📉 Quantifying Policy Impact: Sheffield Clean Air Zone Analysis

A rigorous time series analysis evaluating the causal effectiveness of Sheffield's 2023 Clean Air Zone (CAZ) intervention on Nitrogen Dioxide (NO₂) concentrations.

---

📊 Executive Summary

This study employs counterfactual inference and interrupted time series (ITS) regression to quantify the impact of the Sheffield Clean Air Zone implemented on February 27, 2023. By analyzing 35,064 hourly observations (Jan 2022 – Dec 2025) and controlling for seasonality and exogenous factors, this analysis demonstrates a statistically significant reduction in traffic-related pollution.

🏆 Key Findings

• Significant Reduction: The CAZ intervention drove a 41% weather-normalised reduction in NO₂ concentrations.
• Structural Break: Statistical testing confirms a highly significant structural break ($p < 0.001$) at the intervention date.
• Weather Robustness: Results hold even after accounting for wind speed, temperature, and precipitation using XGBoost.
• Model Agreement:
  • ITS Regression: Estimated -2.35 µg/m³ immediate drop.
  • Prophet Forecast: Observed values 44.2% below counterfactual baseline.
  • ARIMA Baseline: Observed values 29.9% below baseline.

---

🔬 Technical Methodology

To ensure robustness, this analysis utilizes a multi-model approach to isolate the policy effect from natural variation.

1. Baseline Forecasting (ARIMA & Prophet)

Constructed a "business as usual" counterfactual based on pre-intervention dynamics. The Prophet model captured complex multi-seasonal patterns (yearly, weekly), revealing that traditional ARIMA methods may underestimate the effect size when strong seasonalities are present.

2. Machine Learning Weather Normalisation (XGBoost)

Used Gradient Boosting to model non-linear relationships between weather (wind, temperature) and pollution. This proved that the reduction was driven by emissions changes, not favorable weather conditions.

3. Interrupted Time Series (ITS) Regression

Formal statistical test for causal structural breaks.

• Level Change: Immediate drop ($\beta_2 = -2.35$, $p<0.01$).
• Trend Change: Accelerated rate of improvement ($\beta_3 = -0.014$, $p<0.001$).

---

🚀 How to Run the Analysis

This repository contains a fully reproducible R pipeline. Follow these steps to replicate the study:

Prerequisites

• Language: R (v4.5.2+)
• Key Libraries: forecast, prophet, xgboost, keras3, tidyverse, zoo, lmtest

Execution Guide

Run the scripts in the scripts/ folder in numerical order:

git clone https://github.com/Vedant-ghadi/Sheffield-caz-analysis.git

1. Data Extraction (Scripts 01-02):

   • 01_extract_air_quality.R: Fetches hourly pollution data from Open-Meteo API.
   • 02_extract_weather_merge.R: Retreives historical weather data and merges datasets.

2. Analysis & Modelling (Scripts 03-08):

   • 03_eda_weather_pollution.R: Generates initial exploratory plots.
   • 04_model_arima_baseline.R & 05_model_prophet.R: Trains baseline forecasting models.
   • 06_model_xgboost.R: Performs weather normalisation.
   • 07_model_its.R: Runs the Causal Impact regression.
   • 08_model_lstm.R: (Experimental) Deep learning forecast.

3. Reporting (Scripts 09-10):

   • 09_viz_composites.R: Generates the final composite figures for the report.
   • 10_viz_gallery.R: Creates supplementary diagnostic plots.

---

📉 Visualization Gallery

The Structural Break

The dashboard below visualizes the raw time series against the intervention timeline. Note the clear shift in the data distribution post-February 2023.

Forecast vs. Actuals

The divergence between the counterfactual forecast (dotted) and observed data (solid) represents the "clean air dividend" generated by the policy.

---

👤 Author

Vedant Ghadigaonkar
Data Scientist | Time Series Analysis | Policy Evaluation

• GitHub: @Vedant-ghadi
• LinkedIn: Vedant Ghadigaonkar

---

Data provided by the Copernicus Atmosphere Monitoring Service (CAMS) via Open-Meteo API.