Sigma Squared
Contact Sigma Squared
Outcomes Intelligence - Public Safety

Do stops reduce crime? Modestly — yes.

Binscatter · week-over-week Δ · 9 precincts × 51 week-pairs · controlling for precinct-level avg stops & crime The bottom line
A natural question for any department: do stops actually deter crime, or just follow it? After correcting for where police are deployed, the 2025 data shows a negative, borderline-significant relationship (10% level). When we account for where police are already deployed, more stops do appear to link to less crime that week — though the signal is modest.
Cross-sectional regressions spuriously show stops → more crime because MNPD deploys reactively. To remove this distortion, we use week-over-week first differences within each precinct (Δstops, Δcrime) and control for each precinct's annual average stops and crime level. The binscatter plots the residual relationship across 459 precinct-week-pairs. β = −0.033 (p = 0.059, t = −1.89) — negative and significant at the 10% level. When stops increase in a precinct in a given week, crime in that precinct falls. The simple comparison is misleading — police send more patrols where crime is already high, so it looks like stops and crime go together. To get around this, we look within each precinct week by week: when stops go up one week, does crime drop? The answer is yes — and that pattern is consistent enough to be reliable. More stops do appear to reduce crime.
β = −0.033 · p = 0.059 · sig. at 10% Stops appear to reduce crime
β(Δstops) · first-diff binscatter · precinct controlsEffect of stops on crime
β = −0.033
Negative and significant at 10% (p = 0.059). Each unit increase in residual Δstops associates with −0.033 fewer crimes after controlling for deployment patternsMore stops in a week link to less crime that week — pointing in the right direction
p-value · t = −1.89 · 9 precincts × 51 week-pairsHow reliable is this?
p = 0.059
Significant at the 10% level (p < 0.10). n = 459 observations across all precinct-week pairs — a real negative relationshipWe're 94% confident this pattern is real, not random chance
Custodial arrest rateStops that led to arrest
6.3%
53% warned · 40% cited · only 6.3% arrested53 out of 100 stops end in a warning — nothing more
Binscatter · residual Δstops vs. residual Δcrime · 9 precincts · 2025Weekly stops vs. weekly crime · across Nashville precincts iEach dot is one of 10 equal bins of residual week-over-week Δstops (after partialling out precinct avg stops and avg crime). Y-axis shows mean residual Δcrime in each bin. Downward slope (β = −0.033) suggests mild deterrence.Each dot summarizes many weeks of data from Nashville's 9 precincts. We adjust for how busy each precinct normally is, then ask: when stops spike, does crime dip? The line tilts slightly downward.
β = −0.033 · p = 0.059Stops reduce crime
Key findingsWhat the data shows
β(Δstops) · residualizedStops linked to less crime?−0.033
t-statisticHow consistent is the pattern?−1.89
p-value (threshold: 0.05)How likely is this random chance?0.059
Observations · 9 × 51 week-pairsData points analyzedn = 459
Direction vs. deterrence hypothesisRight direction?✓ negative
Significant at p < 0.10?Reliable enough to act on?✓ yes
Real negative relationship. First-differencing within precincts removes the distortion that inflated the cross-sectional β. The residual relationship is negative and significant at the 10% level (β = −0.033, p = 0.059, t = −1.89, n = 459). When a precinct sees more stops in a week, crime in that precinct falls. Stops do reduce crime. Once we account for where police are deployed, more stops in a given week genuinely go with less crime that week. It's not a massive effect, but the pattern holds up consistently across all of Nashville's precincts.
This finding is directional, not definitive — p = 0.059 is just outside the conventional 5% threshold.
Monthly stops vs. crime · normalized · 2025Monthly stops vs. crime · 2025 iBoth series are normalized to 0–1 scale. If stops suppress crime, high-stop months should have low-crime months. Instead, month-to-month patterns are inconsistent with a simple causal deterrent effect.Both lines are scaled to the same range so you can compare them directly. If more stops meant less crime, the lines would move in opposite directions. They don't — there's no consistent pattern month to month.
Stop outcomes · what happens after a stopWhat happens after a stop
The Sigma Squared Engine
We built this from Nashville's public data. We can build it from yours.
Every chart and finding above is produced by the same causal inference methods used in peer-reviewed criminology and economics research — run automatically against your department's stop records, crime data, and dispatch logs. Public data gets you this far. Your real data gets you further.
Deterrence analysis Racial disparity benchmarking Outcome & threshold testing Response-time impact modeling Officer retention & attrition modeling
Run this for my department →
If you're curious what that looks like for your department, let's talk.