About the data
No single series covers US homebuilding from 1889 to today. This site is a mosaic: the modern Census series where they exist, and hand-transcribed tables from scanned Census publications where they don't. Each breakdown is shown only for the years its source actually covers, on the basis its source actually used — the table below is the honest map. Everywhere on this site the unit of count is the housing unit, never the building or the project: a 200-unit apartment building started contributes 200 to that year's 5+ multifamily figure.
| Series | Years | Source | Notes |
|---|---|---|---|
| Total, private, and public starts | 1889–1958 | Historical Statistics of the United States (1975), series N 156–158, hand-transcribed | Nonfarm housing only; the ownership split begins in 1935. |
| Total, private, and public starts | 1959–1970 | Historical Statistics N 156–158 and Census New Residential Construction | All housing including farm homes; the two sources agree. |
| Publicly owned starts | 1971–1986 | Statistical Abstract 1988, table 1208 (Census series C20), hand-transcribed | The series was discontinued after 1986, when public starts fell to ~2,000 units. |
| Structure type (single-family / 2 / 3+ units) | 1900–1944 | Historical Statistics N 159–161, hand-transcribed | Nonfarm only. |
| Structure type | 1945–1958 | Estimated: old-basis BLS shares (Statistical Abstracts 1951, 1955, 1960) applied to the revised totals | The revised postwar totals were never published by structure type. The unrevised BLS levels run ~30–40% below the revised series, so only their within-year shares are used. These years are flagged as estimates everywhere they appear. |
| Structure type (1 / 2–4 / 5+), US and four regions | 1959–present | Census New Residential Construction (starts_cust.xlsx) | Privately owned starts. Counted in housing units, not buildings — a 200-unit building started is 200 units in the 5+ category. The 2–4 vs. 5+ split begins in 1964; regions carry total and single-family only. |
| Purpose of construction (who built it) | 1974–present | Census Survey of Construction (starts_quarterly_cust.xlsx) | Single-family: built for sale, contractor-built, owner-built, built for rent (derived as residual). Multifamily: for sale vs. for rent. |
| Public starts by program (federal / military / local authority) | 1949–1970 | Historical Statistics N 171–179, hand-transcribed | |
| FHA- and VA-inspected private starts | 1935–1970 | Historical Statistics N 180–181, hand-transcribed | Agency compliance-inspection counts, a proxy for federally backed financing. |
| LIHTC low-income units placed in service | 1987–present | HUD Low-Income Housing Tax Credit database | Counted at placed-in-service (roughly completion), not start. New construction = HUD types 1 and 3; roughly a tenth of records lack the type field and are allocated by each year's known split. The most recent years are omitted for reporting lag. |
| State building permits (units authorized) | 1980–present | Census Building Permits Survey, state annual files | A different measure from starts: units authorized in permit-issuing places (nationally ~8% above starts in 2024). Misses construction where no permit is required — a meaningful undercount in states with large unpermitted unincorporated areas, notably Texas — and the permit-place universe itself expanded in 1994 and 2004, so the series is not on a perfectly constant coverage base. Per-capita rates use FRED's annual state population series over the whole state, permitted or not. |
| Metro building permits | 1990–present | Census Building Permits Survey county files, aggregated to 2023 OMB metro delineations | Constant geography: today's metro definitions applied to the whole series, with county-code changes crosswalked (Dade→Miami-Dade, and others). Connecticut's five metros are excluded — its county geography changed incompatibly in 2022. Permits a state reports without county attribution (material in New Mexico) can't reach any metro, so metro totals lean slightly low. State shares divide by the state-file US total and metro shares by the county-file total; the two differ by up to ~3% in 1991–1993 (the state files are the low ones there) and by under ~1% otherwise. |
| Households (for per-1,000-household rates) | 1940–present | Census table HH-1 (Current Population Survey) | Missing 1941–1946 (shown as gaps). A person-weighted survey count, not the housing-inventory measure — adequate for a rate denominator, but small level breaks exist at survey redesigns (1993, 2001, 2011, 2021, 2025). |
Things to keep in mind
- Nonfarm vs. total. Series before 1959 exclude farm housing; from 1959 they include it (and Alaska and Hawaii). The jump this causes is small but real — about 1.5% in the overlap year.
- The postwar structure-type gap. When the Census revised the 1945–1958 starts series upward (the old BLS estimates had missed roughly a third of all building), it never revised the structure-type detail. Shares from the old series applied to the new totals are an estimate, shaded on the chart and flagged in the tooltip. At 1959 — the one year both bases exist — the old shares overstate the multifamily share by about 0.9 points at an 11% revision gap; the gap reaches ~56% by 1945, so treat the earliest multifamily shares as upper bounds.
- The subsidized middle era is missing. Between the public-housing era and LIHTC sit the privately owned subsidized programs of 1961–1986 — Sections 221(d)(3), 236, and Section 8 new construction — which at their late-1970s peak financed hundreds of thousands of units a year. No assembled annual series is included yet, so the government-role chart understates those years; the FHA and VA lines also stop in 1970 because their source series ends, not the programs.
- From 1987 the spine is private starts only. Publicly owned starts stopped being counted after 1986, when they were about 2,000 units a year; comparisons that cross 1987 are total-vs-private by roughly that margin.
- Starts, placed-in-service, and inspections are different events. The subsidized-housing view deliberately mixes three measures because no single one spans the century; the chart notes say which is which.
- Transcription quality. Every hand-transcribed table passes internal arithmetic checks (ownership, structure, and geography columns each sum to the printed totals). The raw JSON is downloadable: starts.json and funding.json.