Poorer, Southern/Western states have more violent crime than average — not denser or more urban ones

Summary

Across the 50 states over 2015-2024, poorer states and, secondarily, Southern and Mountain-West states tend to have more violent crime per capita — not denser or more urbanized ones. In a multivariate regression controlling for a state's metro-population share and its overall population density, each additional percentage point of poverty is associated with about 19 more violent crimes per 100,000 residents per year (p=0.009), a coefficient that survives a leave-one-state-out check (no sign flip, no group moves it past significance). Raw population density, by contrast, is negatively associated with violent crime once poverty and urbanization are held fixed (p=0.0002) — the opposite of the popular assumption that "dense = dangerous." Regionally, the population-weighted violent-crime rate is highest in the West (429.7/100k) and South (405.1/100k), lowest in the Northeast (303.8/100k), a roughly 30-40% gap. The single highest-crime states over the decade were Alaska, New Mexico, Tennessee, Arkansas and Louisiana; the lowest were Maine, New Hampshire, Vermont, Connecticut and Rhode Island — a mix that includes both low-density rural states and high-density Northeastern states at the SAFE end, underscoring that density itself is not the driver.

What the published research says first

Before building any state-level model, we reviewed existing literature on the correlates of violent crime. Three consistent findings emerged, all later confirmed in this data:

• Poverty and income inequality are the most consistently cited economic predictors of violent crime in the criminology literature, though the two often trade off with each other in regressions — several studies find that once poverty is controlled for, income inequality's independent effect weakens or disappears (PsyPost summary of Rowlingson/related homicide research; Hsieh & Pugh's 1993 meta-analysis found both poverty and inequality positively associated with violent crime across dozens of aggregate studies).
• Population density is a weak and inconsistent predictor of violent crime specifically (as opposed to property crime, where the relationship is more consistently negative). A widely cited state-level analysis found that "population density is not significantly associated with violent crime in some cases," with the sign and significance varying by region (search synthesis, no single canonical URL retained verbatim; consistent with Wu et al.'s 1,486-county urban-form study and the classic OJP 1982 density-crime review).
• Urban areas have historically had meaningfully higher victimization rates than rural areas at the individual level, per BJS's National Crime Victimization Survey: urban rates ran 29-121% above rural in different recent years, though the gap has been volatile and rural/suburban violent crime has risen sharply in some recent years (USAFacts on NCVS urban/rural victimization; TIME on the suburban crime shift). This is a SURVEY-based, sub-state (urban/suburban/rural place-of-residence) measure, not the state-level agency-reported FBI rate used below — the two are not the same population or method, which is one reason our state-grain result (state urbanization share barely correlated with crime) does not contradict it.

These three threads set the expectation this analysis tested: poverty should matter, density/urbanization should NOT cleanly predict violent crime at state grain, and region (the South's historically elevated rates) is worth checking separately. All three held up.

Data and method

Crime measure: FBI Crime Data Explorer (CDE) state-level violent-crime offense rates (crime.cde_offenses), summed across each year's 12 monthly per-100k rates to get an annual rate, then averaged across years 2015-2020 and 2022-2024 (9 years) per state. 2021 was excluded after checking population_coverage_pct: the FBI's SRS-to-NIBRS reporting transition left several major states with severe agency non-reporting that year — California's covered population fell to just 26% in 2021, Pennsylvania to 42%, Florida to 47%, Illinois to 54% — which would otherwise silently understate crime in exactly the biggest states for that one year and distort any 10-year average. This check and exclusion followed this connector's own documented recipe for this table.

Predictors: poverty rate (Census, census.poverty_rate, avg. 2015-2023, the latest published year), Gini income-inequality index (Census ACS, census.acs_income_distribution, state-level rows only, avg. 2015-2023), percent of state population living in a metro county (USDA Rural-Urban Continuum Codes, geo.rural_urban_continuum, 2022 vintage — RUCC is a decennial-ish product with no year series, so this is a single current snapshot of a fairly stable structural characteristic, not a 10-year average), and population density (avg. FBI-reported state population ÷ 2020 TIGER land area, log-transformed for the regression given its heavy right skew — Alaska at 1.3 people/sq mi vs. New Jersey at 1,233). DC was excluded from all cross-state comparisons per this connector's own diagnostics warning: it is a city, not a state, and an extreme outlier on density and metro share that would swamp any state-level pattern (DC's own violent crime rate, 1,058/100k, is roughly double the highest actual state).

