National Crime Victimization Survey (NCVS)

It is the nation's primary measure of criminal victimization, and the longest-running one: a federal statistical program that has interviewed American households about crime continuously since 1973. Administered by the U.S. Census Bureau for the Bureau of Justice Statistics, it puts roughly 240,000 people age 12 or older in about 150,000 households on the record each collection year. Households join a rotation panel for three and a half years - seven interviews - so the same address reports its own burglaries and assaults repeatedly, which is what makes year-over-year household follow-through possible at all.

The instrument covers nonfatal personal crime (rape/sexual assault, robbery, aggravated and simple assault, personal larceny) and household property crime (burglary/trespassing, motor vehicle theft, other household theft), plus the context police counts never capture: whether the incident was reported, offender characteristics, weapons, injuries, losses, and the protective measures households took. Within Datadory's catalog of 1,744 datasets this record scores 8/10 against a mean of 7.81. Get a sample of this dataset and judge the household-security columns against your own market model.

One spine, two sides - household crime and personal crime:

# one household-crime interview record, shaped exactly as delivered
record_level       : household incident
panel_key          : same address, tracked across 7 interview waves
crime_type         : burglary / trespassing
entry_outcome      : completed entry vs attempted entry, as classified
reported_to_police : no
loss               : household property, valued at interview
protective_measures: recorded as the household answered

# the same spine switched to a personal crime
record_level       : personal incident
crime_type         : simple assault
victim_age_sex     : from the person demographic block
offender           : number, perceived age/sex, victim-offender relationship
weapon             : presence and type as described by the victim
injury             : treatment sought, as reported

# ...repeats across ~240,000 persons and ~150,000 households per collection year,
#     stacked into a series reaching back to 1973

Read the anatomy rather than the magnitudes. Every line is something the questionnaire asks: the panel key that keeps one address identifiable across seven interview waves, the crime-type classification, the outcome of the entry attempt, the reported_to_police flag that makes the unreported share computable, and the protective-measures answer that turns a burglary row into security-adoption evidence. Multiply the shape across roughly 240,000 persons and 150,000 households a collection year, stacked back to 1973, and a household-risk study stops being a survey-design problem and becomes a filter.

Six jobs, one row: panel identity, incident classification, the police-reporting flag, victim and household characteristics, offender and incident detail, and the weights. The table below carries the columns every analysis touches, defined at the level the instrument documents them.

One honest caveat, stated up front: variable names and codes move slightly between instrument years, and this record's dictionary carries them at inferred confidence until pinned. The exact column names for your target year sit folded under additional fields on request and get confirmed against that year's own documentation when your sample is cut - before anything ships, not after.

Geography - United States throughout, designed to produce national estimates. Subnational work is possible but disciplined: the publisher maintains a dedicated state-level analysis methodology, and serious state cuts pool years under it rather than pretending every row carries a state code.

Temporal - the collection span runs 1973 through 2024. Trend estimates are typically computed from 1993-forward files, the era of the modern instrument design. And 2024 deserves its own sentence: a split-sample year pairing the legacy instrument with a redesigned one, whose estimates publish separately.

Granularity - person-level and household-level interviews with incident-level records underneath: roughly 240,000 persons, about 150,000 households, and about 109,000 personal interviews in the 2024 collection year alone. Any national total you compute reconciles against the published estimates once the weights are applied.

API, files, or your warehouse. Daily, weekly, or hourly.

Pick the channel your stack already speaks and set the cadence to match the decision being fed - flat-file loads for teams anchoring a research bench on the full 1973-forward history, direct pipes into Snowflake, BigQuery or Redshift for analysts running household-risk models in SQL, scoped extracts for products surfacing burglary-rate signal inside a dashboard. Cadence changes are a settings conversation, not a re-integration project.

Normalization happens before anything reaches you. Instrument-year files arrive merged into one consistent schema with types resolved against the publisher's own documentation, joined to the field reference above unchanged. Every shipment includes sample rows for validation and the coverage profile mapped to the years, crime types and cuts you named.

Five jobs it does better than any police-count substitute.

Size the residential security market from what actually happened. Household burglary and trespassing rates by year, counted at the door rather than at the precinct, are the demand baseline for alarm-market models - complete with the break-ins that generated no claim and no report.

Measure the protection gap. Protective-measures responses sit beside victimization rates in the same instrument, so penetration of locks, alarms and defensive behaviour reads against the very crime it was meant to deter.

Quantify the dark figure. The reported-to-police flag makes the unreported share a group-by, per crime type, per year.

Benchmark official counts. Set victimization rates beside agency counts and separate real crime movement from reporting behaviour - the reconciliation every serious crime-data argument eventually needs.

Train models on labelled household risk. Decades of incident records with published weights and a built-in reporting label give classifiers ground truth private panels cannot approach.

Market Researchers & Consultants get the victimization-plus-protection baseline for residential security market studies. Data Scientists & ML Engineers get decades of labelled incident records - see data scientists use cases. Competitive Intel & Product Teams get household protective behaviour as an adoption signal - see competitive intel product teams use cases. Developers & Builders get a stable documented spine that joins onto household reference data - see developers builders use cases. Journalists, Academics & Students get the federal reference everyone else charts from, with its weights and caveats intact. The common thread is the vantage point: every police-count series starts after a report is filed; this one starts at the household, report or no report.

Within security alarm services data, this record owns the household's own account - victimization including unreported crime, plus protective behaviour. The neighbours own different jobs: the FBI Crime Data Explorer (CDE) carries agency-reported offence and arrest counts - official, but blind to anything unreported; data.police.uk maps street-level police-recorded incidents for England and Wales; LA Crime Data (2020 to Present) goes incident-deep for one metro - see LA Crime Data vs NCVS.

On the catalog's rubric this record scores 8/10 against a 7.81 average across 1,744 datasets, held back mainly by grain: it is a national survey, not a nationwide incident ledger, so street-level work belongs to its neighbours while rate-and-trend work belongs here.

Four things, stated plainly.

First, variable names are pinned at sampling. The dictionary above is written at the level the instrument documents and carries inferred confidence on exact codes; your target year's variable names get confirmed against that year's documentation before anything ships.

Second, 2024 is a split-sample year. The legacy instrument and a redesigned one ran in parallel, and their estimates publish separately - name which basis your analysis wants and the sample arrives on it.

Third, trend work starts in 1993. Files reach back to 1973, but comparable modern trends are typically computed from 1993-forward material; pre-1993 files differ in design and we say so rather than splice quietly.

Fourth, geography has rules. This is a national-design survey; state-level analysis follows the publisher's own methodology and usually pools years. Your sample states which years were pooled and how.