Unpacking the Unpredictable: Cabinet Politics in Presidential Democracies
Examining how NLP and LLMs can shed light on cabinet responses to stochastic events in Latin American democracies
Dr. Bastián González-Bustamante, Principal Investigator
Project Overview
Focus
Investigating the impact of stochastic events on cabinet stability in presidential systems
Scope
12 Latin American democracies, mid-1970s to early 2020s
Methodology
Utilising machine learning and AI techniques, including LLMs
Theoretical Foundations
Attribute-based Approach
Cabinet characteristics at formation determine duration
Event-based Approach
Stochastic events affect political system and cabinet stability
Presidential Systems
Removing ministers as response to events, optimising support
Research Question
How do various types of stochastic events influence the stability and composition of cabinets in presidential democracies?
Social Protests
Economic Crises
Natural Disasters
Media Scandals
Methodology: Case Selection
Our research examines four key Latin American democracies, selected based on their diverse institutional characteristics and data availability:
Brazil
South America's largest democracy, featuring complex federal system and diverse political landscape
Venezuela
Represents significant institutional changes and varying levels of democratic stability over study period
Costa Rica
Known for consistent democratic governance and stable institutional framework
Mexico
Features transition from dominant-party system to competitive democracy
These cases were selected based on their higher number of monthly observations, diverse levels of governance and institutional (in)stability, and economic diversity
Data Collection and Analysis
Dataset Creation
Novel dataset on ministerial turnover and resignation calls
LLM Application
Use of open-source LLMs to identify stochastic events
Validation
Gold standard creation and measurement validity assessment
Causal Analysis
Combination of survival approach and propensity score methods
Validation Process
Random Sampling
500 media reports per country
Manual Annotation
Several human coders per report
LLM Benchmarking
Compare LLM outputs to the human gold standard
Convergent Validation
Using related cabinet turnover variables
Enlace-Inserción UDP 2025-26