Measuring Political Uncertainty in Autocracies Using Evidence from China

• In Event: Novel Applications of Machine Learning

Abstract

I use and contribute new developments in anomaly detection that have yet to be applied or applied widely in the social sciences to devise a solution to the open problem of measuring political uncertainty in autocracies. Political uncertainty refers to the lack of confidence with which government and policy outcomes can be predicted. Recently, it has surged to unprecedented levels across the globe. In response, scholars have sought to determine its consequences, and found that political uncertainty has substantial adverse economic effects, including that markets tend to be more volatile and investment depressed. However, the existing research on political uncertainty is constrained by the absence of an objective measure, and overwhelmingly focused on democracies. Without an objective measure, political uncertainty is necessarily measured indirectly; most commonly by proxy via plausibly exogenous competitive elections and newspaper indices that count explicit mentions and related terms. Those measures of political uncertainty exclude autocracies where competition is suppressed, ballot boxes are scarce or stuffed, and newspapers are censored.

At the same time that political uncertainty has surged, so too have autocracies in number and influence. Billions are subject to autocrats, who control resources that rival those found in democracies. To advance research on political uncertainty in autocracies, I develop a measure with evidence from China. I first locate a plausibly exogenous source of variation in political uncertainty, namely China’s formerly institutionalized and routinized leadership transitions. I then exploit the Communist Party’s tight control over the news media to uncover otherwise obscured internal dynamics of those leadership transitions such as when leadership selections are actually made and shared prior to their formal announcement. This entails identifying informative geographic and temporal patterns in Chinese archival news coverage of the top candidates for ``future'' leadership positions using statistical and machine learning-based analyses, including anomaly detection with epidemic change-point estimation via a penalized cost approach that I extend for negative binomial and Poisson models, Bayesian detection methods, and an ensemble approach to mitigate the impact of false positives and false negatives from individual procedures; and text classification using unsupervised natural language processing.

I find that prior to leadership selection, China's local officials campaign for candidates who share factional and regional ties; and that the Communist Party signals its leadership selection in the press before their identities are officially announced. I use that information to estimate the period when China's leadership transitions most closely approximate competitive elections to develop a measure of political uncertainty in autocracies. In addition to providing a tool for advancing research on political uncertainty, this paper offers some more general contributions. For the analysis, I develop code to scrape an archive designed to prevent mechanical collection; and construct a novel database with approximately 1.5 million articles from China's state-run national, provincial, and prefecture-level news media. Additionally, the application of anomaly detection should provide a roadmap for wider use across the social sciences.

Author

• Tamar Zeilberger, London School of Economics