count and the rate are the same underlying incidents; only the denominator differs.

Crime data cautions

The crime series carry limitations their sources are explicit about, and readers of the peer comparisons should hold them in mind:

• Counts are preliminary and unofficial. Agencies revise their figures, sometimes months later; the Real-Time Crime Index trades finality for freshness, and revisions flow through these pages as the source publishes them.
• Reporting methodology varies by agency. Offense classification and reporting practices differ from department to department, so a cross-city difference of a few percent may reflect practice rather than crime. Peer comparisons here should be read as indicative, not definitive.
• Rankings are not effectiveness measures. The FBI expressly discourages using crime rankings to judge law-enforcement effectiveness, because crime is shaped by many factors outside any agency's control. City+ therefore does not rank cities by crime levels at all: for these metrics the peer tables carry no current-level rank and no latest-versus-peers column, and cards carry no level-standing label. Each peer's actual rate remains visible in the table, but the only rankings presented are rankings of change, each city measured against its own prior twelve months. Peer values are context; they say nothing about why cities differ, and nothing about any police department's performance.
• Participation changes over time. Agencies join, leave, and lapse in reporting, so some multi-year comparisons span differing levels of participation. The "U.S. (RTCI sample)" reference row aggregates only agencies with complete recent reporting and is the sample's benchmark, not a census of the country.

For these metrics the card leads with the most recent 12 months, the horizon most relevant to operational management, and the badge follows that recent direction ("Improving recently", "Worsening recently"). The multi-year window stays on the card as the labeled context line, and "Verified improver" is still earned only by the multi-year checks plus a recent direction that agrees; twelve good months alone never qualify. Because the underlying values are already 12-month sums, this recent reading is a real trend, not seasonal noise.

Peer tables for these metrics show the short horizon only: each city's latest reading, its 12-month change with a significance dot where the data supports one (see one table), and that change's rank. Multi-year columns are omitted because these metrics are managed on a monthly horizon; the multi-year story stays in the card text and the full History sparkline.

For monthly sources (unemployment, crime, housing), cards show the most recent monthly reading next to the same month a year earlier, and annual-survey peer tables carry each city's latest monthly reading and its 12-month change where a monthly companion series exists. City-level monthly data is not seasonally adjusted, so a 5-month "trend" would mostly measure summer against winter; the year-over-year pairing is the honest short-horizon comparison, and the multi-year trend remains the basis for every improvement claim.

Known shocks are registered per metric and trends never span them. Unemployment is the clearest case: its window starts after the 2020-2021 pandemic spike and rebound, so recovery from the shock is never scored as improvement.

"Peers worth a call"

For each metric with a declared better direction, City+ surveys all cities in the same size group, retains only those whose improvement passed all three checks, and groups the survivors by starting position: bottom, middle, or top third of the peer group.

The grouping guards against mean reversion: the places that "improve" most are often simply the places that started worst, drifting back toward typical values. A city therefore sees only referral candidates that started in a comparable position, and every candidate's starting position is disclosed, so a city that improved from a poor baseline is never mistaken for one that improved from a strong one.

For metrics with a monthly pulse (the crime rates, unemployment), passing the multi-year checks is not enough on its own: a candidate whose own most recent 12 months moved against the better direction is dropped, because a recommendation based on an improvement that has since reversed would be stale. The rows for these metrics lead with each candidate's last-12-months movement and are ordered by it; the multi-year rate stays on the row as the sustained credential.

Referral candidates are places where something genuinely changed, worth a phone call to find out what. They are never presented as evidence that a policy worked: establishing causation requires evaluation-grade study designs and is outside the scope of these methods.

Claims City+ does not make

• No improver claims on neutral metrics. Rising home values benefit owners and burden buyers, so housing prices carry no "better" direction and no referral list.
• No trend claims on modeled estimates. Where the source is a slow-moving statistical model, only the level is shown, and the card says so.
• No mixed vintages. Every metric uses a single source series; datasets are never spliced to construct a longer history.
• No causation. "This peer improved" never becomes "this peer's policy worked".

