Natural Experiment in Politics

The generic natural-experiment concept covers as-if random variation across all of social science — economics, epidemiology, sociology. This entry concentrates on the political-science applications and conventions: close-election regression-discontinuity designs for incumbency and electoral effects, lotteries and boundaries used to study representation and participation, and the design-based emphasis Dunning brought to the field. The identification logic is shared with other disciplines, but the running examples, threshold designs, and substantive questions here are those of political behavior and institutions.

In a randomized experiment the researcher controls and knows the assignment mechanism, guaranteeing that treatment is independent of potential outcomes by construction. In a natural experiment the researcher does not manipulate assignment; instead, an external process is argued to assign treatment as if at random. That as-if claim is an assumption that must be defended with evidence — balance tests, knowledge of the process, manipulation checks — rather than guaranteed by design, which makes natural experiments more vulnerable to hidden selection than true experiments.

When an election is decided by a tiny margin, the outcome is dominated by factors outside any candidate's precise control — turnout noise, weather, counting error — so which side wins is effectively a coin flip. Candidates who barely win and barely lose are therefore comparable on pretreatment characteristics, and the regression-discontinuity logic compares outcomes just above and just below the winning threshold to estimate the effect of holding office. The design is valid only if candidates cannot precisely manipulate their vote share near the threshold, which is checked empirically.