How We Read The Survey

Key Parameters At A Glance

The thresholds that govern what the site will and won't say. Each is explained in the section that follows; they live in one place in the pipeline so the rules are uniform across every page.

Data release

USGS BBS 2025 · 1966–2024

Survey runs used

Standard protocol only (RPID 101)non-standard runs are excluded

Trend needs

≥ 12 surveyed years · ≥ 3 routesbelow this, no trend is shown

Trend smoothing

3-year mean at each end

Signal threshold

≥ 50% long-term changebaseline index ≥ 0.05

Guild collapse

≥ 40% across ≥ 3 species

Fold-change display

≥ 1,000% shown as “×”

Forecast horizon

5 years

Forecast needs

≥ 18 years of record

Backtest holdout

last 5 years

The Breeding Bird Survey

The North American Breeding Bird Survey (BBS) is a roadside count run every year since 1966 by the USGS Eastern Ecological Science Center (formerly Patuxent) and the Canadian Wildlife Service. It is the longest continental record of bird populations that exists, and it is entirely in the public domain.

Each route is about 24.5 miles long with 50 stops spaced half a mile apart. Once a year, in June, a skilled volunteer drives the route starting half an hour before dawn and, at each stop, records every bird seen or heard in a three-minute window. About 3,000 routes run across the United States in a given year.

We use standard-protocol runs only— internally, run type RPID 101. A run done off-protocol (wrong window, substitute observer arrangement, weather-shortened) is excluded so it can't distort a trend. The survey is designed to be reproducible: same routes, same stops, same window, year after year.

What it samples well — and poorly

• Well sampled: common, conspicuous, daytime-singing landbirds along roads — songbirds, doves, many raptors, open-country species.
• Poorly sampled: nocturnal birds (owls, nightjars), most waterbirds and seabirds, marsh species, and anything that breeds far from a road. Read trends for those groups with extra caution.
• Geographic scope: the United States only in this build — 49 states (Hawaii and the U.S. territories are not part of the BBS at all). The data layer is architected so Canadian provinces can be added later.

The Route-Weighted Index

Raw counts can't be compared across years directly, because a different number of routes runs each year. So for every species, geography, and year we compute a route-weighted index:

index = (total birds counted on routes that ran) ÷ (number of routes that ran that year)

This is the simplest defensible BBS index: the average number of birds a route detects. It is a relative-abundance index, not an absolute population. An index of 1.4 has no meaning on its own — only its change over time, or its size relative to another geography or species, carries signal.

Long-term trend

The headline percent change compares a smoothed baseline to a smoothed recent window — the mean of the first three available years against the mean of the last three. Smoothing both ends keeps a single freak year from defining the trend. The baseline year shown on the site is the last year of that baseline window — the point the change is measured from.

When the baseline index is essentially zero, a percent change explodes into a meaningless number, so we apply a tiny floor (0.001) below which we show no trend at all. For genuine population explosions — an increase of 1,000% or more — a raw percent reads dishonestly, so we switch to a fold-change(e.g. a +1,000% rise is shown as “11×”). Declines are bounded at −100%, so only large increases ever get the “×” treatment.

Recent count

On route-level tables we also show the recent count — the raw number of individuals recorded on that one route in its most recent survey year. That is a single-route tally, not a trend, and not weighted or smoothed.

The Geographies We Compute

The same index is computed at four nested scales, each answering a different question:

• National. Every standard-protocol route in the U.S. The continental picture for a species.
• State. Routes within a state. Useful, but state lines are political, not ecological.
• Bird Conservation Region (BCR). The ecologically meaningful unit — large regions of similar habitat and bird communities, defined by the North American Bird Conservation Initiative. A species' trend often makes more sense by BCR than by state, because birds track habitat, not borders.
• Route. A single survey transect. The finest grain we publish — a sample of one road, valuable for local detail but the noisiest scale.

A trend at any scale is only published when that scale clears the data-quality gate below.

