ProPublica’s COMPAS Data Revisited

This repository contains the Rmarkdown program that generates all the Figures and Tables in my July 8, 2019, arXiv paper. In that paper, I examine ProPublica’s COMPAS score and two-year general¹ recidivism dataset. I also include the actual full-length paper here as a markdown output file. This repository also contains several other related files. The files in this repository are the following:

Analysis_ProPub_COMPAS_Revisited.Rmd

• This file contains the Rmarkdown program that generates all the Figures and Tables (in HTML format) in my arXiv paper from July 8, 2019.

Code_to_correct_ProPub_two_year_data.md

• This file contains the few lines of R code one can use to implement the simple April 1, 2014, COMPAS screening date cutoff rule I use in my paper to correct ProPublica’s two-year general recidivism dataset. (This code also appears in the Rmarkdown program listed above, which generates all the Figures and Tables in my paper)

R_session_information.md

• This file contains the R session information, including the R version (3.5.1) and R packages used to generate my paper.

arXiv_ProPub_COMPAS_Revisited.md

• This file contains the markdown output file that renders a (full-length) copy of my arXiv paper. This allows one to view the paper as a webpage on GitHub. Except for formatting differences and the correction of two typos regarding the writing, this is identical to my July 8, 2019, arXiv paper.

formatting_GitHub_vs_arXiv

• Here, I list a few formatting differences between the GitHub markdown version of my July 8, 2019, arXiv paper in this repository and the PDF version on the arXiv website.

minor_clarifications.md

• Here, I list a few minor clarifications regarding some of the writing in my arXiv paper (and the same copy of that paper in this repository).

(ProPublica’s own COMPAS GitHub repository is here)

---

Summary of my arXiv paper

I examine the COMPAS recidivism risk score and criminal history data collected by ProPublica in 2016. ProPublica’s article and data analysis fueled intense debate and research in the nascent field of fair machine learning or algorithmic fairness.

My paper takes a closer look at the actual datasets put together by ProPublica. In particular, the sub-datasets built to study the likelihood of recidivism within two years of a defendant’s original COMPAS survey screening date. (This screening is typically performed the day of, or one day after, the arrest)

I visualize these data by analyzing the distribution of defendants across COMPAS screening dates. Doing so, I find that ProPublica made a substantial data processing error when it created these datasets. Namely, it failed to implement a two-year window sample cutoff for recidivists in such datasets (whereas it did implement such a sample cutoff for non-recidivists).

I show this data processing error in the Figure below for ProPublica’s two-year general¹ recidivism dataset. This is the key Figure in my paper (Figure 4).

[Note: There is also an unrelated, but very visible, drop in COMPAS screens or cases in mid-2013 (for recidivists and non-recidivists alike). This is a separate issue, however, which seems to be present in the original dataset that ProPublica received from Broward County, FL. So this particular issue does not appear to be a data processing mistake by ProPublica, and I do not address this in my paper]

As a result, ProPublica incorrectly kept a disproportionate share of recidivists in the two-year datasets. I estimate this biases the two-year general recidivism rate upward by almost nine percentage points, pushing it from 36.2% to 45.1%. The two-year general recidivism rate ProPublica calculates is therefore over 24% higher than the true two-year general recidivism rate in that same data when it is processed correctly.

In my research paper, I also explore how this data processing error impacts other statistics. Specifically, I look at ProPublica’s confusion matrix analysis of COMPAS Low/High score vs. two-year recidivism status. I find that the biased two-year datasets also have a substantial effect on the positive predictive value (or precision) and the negative predictive value.

On the other hand, the biased two-year datasets have relatively little impact on several other key statistics in the confusion matrix analysis, which are less susceptible to changes in the relative share of recidivists versus non-recidivists. In particular, the accuracy, false positive rate, and false negative rate.

To my knowledge, this is the first endeavor to highlight this data processing error made by ProPublica.

---

Matias Barenstein

First version of my paper on arXiv: June 11, 2019. E-mail: mbarenstein@gmail.com. The author is a staff economist at the Federal Trade Commission (FTC). This study was conducted independently of his work at the FTC. The views expressed in this article are those of the author. They do not necessarily represent those of the Federal Trade Commission or any of its Commissioners.

Bibtex citation:

    @ARTICLE{2019arXiv190604711B,
       author = {{Barenstein}, Matias},
        title = "{ProPublica's COMPAS Data Revisited}",
      journal = {arXiv e-prints},
         year = "2019",
        month = "Jun",
          eid = {arXiv:1906.04711},
        pages = {arXiv:1906.04711},
archivePrefix = {arXiv},
       eprint = {1906.04711},
 primaryClass = {econ.GN},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2019arXiv190604711B},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
      }

Keywords: Fair Machine Learning, Algorithmic Fairness, Recidivism, Risk, Bias, COMPAS, ProPublica

Footnotes

1. In my paper, I examine ProPublica’s two-year general recidivism dataset. General recidivism includes both violent and non-violent offenses. While I focus on this dataset, the two-year violent recidivism dataset that ProPublica also created suffers from the same data processing issue I identify here.