Research

Under Review

  • Selim Yaman, Abdullah Atalan, and Jeff Gill. “Bridging Prediction and Theory: Introducing the Bayesian Partially-Protected Lasso” [Presented at PolMeth 2023]
    Abstract This work examines the application of regularization methods in political science, both theoretically and practically. The focus of the study is on the use of regularization to specify models for predicting vote choice and conflict. Specifically, what is the predictive price for protecting theoretically important variables from shrinkage? While many machine learning applications prioritize maximizing predictive accuracy, political science relies on reliable prior findings as a theoretical foundation. Through reanalysis of various political science studies that utilized Lasso, we illustrate that crucial theoretically-driven variables can be shrunk to insignificance. To address this, we propose a protected Lasso approach in a Bayesian framework that safeguards these variables, balancing theoretical robustness and predictive power. Our analyses, applied to a diverse range of political topics, demonstrate the effectiveness of this approach. Finally, we introduce an R package \texttt{ProtectR} to implement Partially-protected Bayesian Lasso, offering researchers a practical tool for variable selection and prediction that respects the theoretical underpinnings of their models.

Working Papers

  • Selim Yaman “Beyond the Tweets: Leveraging Language Models for Estimating Political Ideology on Twitter” [Presented at CompText 2023]

    Abstract This article investigates the potential of advanced transformer-based language models for estimating the political ideology of Twitter users, presenting an alternative to traditional network-based methods. We argue that these state-of-the-art language models may offer a more accurate and efficient means of capturing ideological signals from tweet content compared to network relations. We begin by reviewing existing methods and their applications in various political science research contexts. Subsequently, we propose that leveraging language models for ideology estimation in non-elite users can yield more efficient and accurate results. By examining the performance of these pre-trained models in comparison to network-based approaches, our methodology aims to contribute to the ongoing debate on the most effective ways to estimate user ideology. Ultimately, our research seeks to develop an accessible tool for estimating user ideology with improved accuracy, fostering broader applications in political science research, and deepening our understanding of the political landscape on social media platforms like Twitter.
  • Selim Yaman “Unveiling the Alliance Effect: How Warlord-Government Relations Shape Local Economies in Conflict Zones” [Presented at MPSA 2021]

    Abstract Extant literature often approaches the relationship between conflict and economic growth through broad, national-level analyses, thereby eclipsing the intricate local-level dynamics. This study shifts the focus to explore how alliances between warlords and central governments can mediate the impact of civil conflict on local economic conditions, specifically within the context of Afghanistan from 1993 to 2013. Utilizing a novel combination of night-time light emission data and conflict records, I uncover that not all regions are negatively affected by war. In fact, areas under the control of warlords who ally with the central government may even experience relative economic stability or incremental growth. These trends are particularly pronounced in urban centers, which are less susceptible to the economic ravages of conflict due to their role as financial lifelines for all parties involved. The research adds nuanced understanding to the war-economy relationship by introducing the variable of political alliances at the subnational level.
  • Selim Yaman “Do Politics Stop at the Water’s Edge? Evidence from Twitter Discussions on Afghanistan Withdrawal” [Presented at AU Work-in-Progress Series]

    Abstract The extent of US political polarization in foreign policy debates in comparison to domestic issues has been a subject of robust scholarly discussion with varying conclusions. A section of the literature posits that the divide in foreign policy mirrors the stark polarization observed in domestic affairs. Conversely, others propose a more unified perspective on foreign matters, suggesting a tendency for Americans to 'rally around the flag' in such contexts. This study, however, proposes a nuanced understanding that political polarization in foreign policy is contingent on the topic under scrutiny. I assert that conservatives and liberals converge on universal values, yet distinct ideological differences surface in relation to specific policy matters. I illuminate these patterns through an exploration of Twitter discussions on the 2021 Afghanistan withdrawal. Utilizing sentiment analysis and topic modeling techniques, our research provides empirical evidence to support the hypothesis of topic-specific polarization.
  • Abdullah Atalan, Selim Yaman, and Jeff Gill “Bayesian Partially-Protected Lasso: Extension and Applications in Political Science”

    Abstract In political science data analysis, methods balancing statistical robustness with theoretical relevance are essential. This article delves into the Bayesian Partially-Protected Lasso (BPL), merging the resilience of Bayesian Lasso with the ability to protect key variables from shrinkage. We: 1) investigate the mathematical underpinnings, differentiating prior distributions for protected and non-protected variables; 2) adapt the method for regularized logistic regression for binary outcomes; 3) test its efficacy using simulated datasets and seven re-analyses of political science studies. Responding to the need for nuanced tools in political science data, we introduce the R package \texttt{ProtectR} for the implementation of the BPL, equipping researchers with a tool for variable selection that honors theoretical foundations. The article concludes with reflections on the results and suggestions for future research.
  • Ahmet Utku Akbiyik, Muhammed Akkus, and Selim Yaman “Dams: Infrastructural Investment as a Tool of Development and Peace”

    Abstract Governments and donors often highlight development spending and public works projects as mechanisms to address the root causes of violent conflict. However, empirical evidence about their impacts is limited. In our study, we investigate the role of infrastructure projects as tools to counter insurgency and civil war violence. Specifically, we focus on dam construction in the Kurdish region of Turkey, analyzing if economic development spurred by infrastructural investments can decrease insurgent recruitment and attacks. We utilize exogenous variation in the conditions required for dam construction to estimate the impact of these investments on violence and rebel recruitment at the district level. Additionally, we employ a difference-in-difference analysis of georeferenced data on dams, irrigation, and conflict. Through this district-level analysis, we contribute to the critical discussion on the influence of development on political violence, providing insights for future policy implementations aimed at reducing political violence in diverse settings.

In Progress

  • Selim Yaman and Mustafa Kocyigit. “Beyond Human Subjects: Large Language Models as Participants in Political Experimentation”
    Abstract We introduce a novel framework that employs large language models (LLMs) to simulate political experiments. Building upon the foundation laid by Argyle et al. (2023), our work advances from AI-generated survey response predictions to the complete simulation of political lab experiments. We demonstrate this through the replication of two seminal political science studies: one examining the impact of inaction inertia in international negotiations, and the other investigating gender differences in candidate emergence. Our simulations, powered by 'silicon samples' generated by LLMs, allow for a detailed simulation of human behavior within controlled environments. The behavioral responses of our LLM-powered agents consistently align with the original studies, validating the applicability of LLMs in modeling complex human behaviors in a political context. Our framework opens up new possibilities, providing an additional tool for researchers to study political phenomena, including the capability to explore scenarios via artificial group experiments and large-scale simulations that were previously challenging due to ethical or logistical constraints. The successful implementation of our framework presents promising directions for further research into political behavior using LLM-powered autonomous agents.