Police-force size tracks crime rates across places, but not within places over 20 years — and county-grain data doesn't exist to test this directly

Summary

The question as asked cannot be answered directly: no dataset — in this connector or, to our knowledge, in any public federal source — tracks police force size at the county level across a 20-year window. The FBI's police-employment series (crime.cde_police_employment) is published only at state grain; county-level agency-by-agency staffing exists only as a scattered, incompletely-reported roll-up (crime.cde_reta / cde_agencies) with no consistent officer-count field. Two things can still be said with real data: (1) at the coarser state-year grain, places with more officers per capita also tend to have more crime — a raw cross-sectional correlation of about r=0.58 (r=0.33 excluding the extreme outlier, Washington DC) — but this is driven mainly by cities/states reactively staffing up where crime is already high, not by police presence causing crime; (2) once state and year fixed effects are removed — i.e., looking only at whether a state's own year-to-year change in officers per 1,000 tracks its own year-to-year change in violent crime — there is no statistically detectable relationship (coefficient -5.0 per officer/1,000, p=0.51, n=712 state-years, 2010-2024). This matches the published quasi-experimental literature: naive raw correlations between police staffing and crime are typically flat-to-positive because of reverse causation, but well-identified studies using exogenous variation (COPS grant timing, terror-alert redeployments) find police reduce crime, with elasticities around -0.3 to -1.3 depending on crime type and study.

Why the question can't be answered at county grain

We ran search_catalog and describe_table across the crime schema and every table with a county dimension. The results: crime.cde_police_employment (officers per 1,000, sworn officer counts) is keyed on state_abbr + year only — no county column. crime.cde_agencies has a county_name field (agency registry, ~19,600 agencies) but carries no officer-count column at all — just agency identity and NIBRS status. crime.cde_reta is agency-level and monthly, with a county derivable through cde_agencies, but it records only offense counts and officers killed/assaulted in the line of duty — never total officer headcount. No table in this warehouse joins an agency's authorized/sworn strength to its jurisdiction's population at county grain. The FBI's own public LEOKA/CDE employment publication is state-level for the same reason: local agencies report officer counts to the FBI's Uniform Crime Reporting Program inconsistently below the state roll-up, and BJS's county-comparable series (the Law Enforcement Management and Administrative Statistics survey) is a periodic sample survey, not an annual full-count series, so it cannot support a continuous 20-year county panel either. This is a genuine, structural data gap, not a query-writing failure — we are naming it rather than substituting a coarser table for the county claim.

What the state-year data show instead (used as the closest available proxy)

Using FBI CDE state-year data 2010-2024 (crime.cde_offenses summed to annual violent-crime rate per 100k, filtered to state-years with at least 80% agency reporting coverage per the recipe for this table's known SRS-to-NIBRS reporting-coverage instability; crime.cde_police_employment for officers per 1,000):

• Every query below applies the predicate p.officers_per_1000 IS NOT NULL. We checked directly what this excludes: SELECT COUNT(*) FROM crime.cde_police_employment WHERE state_abbr <> 'national' AND officers_per_1000 IS NULL returns 0 — all 867 non-national state-year rows have a non-null value, so this predicate is a defensive no-op that excludes zero rows and changes none of the figures below; there is no with/without comparison to show because nothing was removed by it.
• Pooled cross-sectional correlation between a state's average officers-per-1,000 and its average violent-crime rate, 2010-2024: r=0.58 (50 units, 49 states + DC, the 'national' aggregate row excluded via state_abbr <> 'national' because it is not a place and would double count against every state). Excluding DC as well (state_abbr NOT IN ('national','DC')) — DC being an extreme outlier as both the most heavily staffed and most violent jurisdiction — the correlation falls to r=0.33 (49 states). Both figures are reported; this positive correlation is the expected signature of reverse causation, not evidence police presence raises crime.
• Two-way fixed-effects panel (entity=state, time=year, cluster-robust SE on state, n=712 state-year observations, 2010-2024, DC excluded via the same state_abbr NOT IN ('national','DC') predicate for the same outlier reason above — 49 states remain): regressing violent-crime rate on officers-per-1,000 with state and year fixed effects absorbed gives a coefficient of -5.0 (crime-rate points per additional officer per 1,000), std. error 7.6, p=0.51 — not statistically distinguishable from zero.

