Political Analysis

Extractive versus Generative Language Models for Political Conflict Text Classification | Political Analysis | Cambridge Core Extractive versus Generative Language Models for Political Conflict Text Classification

ConfliBERT is open source and is easily deployed and replicable. It is significantly better on comparable, relevant quality metrics and faster than other LLMS that use decoder technologies with graphical processing units (GPUs). Read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Extractive versus Generative Language Models for Political Conflict Text Classification,” P. Brandt, S. Alsarra, F. D’Orazio, @dagmarheintze.bsky.social, L. Khan, S. Meher, @javierosorio.bsky.social, & M. Sianan review and benchmark the ConfliBERT model.

The January 2026 issue of Political Analysis is out and currently free to read. Check it out now through the end of February!

Their BSA method is designed to address concerns about confounders that cannot be addressed by fixed effects. They illustrate this using a Monte Carlo simulation study and an empirical example on the effect of war on tax rates. Read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Bayesian Sensitivity Analysis for Unmeasured Confounding in Causal Panel Data Models,” Licheng Liu and Teppei Yamamoto develop a Bayesian sensitivity analysis (BSA) method for causal panel data analysis.

They find that complex prompting strategies can lead to improved model performance. The authors also offer several recommendations for researchers using LLMs for stance detection in political texts. You can read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Stay Tuned: Improving Sentiment Analysis and Stance Detection Using Large Language Model,” Max Griswold, Michael Robbins, and @sociologian.bsky.social evaluate fine-tuning strategies to improve LLM performance using social media data surrounding the 2020 election.

Political DEBATE: Efficient Zero-Shot and Few-Shot Classifiers for Political Text | Political Analysis | Cambridge Core Political DEBATE: Efficient Zero-Shot and Few-Shot Classifiers for Political Text

The Political Domain Enhanced BERT-based Algorithm for Textual Entailment (DEBATE) is benchmarked against other popular supervised classifiers. Ultimately, DEBATE is both efficient and completely open source. Read the paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Political DEBATE: Efficient Zero-Shot and Few-Shot Classifiers for Political Text,” Michael Burnham, Kayla Kahn, Ryan Yang Wang, and Rachel Peng introduce DEBATE, a new open source foundation model for classifying political documents.

Promotional banner for Political Analysis announcing 'New Issue Online' on a yellow and red background.

NEW ISSUE from @polanalysis.bsky.social -

Political Analysis - Volume 34 - Issue 1 - January 2026 - https://cup.org/4aAPBWB

Analyzing Political Text at Scale with Online Tensor LDA | Political Analysis | Cambridge Core Analyzing Political Text at Scale with Online Tensor LDA

Their method is demonstrated using social media conversations surrounding the MeToo movement and the 2020 presidential election. This method is an alternative to off-the-shelf methods such as LDA, which are computationally inefficient. Read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Analyzing Political Text at Scale with Online Tensor LDA,” @sarakangaslahti.bsky.social, Danny Ebanks, @jeankossaifi.bsky.social, Anqi Liu, @rmichaelalvarez.bsky.social, and Anima Anandkumar introduce a topic modeling method that scales linearly to billions of documents.

They focus on two key political traits, agency and communion, and extract these traits from a large corpus of politicians’ speeches. This approach is validated using human-labeled data and functional tests. You can read the paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Measuring Politicians’ Public Personality Traits Using Computational Text Analysis: A Multimethod Feasibility Study for Agency and Communion,” @lukasbirkenmai1.bsky.social and Clemens Lechner introduce an approach to infer politicians’ personality traits from text data.

Nationally Representative, Locally Misaligned: The Biases of Generative Artificial Intelligence in Neighborhood Perception | Political Analysis | Cambridge Core Nationally Representative, Locally Misaligned: The Biases of Generative Artificial Intelligence in Neighborhood Perception

They use an image classification task to compare assessments of GenAI models to a national and locally representative survey sample. Overall, GenAI is biased toward national averages over local perspectives. You can read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Nationally Representative, Locally Misaligned: The Biases of Generative Artificial Intelligence in Neighborhood Perception,” Paige Bollen, @joehigton.bsky.social, and @msands.bsky.social test which populations Generative AI is most representative of.

