Sezibra/conflict-causal-inference

Causal effect of UN peacekeeping on civilian violence in Africa using Double Machine Learning and Causal Forests. 43-country panel (2000-2022) built on UCDP GED v25.1 and EconML. # Causal Inference for Conflict with Machine Learning ## Research Question Does UN peacekeeping deployment reduce violence against civilians in Africa, and does this effect vary across different conflict contexts? ## Navigation | Section | Description | |---------|-------------| | Motivation | Why causal inference matters for conflict research | | Key Findings | Main results across all methods | | Data | Five merged data sources covering 43 African countries | | Methods | OLS, Double Machine Learning, Causal Forests | | Results in Detail | Figures and interpretation for each analytical step | | Robustness | Five alternative specifications and placebo test | | Notebooks | Four-notebook analytical progression | | How to Reproduce | Setup and replication instructions | | References | Academic sources | ## Motivation Prediction models show what correlates with violence, but they cannot answer whether an intervention *causes* a change in outcomes. This project applies causal machine learning methods to estimate the effect of UN peacekeeping on civilian violence using observational panel data from 43 African countries (2000–2022). ## Key Findings **Selection bias is severe.** A naive comparison shows peacekeeping countries have 2.90 *more* one-sided violence events per month, because the UN sends missions to the most violent places, not because peacekeeping causes violence. **OLS underestimates the protective effect.** After linear controls, OLS still estimates +0.49 events/month. Linear models cannot fully capture the nonlinear confounding. **Double Machine Learning flips the sign.** After flexible ML-based confounder control, the estimated average treatment effect is −0.16 events/month (95% CI includes zero, p = 0.82). The average effect is not statistically significant. **Heterogeneity reveals where peacekeeping works.** Causal Forests show that 8 of 12 peacekeeping countries have negative treatment effects. The strongest protective effects appear in low-inc …