Skip to content

README.md

Latest commit

 

History

History
231 lines (168 loc) · 9.16 KB

README.md

File metadata and controls

231 lines (168 loc) · 9.16 KB

Time Series Causal Analysis of U.S. Air Pollution (2000–2016)

Python Jupyter License Status

Quantifying causal relationships between atmospheric pollutants across U.S. states using Granger causality, OLS regression, and time series econometrics — enabling data-driven environmental policy decisions.

Table of Contents

Business Impact

Air quality degradation costs the U.S. economy hundreds of billions of dollars annually through healthcare expenditures, lost productivity, and regulatory non-compliance penalties. This project delivers a statistically rigorous causal inference framework that directly addresses critical environmental and economic questions:

  • Policy Targeting: By identifying which pollutants causally drive ozone (O3) concentrations, regulatory agencies (EPA, state DEQs) can prioritize enforcement resources on the highest-impact emission sources — reducing compliance costs and maximizing health outcomes.
  • Predictive Risk Modeling: The OLS regression model (R² = 0.596) enables forecasting of ozone AQI changes from upstream NO2 and CO measurements, providing a leading indicator for air quality alerts and public health advisories.
  • Regional Strategy: State-level pollutant trend analysis supports granular resource allocation for environmental compliance programs, infrastructure investment, and industrial permitting decisions.
  • Environmental ESG Reporting: Organizations operating in affected states can leverage causal pollutant maps to quantify emissions impact and demonstrate regulatory compliance within ESG frameworks.

Project Overview

This project conducts a comprehensive time series causal analysis of four major air pollutants — Carbon Monoxide (CO), Nitrogen Dioxide (NO2), Ozone (O3), and Sulfur Dioxide (SO2) — across all U.S. states using EPA monitoring data spanning 2000–2016.

The analytical pipeline progresses from exploratory data analysis through stationarity testing, multicollinearity diagnostics, regression modeling, and Granger causality inference — forming a complete econometric study of inter-pollutant dynamics.

Core Question: Do changes in NO2 and CO concentrations causally precede and predict changes in ground-level ozone — and if so, with what lag structure?

Key Findings

Finding Detail
Granger Causality: NO2 → O3 NO2 Granger-causes O3 at lags 1 and 2 (p < 0.001)
Granger Causality: CO → O3 CO Granger-causes O3 at lags 1 and 2 (p < 0.001)
OLS Model Performance R² = 0.596 on first-differenced monthly data
High Inter-Pollutant Correlation Pearson correlation ~0.9 between pollutant AQI measures
Non-Stationarity Addressed First differencing confirmed stationarity (ADF test)
Multicollinearity Detected High VIF values; addressed via differencing and variable selection
Significant Predictors NO2 and CO first-difference lags are statistically significant drivers of O3_diff

Dataset

  • Source: U.S. Environmental Protection Agency (EPA) — National Ambient Air Quality Standards (NAAQS)
  • Temporal Coverage: January 2000 – December 2016 (monthly aggregates)
  • Geographic Scope: All U.S. states
  • Pollutants Tracked:
    • NO2 — Nitrogen Dioxide (AQI & mean concentration)
    • O3 — Ozone (AQI & mean concentration)
    • SO2 — Sulfur Dioxide (AQI & mean concentration)
    • CO — Carbon Monoxide (AQI & mean concentration)
  • Key Variables: AQI index, pollutant mean (ppm/ppb), state, date

Methodology

The project follows a structured econometric workflow:

1. Exploratory Data Analysis (EDA)

  • Descriptive statistics and null value audit across all pollutant series
  • Correlation heatmap revealing strong inter-pollutant relationships (~0.9)
  • State-wise NO2 time series visualization to identify regional patterns

2. Multicollinearity Diagnostics

  • Variance Inflation Factor (VIF) analysis to quantify predictor redundancy
  • Informs feature selection for downstream regression modeling

3. Stationarity Testing

  • Augmented Dickey-Fuller (ADF) Test applied to all pollutant series
  • First-order differencing applied to achieve stationarity for regression and causality testing