Guilds & Families

Single species wander year to year; ecological guilds move together, and that group signal is where the conservation story actually lives. We assign every species to one of eleven guilds — forest, grassland, wetland, shorebird, waterfowl, seabird, birds of prey, game birds, aerial insectivores, arid-land, and generalists.

How a guild trend is computed Live

A guild's aggregate index for a year is the mean of its member species' indices for that year; the long-term trend is then computed on that mean series exactly like a single species. A guild trend is only shown when at least threemember species are individually eligible, so one well-sampled bird can't stand in for a whole group. This is the same mean-index aggregation the pipeline uses for guild pages and the per-state guild rollups.

Families are the navigation spine

Taxonomic families(warblers, sparrows, hawks, ducks — about 75 of them) are how people actually browse birds, so they organize the site's navigation. Family pages list their member species ranked by each species' own trend; the family browse view shows a convenience mean-index summary. We do not treat a family as an ecological unit the way we do a guild, because a single family can span very different ways of living.

The Signal Engine

A signal (or flag) is a trend notable enough to surface as a card. It is editorial attention, not a regulatory finding or a conservation listing — a prompt to look closer, computed by fixed rules so the same shape always flags the same way. Two signal types ship today; several more are designed and deferred.

Long-arc decline / increase Live

Fires when a species' long-term change at a geography is at least ±50%, the series clears the quality gate (≥ 12 years, ≥ 3 routes), and the baseline index is at least 0.05. That last floor matters: a jump from a near-zero base produces a huge percent that means almost nothing, so we require a non-trivial baseline before a percent change is allowed to flag. Severity is simply the direction — decline or increase.

Guild collapse Live

Fires when a guild's aggregate index is down 40% or more across at least three member species. This is the structural-decline signal — the grassland-bird and aerial-insectivore story — and it only fires on collapse, never on a rise.

Deferred signals Planned

Designed and specified, not yet shipped: recent step (a sharp shift in a short recent window), range shift(the centroid of a species' detections moving ≥ 100 km, typically poleward), and colonization / extirpation (a species newly establishing in, or vanishing from, a geography after years of the opposite). None are technical blockers — they are scope choices for the first release.

To keep cards informative rather than overwhelming, “notable” sections show a balanced mix of declines and increases ranked by magnitude, and an explicit line when there is nothing notable — the absence of a signal is itself meaningful, never a quietly missing section.

Forecasts & Backtest

Where a record is long enough, we project the annual index five years forward with uncertainty bands. A forecast is a projection of the recent trend, not a prediction. Habitat loss, climate, land-use change, and disease are not modeled — if the drivers change, the projection won't.

The model Live

The current build fits an ordinary least-squares linear trend to the annual index and draws prediction bands from the spread of the fit's residuals — an 80% and a 95% band. A forecast is only produced when a series has at least 18 years of record. Production will swap this for Prophet (trend with changepoints, no seasonality) behind the same interface; the bands and backtest reporting stay the same.

The backtest Live

Every forecast carries an honesty check. We hold out the most recent 5 years, fit the model on everything before them, forecast those years, and compare to what actually happened. We report two numbers:

• Backtest error (MAPE) — the typical size of the miss, as a percent. Lower is more trustworthy. Long, smooth series score well; short or volatile ones score poorly, and a large MAPE is a signal to treat the projection loosely.
• 95% coverage — the share of those held-out years that actually landed inside the 95% band. Near 1.0 means the uncertainty bands are honest; well below means the bands are too tight.

We publish both rather than hide them, because a projection without its track record is just a confident line.

Computed vs Cited Trends

There are two legitimate ways to read a BBS trend, and they answer slightly different questions.

• Our computed index — the route-weighted index and smoothed-endpoint change described above. Transparent and reproducible from the published data; every trend on the site is currently labeled computed.
• USGS modeled estimates Planned — USGS also publishes official trend estimates from a hierarchical model, with credible intervals and per-region credibility codes that flag where the data are too thin to trust. Surfacing the cited USGS figure alongside our computed one, with a plain-English reconciliation, is the next ingest.