Where is this changing: the tract map

On every city page, metrics with census-tract data carry a "Where is this changing?" panel: the city's tracts drawn over a street basemap, each colored by its change over the same two survey periods as the card above it, relative to the tract's own starting level. Green means the tract moved in the metric's better direction and red the other way (blue and orange for metrics with no better direction); deeper color means a larger move. Hovering a tract shows its numbers, its margin verdict, and its full history; clicking it opens that tract in the Explorer. A companion table beside the map ranks tracts on the same skeleton as the peer tables (Change rank, Latest, Change, Sig.), listing the twenty improving fastest and the twenty moving the wrong way fastest. Hovering a table row highlights that tract on the map, and hovering a tract highlights its row. Tract boundaries are Census TIGER/Line 2024, simplified for display.

One state needs a stitch: Connecticut renumbered every census tract in 2022 when it replaced counties with planning regions, without changing any boundary. The source data keys the years before the change under the old tract id and the years after under the new one, so for Connecticut cities each tract's two halves are rejoined into one series (the id pairing is validated geometrically and published in the artifact as tract_id_aliases.tsv). No values are altered or interpolated.

Rules governing the map:

• Every tract is judged against its own published margin. Tract estimates are noisier than citywide ones; a tract whose change is within its margin keeps its color (the spatial pattern survives noise better than any single tract) but is starred and says so on hover.
• Tract membership follows the majority rule. A tract belongs to a city when at least half of it lies inside, by Census Block Assignment File shares; tract boundaries and city limits never match exactly.
• Newer tracts still get a verdict. Census tract boundaries changed in 2020, so some tracts lack the window's first years. Those with at least 3 in-window years spanning at least 2 calendar years are colored by the change over the years they do have, drawn with a dashed outline and labeled with their shorter window; tracts with too little data stay gray. Nothing is imputed.
• Small tracts are flagged. Tract-level survey estimates are noisy; the hover card marks tracts whose population base falls below the tract_small_denom threshold.
• Computed metrics use the same recipes at tract level (each tract's own rate from its own components, same-year values), and a citywide median is simply shown as each tract's own median change.

Not yet at tract level: unemployment (BLS publishes no tract series), the crime rates (the Real-Time Crime Index is agency-level; Baltimore's own police data does reach tract level in this Commons and is the planned next step), and the Zillow housing metrics. The obesity card links to a tract map of levels but makes no change claims (modeled estimates).

State and national reference values

Three surfaces carry quiet reference values: the peer tables pin a row for the city's state and one for the United States (marked "ref"), the tract panel states both latest levels under its map, the tract popup's "by the numbers" table adds state and US columns, and each demographic group in a breakdown panel shows the same group's statewide and national value in small type. All of them are context, not competitors: a reference never takes a rank, never joins a percentile pool, never sorts with the peer rows, and never changes a verdict, badge, or sentence. They exist so a number always has a yardstick, a city can lead all its peers and still trail its own state, and both facts should be visible at once.

Where they come from: for survey and BLS metrics, the same variable at state and national level from the same source family; for Zillow housing metrics, Zillow's own state and national series; and their trend cells use the identical estimator and display grammar as the peer rows. Two honest exceptions. Crime: the Real-Time Crime Index publishes no state series, so crime metrics carry no state reference, and their "U.S. (RTCI sample)" row is RTCI's own full-sample aggregate, the average across its national reporting sample of agencies, which is the source's national benchmark but not a census of the country, and skews toward the kinds of agencies that report. Rent: the rent index is published nationally but not by state, so rent metrics carry a US reference only. A reference can also lag the cities by one release; its cell always shows its own month.

One table: level and change together

Every comparison table, peer cities, tracts, and demographic groups, uses the same skeleton so a reader never has to reorient: Latest (each place's newest value, its period named), Today's rank (current values ordered best-first in the metric's better direction), Change (each place's change over its own window: two non-overlapping survey periods, or the last 12 months for monthly sources), Sig. (a dot, below), Change rank (every measured change ranked, largest improvement first), and History. Tables open sorted by Change rank. A star on a change means that place's two survey periods share one year; direction is trustworthy, size approximate.

