Ranked Choice Voting (RCV), also known as Instant Runoff Voting, is increasingly adopted in the U.S. and internationally to better capture voter intent. However, RCV's iterative elimination process introduces mathematical complexities and path dependencies. A critical question for election officials and pollsters is: how reliably can the final RCV winner be predicted from a small sample of ballots? This paper investigates this question, revealing a stark contrast between pessimistic theoretical bounds and optimistic empirical performance.
The Theoretical Fragility of RCV
Theoretically, RCV is highly sensitive to small changes in vote counts. Because candidates are eliminated round-by-round based on the fewest first-choice votes, a slight shift in a sample can lead to a different candidate being eliminated early, which in turn cascades into a completely different final winner.
The authors prove that in the worst-case scenario, RCV sampling is remarkably fragile. For any constant $k$, one can construct an election where a sample of $n-k$ votes—nearly the entire population—still results in a 0% probability of correctly identifying the winner. This "sampling paradox" highlights that increasing the sample size does not always monotonically improve prediction accuracy in theoretical constructions.
Despite this fragility, the authors establish a lower bound for the probability of a correct prediction with a sample of size 1, which is at least $1/2^{m-1}$ for $m$ candidates.
Practical Robustness in Real-World Elections
While the theoretical results are discouraging, the empirical analysis of real-world data tells a different story. Using 275 individual elections from sources like PrefLib and the Harvard Dataverse (including data from Alaska, Maine, and Berkeley), the researchers found that RCV is remarkably stable in practice.
In real-world scenarios, RCV is generally the best predictor of its own final outcome. Even with a sample as small as 5% or 10%, the RCV winner can often be identified with high confidence. This suggests that the "knife-edge" conditions required for theoretical failure are rare in actual human voting patterns.
As shown in the empirical results, RCV (represented by black squares) consistently reaches high accuracy faster than or comparable to other voting rules like Borda or Copeland when used as proxies. In many cases, such as the Alaska House of Representatives data, the RCV winner is identified correctly almost immediately as sample size increases.
Methodology: Ensembles and Statistical Cultures
To further improve predictions, the authors explored ensemble learning methods: Summation and Multiplicative Weights (MPW). These methods leverage the "Map of Elections" framework to identify which "statistical culture" (e.g., Mallows or Urn models) a real-world sample most closely resembles. By weighting different voting rules based on their performance within that identified culture, the ensemble can potentially outperform simple RCV sampling.
While these ensemble methods provide gains in synthetic environments—where rules like "Plurality Veto" can sometimes be better proxies for RCV than RCV itself—their advantage on real-world data is modest because RCV is already so robust.
Conclusion
The study demonstrates a significant gap between the mathematical theory of RCV and its practical application. While RCV is theoretically prone to sampling paradoxes and extreme sensitivity, real-world elections exhibit a level of stability that makes early winner prediction highly feasible. For election officials and pollsters, these findings provide reassurance that small samples, such as those from exit polls or early returns, are generally reliable indicators of the final RCV outcome.
Summary updated