Until that lands, treat every figure here as our own computed index, not an official USGS trend estimate.

Data-Quality Gates & Suppression

The most important methodological choice is knowing when to say nothing. The survey is uneven — some species×geography combinations have decades of dense coverage, others a handful of sparse years. We suppress rather than publish a number we don't trust:

• No trend below 12 surveyed years or 3 contributing routes at that geography. The cell reads as insufficient data instead of a fragile percent.
• No signal unless the baseline index is at least 0.05 — the floor that stops near-zero baselines from manufacturing dramatic percent changes.
• No forecast below 18 years of record, and no backtest below 13.
• No trend when the baseline is within rounding distance of zero (the 0.001 floor), where a percent change is mathematically meaningless.

These are deliberately stricter at finer geographies, because a single route is a far smaller sample than a whole region.

What To Be Careful About

• It's a trend, not a census. The index tracks direction and relative size, never a head-count of living birds. “Down 40% since 1980” is supported; “there are N birds” is not.
• Roadside bias. Routes follow secondary roads and undersample roadless habitat, wilderness interiors, and wetlands.
• Observer effects. Detection skill varies between volunteers and changes as an observer ages or as a route's observer turns over. The protocol controls for a lot of this, but not all of it.
• Detectability differs by species. A loud, daytime songbird is counted far more reliably than an owl, a marsh bird, or a high-flying raptor. Low-detectability groups carry more noise.
• Sparse series are suppressed, not smoothed over. Where the record is too short or thin, we show no trend rather than a confident-looking guess.

Conservation Overlay

For wildlife-trend data, a conservation overlay is the expected, legitimate framing — the inverse of the stance our crime work takes, because here the relationship between population trend and conservation status is exactly the question readers arrive with.

Today Live, species pages carry a curated status label where one is known, plus a note that ties the continental picture to our own measured trend and the species' guild context. Nothing is asserted unless it is curated — we do not auto-generate a conservation status from the trend. Guild pages carry hardcoded context for the well-documented decliners (grassland birds, aerial insectivores, shorebirds).

Planned Planned, federal sources first: USFWS Birds of Conservation Concern as the federal anchor, then Partners in Flight continental scores and state Species of Greatest Conservation Need lists — each attributed explicitly here, the way we document every source.

What We Deliberately Exclude

• eBird and other non-federal sightings. eBird is superb but carries Cornell licensing and very different sampling properties. We stay on the public-domain federal corpus so every number is reproducible and licensing-clean.
• Absolute population claims. We never convert the index into a count of living birds. That exceeds what a roadside relative-abundance survey can support.
• Real-time or rare-bird alerting. This is an analytics-and-narrative layer over a slow annual record, not a sightings service.
• Individual-route advocacy. A single route is a sample; we surface it for detail, not as a verdict on a specific place.
• Causal attribution. We describe what the trend is, not a definitive why. Habitat, climate, and prey are discussed as context, never asserted as proven cause for a given species.

Sources

USGS North American Breeding Bird Survey, 2025 Release (survey years 1966–2024). A work of the United States Government, in the public domain. We read the ScienceBase release files directly — route metadata (location, BCR, activity), the species list (AOU code → English name, order, family), and the per-state route×year×species counts, standard-protocol runs only.

Source of record: USGS BBS on ScienceBase. We do not use eBird or other non-federal sightings data.

Data Rights & Attribution

The Breeding Bird Survey is a federal public-domain work (17 U.S.C. §105): no licensing fee, royalty, or commercial-use restriction on the underlying observations. Our operating rules are about attribution and civility, not licensing:

• We attribute USGS on the methodology page and in a source line on every data page, with the reporting release named and a link to the canonical record.
• We do not paywall federal data or relabel federal observations as proprietary. What we add is the normalization, signal detection, forecasting, narrative, and presentation.
• We use the bulk release files rather than scraping APIs, to minimize load on federal infrastructure.