4. OLS Regression Modeling

  • Ordinary Least Squares regression on first-differenced series
  • Target variable: O3_diff (monthly change in ozone AQI)
  • Predictors: lagged differences of NO2, CO, SO2 concentrations
  • Model evaluation: R², coefficient significance (t-statistics, p-values)

5. Granger Causality Testing

  • Bivariate Granger causality tests at lags 1 and 2
  • Tests whether NO2 and CO time series contain predictive information about future O3 values beyond O3's own history
  • Statistical threshold: p < 0.05

Technical Architecture

Data Ingestion (EPA CSV)
        │
        ▼
Exploratory Data Analysis
   ├── Descriptive Statistics
   ├── Null Value Analysis
   └── Correlation Heatmap
        │
        ▼
Multicollinearity Check (VIF)
        │
        ▼
Stationarity Testing (ADF)
        │
        ▼
First-Order Differencing
        │
        ├──► OLS Regression Model (R²=0.596)
        │
        └──► Granger Causality Tests
                 ├── NO2 → O3 (p<0.001, lags 1-2)
                 └── CO  → O3 (p<0.001, lags 1-2)

Results

Granger Causality Summary

Hypothesis: NO2 does NOT Granger-cause O3
  Lag 1: F-statistic significant, p < 0.001  → REJECTED
  Lag 2: F-statistic significant, p < 0.001  → REJECTED

Hypothesis: CO does NOT Granger-cause O3
  Lag 1: F-statistic significant, p < 0.001  → REJECTED
  Lag 2: F-statistic significant, p < 0.001  → REJECTED

Interpretation: Changes in nitrogen dioxide and carbon monoxide concentrations statistically precede and predict changes in ozone levels, providing strong causal evidence that these primary pollutants drive secondary ozone formation — a finding with direct implications for emission source prioritization.

OLS Regression Model

Dependent Variable:  O3_diff (First-differenced Ozone AQI)
R-squared:           0.596
Significant Predictors:
  - NO2_diff_lag1    (p < 0.05)
  - CO_diff_lag1     (p < 0.05)
  - NO2_diff_lag2    (p < 0.05)

Repository Structure

Time-Series-Causal-Analysis-Project/
├── Time_Series_Analysis.ipynb   # Full analysis pipeline (EDA → Causality)
├── Group6_Final_ReportNew.pdf   # Academic research report (UTD, 2023)
└── README.md                    # Project documentation

Getting Started

Prerequisites

pip install numpy pandas matplotlib seaborn statsmodels scikit-learn linearmodels

Running the Analysis

  1. Clone the repository:
git clone https://github.com/ManojMareedu/Time-Series-Causal-Analysis-Project.git
cd Time-Series-Causal-Analysis-Project
  1. Launch Jupyter Notebook:
jupyter notebook Time_Series_Analysis.ipynb
  1. Execute cells sequentially — the notebook is self-contained and produces all EDA plots, model outputs, and causality test results inline.

Note: The notebook was originally developed on Google Colab with TPU acceleration. For local execution, the !pip install linearmodels cell handles all additional dependencies.

Tech Stack

Category Tools
Language Python 3.x
Data Manipulation Pandas, NumPy
Statistical Modeling Statsmodels (OLS, ADF, Granger), Scikit-learn
Visualization Matplotlib, Seaborn
Environment Jupyter Notebook, Google Colab (TPU)
Domain Time Series Econometrics, Causal Inference, Environmental Analytics

Authors

Developed as a graduate research project at the University of Texas at Dallas (2023).

Team: Manoj Mareedu, Prriyamvradha Parthasarathi, Premi Jawahar Vasagam, Mira Radhakrishnan, Sofia Rajan, Martin Navarro, Vyshnavi Gangineni, Siva Renuka Chowdary Nandigam

Future Enhancements

  • Incorporate ARIMA / SARIMA / VAR models for multivariate time series forecasting
  • Extend Granger causality to bidirectional and panel data frameworks
  • Interactive geospatial dashboard (Plotly/Dash) for state-level pollutant monitoring
  • Real-time EPA data integration via API pipeline
  • Deep learning approaches (LSTM, Temporal Fusion Transformer) for AQI prediction

This project demonstrates applied expertise in time series analysis, econometric causal inference, and environmental data science — skill sets directly applicable to data science, ML engineering, and quantitative analytics roles.