Democracy: Global, Historical Measures Based on Observables
John Gerring @ UT Austin
Daniel Pemstein @ North Dakota State
Svend-Erik Skaaning @ Aarhus University
Daniel Weitzel @ Colorado State
Abstract:
Most crossnational indices of democracy rely centrally on coder judgments, which are susceptible to bias and error, and also require expensive and time-consuming coding by experts. This project lays out an approach to measurement based on observables that does not sacrifice the nuanced quality of subjectively coded democracy indices.
First, we gather data for a wide range of observable outcomes that aim to capture different aspects of the democratic process. Next, we employ a random forest model in which an existing democracy index, Z, is the outcome and factual indicators, Xi, are the predictors. The model that provides the best fit to the outcome is understood as an alternate index, Zi, for that conceptualization of democracy.
Naturally, there is some information loss from Z to Zi. We show that the loss is minimal for indices that ascribe to an electoral model of democracy. Despite information loss, an index based on observables may be advantageous for some purposes. It is free of idiosyncratic coder errors arising from misinformation, slack, biases for or against a regime, or data entry mistakes. It is free of systematic bias that may arise from coders’ inferences about a country’s regime status, e.g., from its recent economic performance, episodes of civil unrest, public policies (right- or left-wing), alliances (e.g., with the West or against the West), and the time-period under review (e.g., historical or contemporary). The data collection procedure and mode of analysis is fully transparent and replicable, and the procedure is cheap to produce and easy to update. For any index, Z, an observable index, Zi, can be generated that offers coverage of all polities with sovereign or semisovereign status, surpassing the sample of any extant index. We show that expansive coverage makes a big difference to our understanding of some causal questions.
Bios:
John Gerring (PhD, University of California at Berkeley, 1993) is Professor of Government at University of Texas at Austin, where his teaching and research centers on methodology and comparative politics. He is co-editor of Strategies for Social Inquiry, a book series at Cambridge University Press, and serves as co-PI of Varieties of Democracy (V-Dem) and the Global Leadership Project (GLP).
Dan Pemstein is a comparative political economist and methodologist who studies democratic institutions. Much of his current research examines challenges that digital networks pose to democracy and develops tools to better measure democratic institutions. He also has an ongoing research program that explores the interplay between legislative behavior, political careers, and party organization. He teaches courses on comparative politics, political economy, global public policy, and research methods.
Svend-Erik Skaaning is professor of political science at Aarhus University, Denmark. His research interests include the conceptualization and measurement of democracy, civil liberties, and the rule of law. His recent publications include the books Democracy (Reflections) and Democratic Stability in Times of Crisis.
Daniel Weitzel is a post-doctoral researcher at the University of Vienna and incoming assistant professor at Colorado State. His research interests focus on comparative politics and quantitative and computational methods.
Summary:
Key question: how do we measure democracy?
Democracy indicators/indexes
Various levels of judgment by expert coders
Cover various time spans and world regions
Scales: ordinal, interval or binary
Reliance on human coding
Limits scope due to funding limitations
Bias due to individual judgment
V-Dem approach: data-driven quantification
Gather data for many observable outcomes that capture democratic processes
Random forest to predict current indices from these observables
Benefits
Free of idiosyncratic coder error
Free of some systematic biases (e.g. more coders are left-wing)
Data collection process is fully transparent, cheap, extensible
Data:
Input:
NeedFeatures that are
Relevant for democracy (institutional aspects)
Observability: easy to collect without much human judgment
Coverage: available for most of world over long time periods (since 1789)
44 measures collected,
13 used for minimal model
Measures of turnover in power
Vote share of top parties
Share of eligible voters, female voters
Are elections being held
Multi-party elections
Years of continuous elections
Easy to collect, require no judgment
They were able to collect data for many more location/years than previously existed
Target:
Primary: Polyarchy scores (25,759 country-year observations)
Additional: Target indices they predict: UDS (continuous), Polity2, Freedom House, BMR (categorical)
Their Random Forest model was very accurate in predicting Polyarchy
High error examples:
East Germany in 1990 (very unusual case)
Nigeria is consistently under-predicted (they predict less democracy)
Looking over time in individual countries, they track trends and changes in circumstances very well, even over changes in government systems
Most important features measure the party turnover rates and cross-party vote share distribution