In 2024-26, a team of IMD researchers analyzed approximately 4,000 real-world ranked ballot elections, including some 2,000 political elections from the U.S., Australia, and Scotland; this is believed to be the largest study of such elections ever performed. The analysis was supplemented by the investigation of millions of synthetic elections generated from survey data in the Cooperative Election Study, creating a robust simulation-based complement to the real-world data.

Election methods that were examined include plurality, plurality with runoff, instant runoff voting (IRV), Bucklin voting, several versions of the Borda count, top-k methods, and various Condorcet (completion) methods.

Phase 1 results: The general finding of the first stage of the research project is that the best performing methods are IRV and Condorcet methods. These kinds of methods are the least likely to be susceptible to various kinds of spoiler effect, are mostly resistant to undesirable forms of strategic voting, and are unlikely to elect “weak” or “fringe” candidates. Though IRV and Condorcet methods can elect different candidates, we found that they agree an overwhelming amount of the time in the real-world elections we sampled. From this point of view, whether IRV or a Condorcet method is used will make little practical difference. However, while it could appear from our synthetic generated elections that Condorcet is slightly “better” than IRV, we believe that IRV has some logistical and pragmatic advantages over Condorcet.

• Read the Phase 1 report.
• Data and visualizations website.
• GitHub repository with datasets and code.

Phase 2 results: The team analyzed approximately 3,000 real ranked-choice elections and 25,000 synthetic elections generated from real ones to study strategic voting for Condorcet methods and IRV. Most Condorcet methods are vulnerable to strategically shortened ballots and artificially low rankings, while IRV has different weaknesses. Two Condorcet methods—the diversity score method with a 5% threshold and the team’s new threshold-free “friendly fire” method—show particularly strong resistance to manipulation.

The project also studied single-peakedness, social utility, and statistical fingerprinting of real elections. Although real voter preferences rarely satisfy perfect single-peakedness, many elections come surprisingly close. Condorcet methods perform well on social utility under idealized assumptions, but their advantage weakens with more realistic voter behavior. Statistical fingerprinting helps assess whether synthetic elections resemble real ones.

• Read the Phase 2 report.
• GitHub repository with datasets and code.