What the published, quasi-experimental literature finds

The state-year panel above cannot resolve the police-crime relationship causally — officer counts and crime are jointly determined (departments staff up in response to crime, and crime waves can also strain enforcement capacity), and self-reported officer counts are measured with error, both of which bias naive correlations. The literature gets around this with genuinely exogenous shifters of police staffing:

• Chalfin & McCrary, "The Effect of Police on Crime: New Evidence from U.S. Cities, 1960-2010" (NBER w18815, published version Review of Economics and Statistics 2018) — using city panel data and instrumenting for police employment, find police-crime elasticities of about -1.3 for violent crime and -0.8 for property crime; an additional officer is estimated to prevent roughly 1.9 robberies and 0.11 murders per year on average.
• Chalfin & McCrary (2018 companion methods paper) show that OLS estimates of the police-crime relationship are contaminated mainly by measurement error in reported police counts, not simply simultaneity, and correct for it using the Census Annual Survey of Governments as an independent second measure of police staffing.
• Evans & Owens / Mello, "More COPS, Less Crime" — using the timing and size of federal COPS hiring grants (an instrument largely uncorrelated with a city's contemporaneous crime trend) as a natural experiment, finds implied elasticities in a similar range, with substantial cost-effectiveness for violent crime reduction.
• A systematic review and meta-analysis of the broader (mostly non-experimental) literature found the overall macro-level relationship between police force size and crime is negative but small, and often statistically insignificant when studies do not correct for endogeneity — consistent with our own null within-state result above, which used no instrument.

Net picture: when identified with exogenous variation, more police officers per capita causally reduce crime, with an elasticity in roughly the -0.3 to -1.3 range depending on crime type and study design; but a simple correlation of officer staffing against crime rates is confounded by reverse causation and typically comes out flat or positive, exactly what we found in the state-year data here.

Caveats

• Grain: everything computed here is at state-year, not county-year, because no county-grain officer-staffing series exists in this connector or (to our knowledge) in any consistently-reported federal source across a 20-year window. This is named, not glossed over.
• The within-state panel regression above has one predictor and no controls (poverty, unemployment, demographic composition, urbanization) and is not instrumented — it is a description of the raw within-state co-movement, not a causal estimate; the cited literature's instrumented estimates are the causal evidence.
• Every exclusion applied and what it changed, restated together: state_abbr <> 'national' removes the national rollup row (not a place, would double count); state_abbr NOT IN ('national','DC') additionally removes DC as a single-city outlier, moving the pooled correlation from 0.58 to 0.33 — both values reported, and DC is plotted as a separate marker on the chart rather than dropped from the figure; p.officers_per_1000 IS NOT NULL removes zero rows (verified: 0 of 867 rows are null), so it changes nothing.
• FBI CDE reporting coverage varies by state-year due to the SRS-to-NIBRS transition (2018-2021); state-years below 80% population coverage were excluded from the crime-rate calculation to avoid a reporting-coverage artifact.

What This Report Does Not Answer

• County grain: No table in this warehouse (or, per literature search, any consistent federal source) carries police officer headcount at county grain across a multi-year window. crime.cde_police_employment is state-year only; crime.cde_agencies has county but no officer-count column; crime.cde_reta is agency-level but carries no officer headcount. Answered instead at state-year grain, the finest grain the data supports, with the gap stated explicitly.
• Twenty-year window: crime.cde_police_employment's observed coverage is 2010-2026 (about 15 usable years after coverage filtering), not 20; extending further back is not possible with this table. The published causal literature (Chalfin & McCrary) does reach back to 1960 at city grain, cited for the longer-run picture.

Every query behind this report

7 warehouse calls ran in this session, in order. Each is reproducible against the same snapshot.

