Geospatial Analysis of Police Lethality in São Paulo

Authors:
Luis Fernando Ramirez | Monserrat López Perez | Rodrigo Dornelles

Last updated: 2025-05-12

This analysis was made as part of the course GRAD-E1457 - Geospatial Analysis for Data Science, taught by Prof. Alexandra Kapp in the Hertie School Master of Data Science for Public Policy Program

📝 Project Summary

Brazil’s state of São Paulo reports hundreds of police-related deaths each year, but the where and why behind those deaths remain poorly understood. Using 2022 census micro-data, detailed shapefiles, and incident-level police-mortality records (2013-2025), we:

Geocode individual deaths and assign them to districts.
Distribute census population and household income down to census-sector slices that align with district borders.
Map raw counts, per-capita death rates, and Local Indicators of Spatial Association (LISA) clusters.
Model lethal-force risk with spatial-lag and spatial-error SAR models, testing demographic, income, and urban-form predictors.

The full workflow is scripted in reproducible R Markdown (R/PoliceKillingsFinal.Rmd) and documented below.

📁 Repository Layout


├── R/                         # Helper R scripts & functions
├── R/PoliceKillingsFinal.Rmd  # Main analysis / report (this file is the source of the results)
├── data-raw/                  # Original raw data (not tracked by Git LFS)
│   ├── MortesDecorrentesIntervencaoPolicial_2025.xlsx
│   ├── SIRGAS_SHP_*           # IBGE & PMSP shapefiles (districts, census blocks, favelas …)
│   └── bq_data_income_setor_censitario.csv
├── data/                      # Clean, processed data (generated by the R markdown)
│   ├── mortes_policia_clean.rds
│   └── distritos_pop_2022.rds
├── images/                    # Analysis figures exported by R (populated by the Rmd)
│   └── overview_maps.png
├── poster/                    # Conference poster
│   ├── final_poster.pdf
│   └── final_poster.jpg
└── README.md                  # ← you are here

⚙️ Quick-start

Clone the repo:

git clone https://github.com/rfdornelles/geospatial_assignments.git
cd geospatial_assignments

Open R (≥ 4.2) and knit R/PoliceKillingsFinal.Rmd:

rmarkdown::render("R/PoliceKillingsFinal.Rmd")

The first chunk automatically installs and loads all required packages:

c("here", "dplyr", "tidyr", "readr", "ggplot2", "sf", "lubridate",
  "viridis", "ggthemes", "spdep", "spatialreg", "patchwork",
  "janitor", "readxl", "stringr", "broom", "grid")

Outputs:
- Cleaned RDS files → data/
- Figures → images/
- An HTML report mirroring the Rmd → R/PoliceKillingsFinal.html

🗺️ Key Results

Download full poster → poster/final_poster.pdf

📊 Reproducing the Spatial Models

The SAR models use queen-contiguity neighbours via spdep:

nb <- spdep::poly2nb(deaths_map)
lw <- spdep::nb2listw(nb, style = "W")
lag_model <- spatialreg::lagsarlm(
  deaths_log ~ pct_parda + pct_preta + pct_masc + avg_age +
               pct_pm + pct_favela + income_per_capita + pop_density,
  data = model_data,
  listw = lw
)
summary(lag_model)

See Section 6 & 7 of R/PoliceKillingsFinal.Rmd for full diagnostics, Moran’s I on residuals, and coefficient tables.

👩‍💻 Contribution Statement

All authors contributed equally. - Modeling: Luis Ramirez - Visualization: Monserrat López - Data wrangling: Rodrigo Dornelles

📄 License & Citation

Code: MIT License (see LICENSE).
Cite this project as:

@misc{ramirez2025police_lethality_sp,
  author   = {Luis Fernando Ramirez Ruiz and Monserrat López Perez and Rodrigo Dornelles},
  title    = {Geospatial Analysis of Police Lethality in São Paulo},
  year     = {2025},
  url      = {https://github.com/rfdornelles/geospatial_assignments}
}

🙏 Acknowledgments

SSP-SP for releasing incident-level lethality data.
IBGE 2022 Census team for open micro-data & shapefiles.
R-Spatial community for sf, spdep, and spatialreg.

_{^{Questions or ideas? Open an issue!}}