County-level empirical study

Economic Policy, Poverty, and Crime in U.S. Counties

Does raising the minimum wage reduce crime? Does poverty cause crime—or does crime deepen poverty? This project uses 25 years of county-level U.S. data and multiple causal inference methods to test both directions: whether economic policy changes affect recorded crime rates, and whether poverty and crime reinforce each other.

— Counties

— Years

— Observations

10 Data Sources

Overview

Key Findings

After applying multiple estimation strategies and a systematic credibility framework, no policy lever examined in this study produces a clear, consistent causal effect on recorded crime. The bidirectional poverty–crime analysis finds associations in both directions, but placebo and overlap checks limit causal interpretation.

No Clear Consensus

No single economic policy variable shows a consistent causal effect on crime across all estimation methods. Every lane carries a "caution" verdict from the credibility framework.

Both Directions Tested

The project tests poverty → crime and crime → poverty on equal methodological footing. Associations appear in both directions, but neither passes the full battery of credibility checks.

Method Sensitivity

Results shift meaningfully between conventional two-way fixed effects and double/debiased machine learning, and between national samples and local identification strategies.

Identification Matters

Tighter research designs—border-county comparisons, staggered-adoption estimators, support trimming—tend to weaken or fail to confirm initial broad-sample estimates.

Results

Policy Lanes

Each policy is tested against both violent and property crime rates. Credibility checks are shown as colored indicators: green = pass, amber = warning, red = fail, gray = not available. Hover over a check for details.

Minimum Wage

Primary research lane. Tests whether state and federal minimum wage changes affect county-level recorded crime.

Loading results...

Earned Income Tax Credit

Secondary lane. State EITC rate variation is used for identification. Results are method-sensitive.

Loading results...

Exploratory Lanes

SNAP broad-based categorical eligibility and TANF benefit levels. These lanes have weaker identification and are included for transparency, not causal claims.

Loading results...

Bidirectional: Poverty and Crime

Does poverty cause crime, or does crime deepen poverty? This analysis tests both directions on equal methodological footing: the same panel, the same estimators (county/year FE and DML), and the same overlap diagnostics. Lagged treatment variables are used to separate cause from effect. Associations appear in both directions, but placebo-lead concerns and high covariate imbalance limit causal interpretation. These results are exploratory.

Credibility

Diagnostic Framework

Every policy lane is evaluated against a battery of identification and robustness checks. A lane must pass all checks to qualify as headline-eligible. Currently, no lane meets that bar.

Event-Study Pre-trends Joint F-test on pre-treatment coefficients. A pass (p > 0.05) supports parallel trends.

Covariate Overlap Standardized mean differences between treated and control tails. Max |SMD| < 0.75 is a pass.

Support Trimming DML re-estimated on the subset with good propensity-score overlap. Warns if the estimate shifts materially.

Border-County Design Adjacent cross-state county pairs compared within border pairs. Tests whether local comparisons reproduce national estimates.

Negative-Control Outcomes Policy regressed on outcomes it should not affect (e.g., median age). Significant results suggest residual confounding.

Staggered-Adoption ATT A stacked not-yet-treated event-study estimator. Avoids bias from heterogeneous treatment timing.

Methodology

Methods

The analysis uses a county-year panel spanning 2000–2024. Multiple estimation strategies are applied to each policy lane, with systematic robustness and falsification checks.

Two-Way Fixed Effects (TWFE): County and year fixed effects absorb time-invariant county characteristics and common year shocks. Event-study specifications test for pre-treatment trends.
Double/Debiased Machine Learning (DML): Uses cross-fitted random forests to partial out high-dimensional confounders (demographics, housing, employment, income) before estimating the treatment effect.
Border-County Design: Restricts the sample to adjacent county pairs that straddle a state border, comparing outcomes within pairs over time to control for local unobserved factors.
Staggered-Adoption Estimator: A stacked approach using not-yet-treated counties as controls, avoiding contamination from heterogeneous treatment timing that can bias standard TWFE.
Robustness Checks: Each lane is re-estimated with alternative specifications including population weighting, county-specific detrending, strict coverage filters, placebo leads, and support-trimmed samples.
Falsification / Negative Controls: Policies are regressed on slow-moving demographic outcomes they should not plausibly affect. Significant results flag potential confounding.

Data

Data Sources

All data comes from publicly available U.S. federal sources. Coverage is measured as the share of county-year observations with non-null values.

Source	Variables	Coverage
Loading...

Limitations

Caveats

Crime data reflects recorded crime from the FBI Uniform Crime Reporting program, not actual crime. Reporting practices vary across agencies and over time.
County-level crime coverage is not universal. The FBI UCR county fallback used here aggregates agency-level data, and some counties have incomplete reporting in some years.
No independent external benchmark for county-year crime rates is available in this project. Coverage-sensitivity checks are used instead.
Results are sensitive to the choice of estimator. TWFE and DML can disagree on sign, magnitude, and significance for the same policy lane.
This is an observational study. Even with fixed effects, machine learning, and border designs, residual confounding cannot be ruled out. The negative-control tests reinforce this concern.
ACS demographic controls are only available from 2005 onward, creating a coverage gap in early panel years that may affect DML and overlap diagnostics.