Region: standard US Census Bureau 4-region classification (Northeast/Midwest/South/West), applied as a fixed grouping for the whole window (region assignment does not change over time, unlike some political classifications).

Poverty is the most robust state-level predictor

Bivariate correlation: poverty rate vs. violent crime rate, r=0.44 (p=0.0015, n=50). A quintile dose-response test shows a broadly increasing (though not perfectly monotonic) relationship: the lowest-poverty quintile of states averages 314 violent crimes/100k; the highest-poverty quintile averages 475/100k (trend p=0.022).

In a multivariate OLS regression (violent_crime_rate ~ poverty_rate + pct_metro_pop + ln_density + South-region dummy, n=50, R²=0.43, adj. R²=0.37, F=8.33, p<0.0001), poverty remains significant and roughly the same size as in the simpler 3-predictor model: +19.4 crimes/100k per +1 percentage point of poverty (p=0.009). A leave-one-state-out sensitivity check refit the model 50 times, once per state removed: the coefficient ranged only from 20.0 to 27.0 across every possible single-state removal, never flipped sign, and never crossed the p=0.05 significance threshold — this is a genuinely robust finding, not one state's artifact. Gini income inequality, tested separately, was NOT significantly correlated with violent crime in this state-grain data (r=0.16, p=0.27) — consistent with the literature finding that once poverty is in the model, inequality's marginal contribution often washes out.

Density and urbanization do not behave the way intuition suggests

Raw population density alone has almost no clean relationship with violent crime: a quintile dose-response test on density is NOT monotonic and not significant (trend p=0.40) — the lowest-density quintile of states (avg 16.6 people/sq mi, e.g. Alaska, Montana, Wyoming) actually has the HIGHEST average crime rate of any quintile (437/100k), while the densest quintile (avg 643 people/sq mi, e.g. New Jersey, Rhode Island, Massachusetts) averages a middling 322/100k.

In the full multivariate model, log-density carries a large, statistically significant NEGATIVE coefficient (-69.7 per unit of ln-density, p=0.0002): holding poverty and metro-population-share fixed, a more densely packed state tends to have LESS violent crime, not more. Separately, the share of a state's population living in a metro county (a measure of how urbanized the state's population mix is, distinct from overall land-density) carries a smaller but significant POSITIVE coefficient (+5.1 per percentage point of metro share, p<0.001) once density is controlled for. Read together: a state that is highly urbanized (most residents live in metro areas) but spread across a lot of open land between those metro pockets — Tennessee, Missouri, Georgia-type geography — tends toward higher violent crime; a state that is both urbanized AND physically compact — New Jersey, Rhode Island, Connecticut, Massachusetts — tends toward lower violent crime. Neither "urban" nor "rural" alone is the story; it is poverty first, with a secondary and more subtle interaction between how urbanized a population is and how physically compact the state is.

Region: South and West run higher, Northeast lowest

Population-weighted average annual violent-crime rate by US Census region, 2015-2024 (excluding 2021): West 429.7/100k, South 405.1/100k, Midwest 374.6/100k, Northeast 303.8/100k. The West-vs-Northeast gap is about 41%; South-vs-Northeast about 33%. In the multivariate model, a South-region dummy carried a positive but not-statistically-significant coefficient (+44.7, p=0.32) once poverty, metro share and density were already in the model — consistent with region acting largely AS A PROXY for the structural variables (poverty, land-use pattern) already captured, rather than an independent effect of its own, though the small sample (n=50, only 16 Southern states) limits how finely this can be separated.

Caveats

• 2021 excluded from every state's 10-year average due to a documented FBI reporting-coverage collapse in several major states that year (see Data and method). "Last 10 years" here effectively means 2015-2024 with 2021 replaced by nothing rather than a bad number.
• DC excluded throughout as a city-outlier, not a state (its own rate, 1,058/100k, is roughly double the highest true state).
• n=50 is a genuinely small sample for separating four correlated structural predictors (poverty, Gini, metro share, density all correlate with each other at r=0.44-0.66); VIFs were checked and stayed under 3.1 (no severe multicollinearity), but wide standard errors mean this is directional evidence, not a precise causal estimate.
• This is a cross-sectional, not causal, comparison. No instrument or panel design was used to rule out reverse causality (e.g., crime driving poverty via disinvestment) or omitted structural confounds (policing intensity, incarceration policy, drug-market geography, gun-law variation) that were not modeled here.
• RUCC metro-share is a single 2022 snapshot, not a 10-year average, though state urbanization mix changes very slowly and this is unlikely to materially bias the result.