The Sig. dot is the significance verdict. Filled green: a real improvement. Filled red: a real worsening. Hollow: within the margin, so ranks among hollow-dot rows are ordering values inside their margins. What the dot tests depends on what the source publishes: survey metrics test against published margins of error (90% confidence); the crime rates test the 12-month change in implied incident counts (rate × population / 100k) against chance fluctuation: z = (n₂ - n₁) / √(n₁ + n₂), filled when |z| > 1.645 (the same 90% level). Unemployment and the housing indexes carry no dot: their sources publish no error model, and constructing one would introduce assumptions the sources do not support.

The tract tables and group breakdowns use the same columns (see the tract view and breakdowns).

Break down by group: race, age, sex

Metrics drawn from the Census American Community Survey carry a "Break down by group" panel: the same indicator computed separately for the demographic groups the source table publishes, each with its current value, a proportional bar, an estimate of the people behind the number, its full history, and its change between the same two survey periods as the card above it. A gap line names the current highest and lowest groups, the size of the difference between them, and whether that pair's difference narrowed, widened, or held roughly steady between the two periods; the pair is held fixed so the sentence never mixes different groups across years.

Today's rank and Change rank place each group against the same group across the peer set shown on the page (Black poverty here against Black poverty in the peer cities, never against a different group), colored when the city sits in the best or worst third; a peer whose group is too small to report reliably sits out. Est. people (or households) is the count behind the rate: the survey's numerator (people below poverty, uninsured people, owner households) rounded to the nearest 1,000, 100, or 10 by size, or for median income the number of households in the group. Within the list, the group faring worst carries a red bar and the best a green one.

Rules governing the breakdowns:

• Every group's rate is self-contained. A group's numerator and denominator always come from the same source table (Black poverty = people below poverty among Black residents for whom poverty status is determined), never from another table's population count, so category quirks in one table can never contaminate a rate.
• Every variable was verified before shipping. The statvar list behind each group is generated and checked against the source at authoring time over a probe roster of cities across all four size tiers (published in the artifact as breakdowns.yaml with its coverage stats). Groups the source does not publish at city level simply do not appear; that is why unemployment (BLS, no city-level demographic series) and education-by-race have no panel or fewer dimensions.
• Small groups are suppressed, by name. A group whose universe in this city falls below breakdown_small_denom (500 people) is not drawn or used in the gap, and the panel says which groups were withheld and why. Margins at this grain are wide and unpublished, so no significance dot is shown and the Change-rank ordering is soft.
• Race categories overlap and are labeled so. The race view shows White (non-Hispanic) and Hispanic (any race) alongside the race-alone groups, which can include Hispanic members; it is a lens, not a partition that sums to the city total.
• Householder semantics are named. Median household income and homeownership classify a household by the race of its householder, and the panel says so.
• ACS five-year estimates are smoothed. Adjacent years share sample, so group trends move more slowly than reality and the 2020 collection problems affect these series like every other ACS series on the page.

The "What's behind this?" analysis

Each indicator card carries a "What's behind this?" button that opens a guided analysis of that card. The question itself is generated deterministically from the card's certified state (a certified trend, a certified reversal, or an early direction signal), and the interface accepts no free-form input.

The analysis is composed from evidence computed server-side before any language model is involved: the card's own trend verdict, a decomposition across census tracts and demographic groups where those data exist, the pattern across all three peer sets, and, where the metric registry declares them, hypothesis checks evaluated against held data over the card's own window (for example, income-distribution measures behind the inequality card). All figures, changes, and percentages are precomputed; the language model narrates the computed evidence and may name general, well-documented mechanisms from urban and economic research as possibilities, but it is prevented from generating specific citations, statistics, or study references, and its output is checked so that any figure not present in the retrieved evidence is removed. Each evidence block links to its underlying series, names its source, and states the window it covers. Nothing is retrieved from the open web, and causal claims are refused for the same reason the referral list refuses them. These analyses use the site's standard question allowance.

The exact thresholds in force

All of the above runs on named, published thresholds. Changing any of them is a visible, versioned change (current build: 2026-08-17.1), and the machine-readable registry behind this page is public.