Segurança Urbana e Machine Learning: Avaliando o Impacto da Iluminação Pública e de Outros Fatores na Criminalidade Brasileira

This project investigates the relationship between urban environment characteristics and lethal violence in Brazilian municipalities in 2022, with special emphasis on the presence of public lighting on streets. Using aggregated data from the 2022 Population Census (SIDRA/IBGE), municipal indicators were constructed for dwelling surroundings (public lighting and sidewalks), urban–rural status, sex and race composition, education (average years of schooling, literacy and school attendance), labour market conditions and average nominal income. These variables were combined with homicide data from IpeaData to compute municipal homicide rates per 100,000 inhabitants. The empirical strategy first estimates linear regression (OLS) models relating homicide rates to the share of residents living on streets with public lighting, controlling for a broad set of socioeconomic and demographic covariates. Next, the analysis applies the Double Machine Learning (DML) framework, using three machine learning algorithms — Random Forest, Lasso and XGBoost — to flexibly model the effect of the covariates and obtain orthogonalized estimates of the impact of public lighting. Results show that the proportion of residents on lit streets is positively and significantly associated with homicide rates, with coefficients around 0.51 across all specifications, whereas higher availability of sidewalks and better education indicators are linked to lower levels of lethal violence. The consistency between OLS and DML estimates suggests that this association is robust, although it cannot be interpreted as strictly causal, highlighting the relevance of reverse causality and unobserved institutional factors when designing evidence-informed urban security policies.