This project investigates systemic racial bias in police traffic stops using a novel outcome-sensitive predictive modeling approach. Rather than analysing stop frequency or hit rates, it tests whether the outcome of a stop (e.g., citation, arrest) can be predicted using race-neutral features, and how prediction accuracy changes when race is included.
Conducted as one of two capstone research projects during the Master of Data Analysis at Queensland University of Technology (QUT).
Alex Conroy
Master of Data Analysis — QUT
📚 Independent Research Dissertation
📍 Project 1 of 2
Can we model the decision to arrest or cite a driver without knowing their race — and does model accuracy improve if race is included?
This reframes the problem from stop bias to decision-making sensitivity, helping avoid confounders present in traditional hit rate or stop frequency analysis.
- Stanford Open Policing Project (SOP): Aggregated traffic stop data across U.S. states
- U.S. Census Bureau ZCTA Crosswalk: Demographic context by ZIP code
- Google Maps Geocoding API: Reverse-geocoding of stop locations (ZIP → ZCTA)
- Reverse geocoded ZIP codes to ZCTA using Google Maps API
- Merged with U.S. Census population and race/ethnicity data
- Constructed multinomial logistic regression models using
nnet::multinom() - Compared outcome prediction accuracy with and without racial features
- Evaluated across states (e.g. Colorado, North Carolina) and time windows
| Metric | With Race | Without Race |
|---|---|---|
| Predictive Accuracy | ↑ Improved | ↓ Decreased |
| State-Level Consistency | High | High |
| Interpretability | Transparent | Transparent |
✅ Racial identifiers increase predictive power, implying bias in outcome decisions
✅ Shows decision-level sensitivity to race, not just stop frequency bias
This project uses both Python and R, with each language selected based on its strengths:
- 🔍 R was chosen for modeling due to the availability of the
nnet::multinom()function for multinomial logistic regression and its strength in explainable statistical modeling. - 🌐 Python was used for reverse geocoding via the Google Maps API, as well as for ZIP/ZCTA preprocessing and dataset merging.
This hybrid approach reflects real-world, tool-agnostic decision making in applied data science projects.
racial-bias-traffic-stops/ ├── data/ │ ├── sop_subset.csv # Subsampled SOP stop data │ └── census_zcta_mapping.csv # ZIP-to-ZCTA crosswalks │ ├── notebooks/ │ ├── zcta_cluster_analysis.ipynb # Geospatial and census merging (Python) │ ├── reverse_geo_code.ipynb # Google Maps geocoder (Python) │ ├── analysis/ │ └── SOPModel_all_test.Rmd # Multinomial regression modeling (R) │ ├── report/ │ ├── Report DRAFT.pdf # Draft dissertation write-up │ ├── IFN704_Project_Proposal.pdf # Initial research proposal │ └── IFN704_Alex_Conroy_Presentation.pdf # Final presentation slides │ ├── research.txt # Raw notes and research log ├── README.md
- Python: pandas, requests, geopy, geocoding APIs
- R:
nnet,car,dplyr,ggplot2 - Jupyter + RMarkdown: Mixed-language data science pipeline
- PDF Reporting: Academic communication and presentation
- Models cannot infer causality or intent
- Data only includes recorded stop outcomes, not reasoning
- Geocoding errors or data gaps may introduce spatial noise
- SOP data completeness varies by state
- Expand to additional states or newer SOP releases
- Apply causal inference or counterfactual analysis
- Integrate police bodycam metadata or context from court outcomes
- Develop fairness-aware models using sensitive feature auditing
This project uses public data from the Stanford Open Policing Project and the U.S. Census Bureau.
No proprietary or confidential datasets were used.
This project shows how transparent, explainable modeling and thoughtful geospatial analysis can support conversations around bias, ethics, and decision fairness — essential for data science in government, policy, and social impact sectors.