2022 Bakerman_Dynamic Logistic Regression and Variable Selection Forecasting and Contextualizing Civil Unrest

Briefing Document: Dynamic Logistic Regression and Variable Selection for Forecasting Civil Unrest

Authors

Jordan Bakerman, Karl Pazdernik, Gizem Korkmaz, Alyson G. Wilson

North Carolina State University, Pacific Northwest National Laboratory, University of Virginia

Overview

The paper discusses the application of dynamic logistic regression and variable selection methodologies to forecast civil unrest events, utilizing real-time data sourced from Twitter. The authors propose a Bayesian framework that enables efficient model parameter updates and variable selection based on the evolving information landscape.

Introduction

Civil unrest can manifest in various forms, from peaceful protests to violent conflicts. Accurate forecasting of such events is essential for resource allocation and public safety. The increasing use of Twitter as a communication platform has provided researchers with a rich source of real-time data. By mining user-generated content and employing daily word counts, researchers can identify predictive terms that signal potential unrest. This paper emphasizes the need for predictive accuracy and variable selection to contextualize the reasons behind protests.

Methodology

Dynamic Model Fitting

The authors introduce a dynamic framework to model civil unrest. Dynamic models accommodate non-stationary time series data and allow for real-time updates to parameters as new information emerges. The dynamic logistic regression model is formulated using Bayesian techniques, particularly through the Kalman filter and Markov chain Monte Carlo (MCMC) methods.

1. Dynamic Linear Model (DLM): The authors outline the structure of a univariate DLM, specifying observation and state equations, which are crucial for the modeling process.
2. Dynamic Generalized Linear Model (DGLM): The observation equation is defined using a Bernoulli distribution, enabling the model to predict binary outcomes.

Variable Selection

Variable selection is crucial for understanding the underlying reasons for protests. The authors propose a new method using penalized credible regions, which allows for the identification of significant predictors while avoiding the limitations of traditional shrinkage priors. This dynamic variable selection framework can adaptively include or exclude predictors over time based on their relevance.

Simulation Studies

The effectiveness of the dynamic logistic regression model and variable selection approach is evaluated through simulation studies. The paper compares its performance against static models, including LASSO regularization and dynamic model averaging. Key metrics used for evaluation include precision, recall, and F1 scores, reflecting both predictive accuracy and variable selection performance.

Results

Prediction Performance

The dynamic logistic regression model consistently outperformed both the static LASSO model and the dynamic model averaging approach across various metrics. Key findings include:

• Precision and F1 scores were significantly higher for the dynamic model, often exceeding 95%.
• The LASSO model exhibited a tendency to over-predict events, resulting in lower precision.

Variable Selection Performance

The authors demonstrate that the dynamic logistic regression model excels in variable selection as well, effectively identifying true predictors while excluding irrelevant variables. The dynamic model's flexibility allows it to respond to changing social contexts, making it more adept at capturing the fluid nature of civil unrest.

Application to Civil Unrest Forecasting in Latin America

The proposed methodologies were applied to real-world data from Argentina, Brazil, Colombia, Mexico, Paraguay, and Venezuela, spanning from November 2012 to August 2014. The analysis involved:

• Utilizing protest-related Twitter terms as model features.
• Conducting forecasts for one-day and five-day intervals.

Results of the Application

• The dynamic model achieved high precision and recall rates, with F1 scores ranging from 97% to 100% in several countries.
• Insights gained included the identification of key terms that predicted civil unrest, offering context to the underlying causes of protests.

Conclusion

The paper presents a robust framework for forecasting civil unrest using dynamic logistic regression and variable selection methodologies. The incorporation of Bayesian techniques and real-time data from social media allows for improved prediction accuracy and contextual understanding of protest dynamics. The dynamic model outperforms traditional static approaches, demonstrating significant advantages in both prediction and variable selection.

Key Takeaways

1. Dynamic Nature of Civil Unrest: The paper underscores the importance of a dynamic modeling approach to capture the evolving patterns of civil unrest effectively.
2. Real-Time Data Utilization: Leveraging Twitter data allows for timely insights into potential unrest, enhancing predictive capabilities.
3. Bayesian Framework: The proposed methodologies provide a statistically sound approach to model fitting and variable selection, enabling continuous adaptation to new information.

This research offers significant implications for policymakers, security agencies, and social scientists interested in understanding and responding to civil unrest.