Chrisolande/Bayesian-Structural-Time-Series-for-Causal-Impact

Causal impact of COVID-19 on Kenya food prices (WFP) using BSTS, DiD, RDD & IV. Triangulated Bayesian & frequentist inference. # Causal Inference: COVID-19 & Food Price Volatility in Kenya [WIP] ## 📌 Project Overview This project investigates the causal impact of COVID-19 lockdowns (March 2020) on food prices across 63 Kenyan markets. Using a single-source dataset from the **World Food Programme (WFP)**, the study identifies treatment effects through geographic variation in supply chain dependence and market remoteness. The core of this project is **methodological triangulation**: using five distinct causal inference frameworks to ensure the estimated effects are robust to unmeasured confounding and model specification. ## 🛠 Research Strategy The project is structured into a Quarto book covering a multi-pronged causal "attack": 1. **Identification via DAGs:** Modeling the 2019-2020 drought and fuel prices as mediators/confounders using `dagitty`. 2. **Geographic Instruments:** Computing `distance_from_nairobi` via Haversine formula to proxy for supply hub dependence. 3. **Triangulated Estimation:** Running Matching, IPW, DiD, RDD, and IV to check for sign and magnitude agreement across estimands. 4. **Bayesian Counterfactuals:** Final synthesis using `CausalImpact` (BSTS) to visualize the "What If" scenario for Nairobi. ## 📦 Tech Stack - **Language:** R - **Causal Frameworks:** `CausalImpact`, `bsts`, `ggdag`, `MatchIt`, `WeightIt`, `fixest`, `rdrobust` - **Data Engineering:** `tidyverse` (dplyr, tidyr, ggplot2, sf/geosphere) - **Reporting:** Quarto Book ## 📈 Analysis Roadmap - [x] **Chapter 1:** Data Ingestion (WFP 2-row header parsing) & Haversine Distance Calculation. - [x] **Chapter 2:** Causal Assumptions & DAG Construction (`dagitty`). - [ ] **Chapter 3:** Matching (Nearest Neighbor on pre-COVID price levels/trends). - [ ] **Chapter 4:** Inverse Probability Weighting (IPW) & Overlap Diagnostics. - [ ] **Chapter 5:** Difference-in-Differences (Callaway-Sant'Anna for staggered timing). - [ ] **Chapter 6:** Regression Discontinuity (200km "Curfew-Window" Cutoff). - [ ] **Chapter 7: …