Prediction: Bias

---

How We Determined Predictive Policing Software Disproportionately Targeted Low-Income, Black, and Latino Neighborhoods

Satellite view of the neighborhood of Northridge, in Los Angeles, Calif., with PredPol predictions heat map overlay

This repository contains the code and data for the analysis we carried out for our investigation, which is described in detail in our methodology. It relies on the R package Targets to generate a reproducible analysis pipeline.

For the sake of convenience, we've already committed the output from the pipeline to the repository. You will find it in the out folder. If you want to re-run the analysis and verify the steps yourself, follow the instructions in the Setup section of the Readme.

Data

In this repository the data used as the input for analysis is stored in the in folder, and the dataframes, plots, and findings that are generated from the analysis are stored in the out folder. There is also a datasheets folder that contains a one-page summary of this analysis for each jurisdiction individually and the raw reports we obtained from PredPol's servers.

in folder

Folder or Filename	Description
all_jurisdictions.csv	2018 ACS 5-year population estimates for race and household income.
all_predictions.csv	PredPol predictions for all departments. These have been parsed from the raw HTML reports described in the methodology.
arrests.csv	Arrest records for 13 departments obtained through public records requests.
arrests-per-capita.csv	Arrests per capita rates for the departments that provided us with records
block-level-prediction-counts.csv	Prediction data aggregated to the Census block level, providing information regarding how many days of the analysis window a particular block received predictions.
block-level-prediction-arrest-counts.csv	Daily arrest and prediction counts at the block level for for 11 jurisdictions in our analysis.
departments.csv	Metadata for the 38 departments in this analysis.
housing.csv	Information regarding local and subsidized housing made publicly available by the U.S. Department of Housing and Urban Development.
stable-bgs.csv	Stable block groups used in the block level race analysis section of the methodlogy.
states-geojson	GeoJSON data for the states that the 38 departments of our analysis belong to.
ucr.csv	Race/Ethnicity arrest information made publicly available by the FBI’s Uniform Crime Reporting Program.
uof.csv	Use-of-force data for six deparments; obtained through public records requests.

out folder

Folder or Filename	Description
dataframes	Dataframes used to generate the charts in the plots folder.
dept_prediction_date_ranges.csv	First and last dates from the prediction reports for each department.
dept_usage_dates.csv	Date ranges for each department in our analysis. Ranges include predictions, software usage start and end dates confirmed by each department, and the start and end dates used in this analysis.
findings.pdf	This PDF contains the code and output for the anecdotal information that was included in the methodology.
jurisdiction-prop-cheatsheet.pdf	PDF containing demographic proportion information as described in the disparate impact analysis of our methodology for each department.
plots	This folder contains all the charts we created during our analysis.
prediction_count.csv

PredPol (now Geolitica) Reports

You can download the raw PredPol reports used as the basis of this analysis here. Note: The addresses listed in the prediction reports are the center of the prediction box. The actual predictions encompass the entire box, not just the address marked at the center.

Datasheets folder

Folder or Filename	Description
datasheet.Rmd	R Markdown that generates the datasheet
geojson	Folder that contains GeoJSON that is used for generating maps
html	Folder that contains HTML version of the datasheets
pdf	Folder that contains PDF version of the datasheets

Setup

To run the analysis

• Install the following packages in RStudio using install.packages:
  • tidyverse
  • targets
  • glue
  • tarchetypes
  • leaflet
• Extract the contents of in.tar.xz.

Run Analysis

• Run library(targets)
• Run tar_make().

Once you folow the setup steps and run tar_make(), the entire pipeline will be run and all the analysis artifacts will be generated. The first time around this usually takes about 10 minutes.

If you are running this repository for the first time and have set it up correctly, there should be no difference between what is already in the out folder and what is generated after running tar_make().

The out folder contains the latest output from the analysis already. If you re-run the pipeline after using tar_make() adding the in folder, nothing should change if you have the latest in folder and codebase.

Data Sheets

Data sheets looking at each department in our analysis individually. These are also available in PDF form at datasheets/pdfs.

Department	URL
alexandriapd	Link
birminghampd	Link
boonecounty	Link
clovisca	Link
cpd	Link
cpso	Link
decaturga	Link
elgin	Link
elmonte	Link
farmersbranch	Link
forsythso	Link
frederick	Link
fresnopd	Link
ftmyerspd	Link
gloucestertwppd	Link
gvpd	Link
haverhill	Link
homewood	Link
jacksonvilletx	Link
jeffcomacc	Link
la	Link
livermore	Link
merced	Link
modesto	Link
nilespolice	Link
ocalapdcom	Link
ocfl	Link
ocoeepd	Link
piscataway	Link
plainfieldpdnj	Link
portagemipd	Link
salisbury	Link
southjordan	Link
tacoma	Link
templeterracepd	Link
tracypd	Link
turlockpolice	Link
westspringfieldpolice	Link