SELECT state_abbr, MIN(year) miny, MAX(year) maxy, COUNT(*) n FROM crime.cde_police_employment WHERE state_abbr <> 'national' GROUP BY state_abbr ORDER BY state_abbr LIMIT 5

SELECT state_abbr, CAST(year AS INTEGER) yr, SUM(offense_rate) AS violent_rate, AVG(population_coverage_pct) AS cov_pct
FROM crime.cde_offenses
WHERE offense_code = 'violent-crime' AND state_abbr <> 'national' AND CAST(year AS INTEGER) BETWEEN 2010 AND 2024
GROUP BY state_abbr, CAST(year AS INTEGER)
HAVING AVG(population_coverage_pct) >= 80
ORDER BY state_abbr, yr

WITH crime_yr AS (
  SELECT state_abbr, CAST(year AS INTEGER) yr, SUM(offense_rate) AS violent_rate, AVG(population_coverage_pct) AS cov_pct
  FROM crime.cde_offenses
  WHERE offense_code = 'violent-crime' AND state_abbr NOT IN ('national','DC') AND CAST(year AS INTEGER) BETWEEN 2010 AND 2024
  GROUP BY state_abbr, CAST(year AS INTEGER)
  HAVING AVG(population_coverage_pct) >= 80
)
SELECT c.state_abbr, c.yr, c.violent_rate, p.officers_per_1000
FROM crime_yr c
JOIN crime.cde_police_employment p ON p.state_abbr = c.state_abbr AND p.year = c.yr
WHERE p.officers_per_1000 IS NOT NULL

WITH crime_yr AS (
  SELECT state_abbr, CAST(year AS INTEGER) yr, SUM(offense_rate) AS violent_rate, AVG(population_coverage_pct) AS cov_pct
  FROM crime.cde_offenses
  WHERE offense_code = 'violent-crime' AND state_abbr NOT IN ('national','DC') AND CAST(year AS INTEGER) BETWEEN 2010 AND 2024
  GROUP BY state_abbr, CAST(year AS INTEGER)
  HAVING AVG(population_coverage_pct) >= 80
)
SELECT corr(c.violent_rate, p.officers_per_1000) AS pooled_corr, COUNT(*) n
FROM crime_yr c
JOIN crime.cde_police_employment p ON p.state_abbr = c.state_abbr AND p.year = c.yr
WHERE p.officers_per_1000 IS NOT NULL

WITH crime_yr AS (
  SELECT state_abbr, CAST(year AS INTEGER) yr, SUM(offense_rate) AS violent_rate, AVG(population_coverage_pct) AS cov_pct
  FROM crime.cde_offenses
  WHERE offense_code = 'violent-crime' AND state_abbr <> 'national' AND CAST(year AS INTEGER) BETWEEN 2010 AND 2024
  GROUP BY state_abbr, CAST(year AS INTEGER)
  HAVING AVG(population_coverage_pct) >= 80
)
SELECT corr(c.violent_rate, p.officers_per_1000) AS pooled_corr_with_dc, COUNT(*) n
FROM crime_yr c
JOIN crime.cde_police_employment p ON p.state_abbr = c.state_abbr AND p.year = c.yr
WHERE p.officers_per_1000 IS NOT NULL

WITH crime_yr AS (
  SELECT state_abbr, CAST(year AS INTEGER) yr, SUM(offense_rate) AS violent_rate, AVG(population_coverage_pct) AS cov_pct
  FROM crime.cde_offenses
  WHERE offense_code = 'violent-crime' AND state_abbr NOT IN ('national') AND CAST(year AS INTEGER) BETWEEN 2010 AND 2024
  GROUP BY state_abbr, CAST(year AS INTEGER)
  HAVING AVG(population_coverage_pct) >= 80
)
SELECT c.state_abbr, c.officers, c.violent_rate FROM (
  SELECT c.state_abbr, AVG(p.officers_per_1000) officers, AVG(c.violent_rate) violent_rate
  FROM crime_yr c JOIN crime.cde_police_employment p ON p.state_abbr=c.state_abbr AND p.year=c.yr
  WHERE p.officers_per_1000 IS NOT NULL
  GROUP BY c.state_abbr
) c
ORDER BY c.state_abbr

SELECT COUNT(*) FROM crime.cde_police_employment WHERE state_abbr <> 'national' AND officers_per_1000 IS NULL