Survey Professionalism: New Evidence from Web Browsing Data | Political Analysis | Cambridge Core Survey Professionalism: New Evidence from Web Browsing Data

They find that survey professionalism is common, but there is limited evidence that survey professionals lower data quality. Professionals do not systematically differ from non-professionals and don’t exhibit more response instability. Read the paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Survey Professionalism: New Evidence from Web Browsing Data,” Bernhard Clemm von Hohenberg, @tiagoventura.bsky.social, Tiago Ventura, @jonathannagler.bsky.social, @ericka.bric.digital, & Magdalena Wojcieszak provide evidence on survey professionalism across three samples.

Meaning Beyond Numbers: Introducing the Plot Staircase to Measure Graphical Preferences | Political Analysis | Cambridge Core Meaning Beyond Numbers: Introducing the Plot Staircase to Measure Graphical Preferences

The plot staircase is introduced as a way of identifying the relative importance of a graph characteristic compared to a baseline. This method is demonstrated using data on economic growth, job creation, and the COVID-19 vaccine rollout. Read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Meaning Beyond Numbers: Introducing the Plot Staircase to Measure Graphical Preferences,” @talbotmandrews.bsky.social, Justin Curl, and Markus Prior examine how visual characteristics influence preferences. They find that people prefer increasing trends.

Codebook LLMs: Evaluating LLMs as Measurement Tools for Political Science Concepts | Political Analysis | Cambridge Core Codebook LLMs: Evaluating LLMs as Measurement Tools for Political Science Concepts

The authors provide a framework to evaluate codebook-LLM measurement, classifying unlabeled documents with LLMs given a human-written codebook. Ultimately, supervised instruction-tuning can substantially improve performance. Read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Codebook LLMs: Evaluating LLMs as Measurement Tools for Political Science Concepts,” @ahalterman.bsky.social and @katakeith.bsky.social show how “off-the-shelf” LLMs have limitations in faithfully following real-world codebook operationalizations.

A Statistical Model of Bipartite Networks: Application to Cosponsorship in the United States Senate | Political Analysis | Cambridge Core A Statistical Model of Bipartite Networks: Application to Cosponsorship in the United States Senate

Bipartite networks are common in social science, but researchers often project data on unipartite networks for analysis. This new model uncovers patterns across node types, uses covariates to explain ties, and fits efficiently. Read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “A Statistical Model of Bipartite Networks: Application to Cosponsorship in the United States Senate,” @adelineylo.bsky.social, Santiago Olivella, and Kosuke Imai develop a statistical model of bipartite networks and offer an open-source software package for researchers.

GNNs are advantageous because they can be trained, saved, and deployed on new data, and they can also generate synthetic data. The paper uses the militarized international disputes dataset to illustrate potential applications. Read the paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Generative AI and Topological Data Analysis of Longitudinal Panel Data,” Badredine Arfi constructs an approach to analysing longitudinal panel data which combines topological data analysis and generative AI applied to graph neural networks (GNNs).

Probabilistic Record Linkage Using Pretrained Text Embeddings | Political Analysis | Cambridge Core Probabilistic Record Linkage Using Pretrained Text Embeddings

The package is demonstrated using several political examples where researchers may wish to join messy data. The fuzzylink package outperforms existing methods and even allows researchers to link datasets across languages. You can read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Probabilistic Record Linkage Using Pretrained Text Embeddings,” @joeornstein.bsky.social introduces the R package fuzzylink and shows how to incorporate pretrained text embeddings into probabilistic record linkage procedure.

Decomposing Network Influence: Social Influence Regression | Political Analysis | Cambridge Core Decomposing Network Influence: Social Influence Regression

The authors apply the SIR model to data on monthly conflict events between countries, highlighting the model’s ability to illustrate complex influence patterns within networks by linking them to specific covariates. You can read the full paper here: www.cambridge.org/core/journal...

Currently in FirstView: In “Decomposing Network Influence: Social Influence Regression,” Shahryar Minhas and Peter Hoff introduce the social influence regression (SIR) model. The SIR model is for relational data that incorporates exogenous covariates into the estimation of influence patterns.