Yes: Denser US Counties Have Consistently Higher Reported Crime Rates Than Rural Ones, With No Sign of the Gap Closing (2019-2023)

Summary

Yes — across roughly 3,000 US counties and 2019-2023, denser counties report meaningfully more crime than the least-dense counties, on both a total-offense and a violent-offense basis, and the gap did not shrink over the window. In 2023 the densest fifth of counties (population density quintile 5) reported about 2,860 Part I offenses per 100,000 residents versus 1,640 per 100k in the least-dense fifth — a 1.7x gap for total crime. A county-year regression of the crime rate on log(population density) finds a coefficient of +290 offenses per 100k per 1-log-unit increase in density (roughly a 2.7x density increase), highly significant (p<0.001, n=10,808 county-years) and stable in sign and magnitude when any single state is dropped from the sample (leave-one-out range 214-309). The pattern is a genuine dose-response across ten density deciles, not just a two-group artifact. This corpus's usable window is shorter than a full ten years (FBI Return A data is only reliably loaded 2018 onward, and 2021 is a documented national reporting-transition dip, not a real crime decline), but within 2019-2023 the density-crime gradient is consistent and does not narrow.

What "crime" and "density" mean here

Crime measure: FBI Uniform Crime Reporting Return A (Summary Reporting System) agency-level monthly counts of the seven Part I offenses (murder, rape, robbery, aggravated assault, burglary, larceny/theft, motor vehicle theft), summed to county-year totals via a name-matched join from agency to county (crime.cde_agencies to geo.counties, ~90% of agencies matched by county name + state) and expressed per 100,000 residents using Census Population Estimates Program county population and density (census.pep_population, 2019 vintage).

Density: county population density (people per square mile), from the same PEP 2019 vintage, used to sort counties into equal-count quintiles/deciles — a standard, if imperfect, proxy for "big dense city" vs. "small town": the densest counties in this sample are large urban cores (e.g., counties containing major cities), the least dense are rural counties.

The gradient is consistent, not a two-bucket artifact

Binning all 10,808 county-year observations (2019, 2020, 2022, 2023 — 2021 excluded, see caveat below) into ten equal-count deciles by density shows a fairly steady dose-response: the total crime rate rises from about 1,640/100k in the lowest decile to roughly 2,000-2,200/100k across the middle deciles to 3,078/100k in the top decile (askengine quantile_binning_test, trend p=0.00027). It is not perfectly monotonic bin-to-bin, but the overall pattern is a rising staircase, not noise.

An OLS regression of the county-year crime rate on log(density) (n=10,808) gives intercept 886 and slope +290 per 100k per log-unit of density (p<0.001, R²=0.029 — density alone explains only a small share of county-to-county variance, which is expected; a great deal of local variation in policing, reporting practice, poverty, and other factors also drives crime rates). A leave-one-state-out sensitivity check (sensitivity_analysis, grouping by state) found the coefficient stayed positive and significant in every one of 49 refits, ranging from 214 to 309, with no sign flip and no state whose removal crossed the p=0.05 threshold — Virginia (state FIPS 51) was the single most influential state (coefficient drops to 214 without it) but the direction and significance of the finding do not depend on any one state.

Has the urban-rural gap widened or narrowed over the decade?

Comparing 2019 to 2023 by density quintile (skipping 2021, see caveat): the total crime rate fell in the four less-dense quintiles (Q1: -15%, Q2: -8%, Q3: -15%, Q4: -8%) but was essentially flat, marginally up, in the densest quintile (+1%). Violent crime rose in every quintile from 2019 to 2023, but by a similar percentage in the densest quintile (+11%) as in the middle quintiles (+6% to +8%) — it did not rise disproportionately faster in dense counties, but it also did not fall there the way total crime did elsewhere. Net effect: the density-crime gap did not close over this window; if anything the persistence of a flat/rising rate in the densest quintile against declines elsewhere kept the absolute gap wide.

2021 data-quality caveat: The FBI's mandatory transition from the legacy Summary Reporting System to NIBRS-only reporting was underway in 2021, and this shows up directly in the raw data: with an essentially stable count of reporting agencies (25,764 in both 2020 and 2021), the nationwide sum of reported offenses fell about 26% in 2021 alone (from ~10.0M to ~7.7M) before recovering in 2022. This is a well-documented national reporting artifact from that transition year (many agencies, including some large ones, reported partial or zero data to the legacy system while converting), not a real one-year crime drop, so 2021 was excluded from all trend comparisons in this report.

How this compares to the published literature

This finding is directionally consistent with existing research. Bureau of Justice Statistics National Crime Victimization Survey data show the 2021 urban violent-victimization rate (24.5 per 1,000) more than double the rural rate (11.1 per 1,000) — a larger urban/rural ratio than this county-level FBI-agency analysis finds, which is expected: NCVS victimization surveys and FBI-agency-reported offenses measure different things (self-reported victimization vs. police-recorded incidents) and use a coarser urban/rural split than a five- or ten-way density split of ~3,000 counties. Commentary on the post-2020 period (e.g., Center for American Progress's review of 2024 crime data) similarly describes cities starting from a higher baseline of violent and property crime and experiencing sharper swings than suburbs/rural areas in the years since 2020, consistent with what the density regression shows here.

Limitations

• Window shorter than 10 years. crime.cde_reta is only reliably loaded from 2018 forward in this corpus (2010-2017 declared but not present), so the analysis covers 2019-2023 (5 years, minus 2021), not a full decade.
• Reported, not actual, crime. All figures are police-recorded Part I offenses. Reporting practices, staffing, and willingness to record vary by agency and can differ systematically between urban and rural departments in ways this analysis cannot separate from true underlying crime differences.
• Agency-to-county matching is name-based (~90% match rate) and 2019 population/density is used as a fixed cross-sectional classifier for the whole window, so it does not capture within-window density changes (e.g., county growth).
• Low R² (0.029) on the bivariate regression means density alone is a weak overall predictor of a county's crime rate — it explains a real and statistically robust average tendency, not the bulk of county-to-county variation, which is driven by many other factors (poverty, policing intensity, demographics, local economic conditions) this analysis did not control for.

What This Report Does Not Answer

• Over the last ten years: crime.cde_reta (FBI Return A) is only reliably loaded 2018 forward in this corpus (2010-2017 declared in the schema but not present in the data); the usable window is 2019-2023 (5 years, with 2021 excluded as a reporting-transition artifact), not a full ten years. This is disclosed explicitly in the Limitations section.