Skip to main content
Policy Simulation

The "Bad Apples" Problem
Can We Identify Them?

Explore how data density bias affects our understanding of complaint concentration, and simulate the impact of early warning system policies.

Simulation Parameters

1.50 complaints/officer

The Data Density Illusion

When complaints are sparse relative to officers, even random assignment creates apparent concentration. The "top 2% account for X% of complaints" statistic is misleading without comparing to this null distribution.

Top 2% Share (Actual)
15.2%
Top 2% Share (Random)
7.2%
Naive Relative Risk
8.8x
Adjusted Relative Risk
2.3x

Complaint Concentration Curves

The gap between the red line (actual) and gray dashed line (random) represents true concentration. The gap between gray dashed and black (uniform) represents data density bias - concentration that appears even under pure randomness.

Understanding the Math

The naive calculation suggests the top 2% of officers are 8.8x more likely to generate complaints. However, even under random assignment, the top 2% would be 3.8x more likely simply due to statistical clustering.

After correcting for data density bias, the top 2% are actually only 2.3x more likely to generate complaints—still elevated, but far less dramatic than the naive estimate.

The Prediction Problem

Officers in the top 2% during probation have only a ~2-4% chance of remaining in the top 2% over the next 10 years. Even with a 5-year observation window, positive predictive values remain modest, making surgical "bad apple" removal challenging.

Persistence Matrix

Positive Predictive Values (%) — Rows: probation ranking | Columns: 10-year follow-up ranking

Pre-PeriodTop 2%Top 5%Top 10%Top 20%
Top 2%2.0%8.0%18.0%37.0%
Top 5%3.5%12.0%22.0%35.0%
Top 10%3.4%8.5%17.0%28.0%
Top 20%3.0%8.3%15.0%26.0%

Example: Of officers in the top 10% during their probationary period, only ~17% remain in the top 10% over the following 10 years. This limits the effectiveness of early identification strategies.

Modest Incapacitation Effects

Chalfin & Kaplan find that removing the top 10% of officers (identified ex ante) and replacing them with median officers would reduce complaints by only 4-6%. This modest effect stems from both prediction difficulty and the fact that replacement officers also generate complaints.

Policy Parameters

Click "Run Simulation" to see estimated policy impact.

Build Your Own Scenario

Adjust the parameters to model your department's situation. The simulation will estimate how much removing "bad apples" might reduce complaints, accounting for prediction error and replacement effects.

Your Parameters

Data Density1.50 complaints/officer

Projected Outcomes

Click "Calculate Impact" to see results

Comparison with Chalfin & Kaplan Findings

ScenarioPaper (Chicago)Your Simulation
Remove top 2%, median replacement-1.2%
Remove top 10%, median replacement-6.1%
Remove top 10%, same district 70-90th-5.2%
Citation

Chalfin, A., & Kaplan, J. (2021). How many complaints against police officers can be abated by incapacitating a few "bad apples?" Criminology & Public Policy, 20(2), 351-370. https://doi.org/10.1111/1745-9133.12542

Interactive analysis by Ian T. Adams, Ph.D.
Simulation results are illustrative. Actual effects will vary based on local conditions.