This paper reports experiments designed to help resolve longstanding controversies about the comparison of voting rules for multi-candidate elections. By paying subjects conditionally on election outcomes, we create electorates with (publically) known preferences. Each electorate then votes, with or without pre-election polls, under one of three voting rules: plurality, approval and Borda. We find that Condorcet losers occasionally win regardless of the voting rule or presence of polls. Duverger's law (the predominance of two-candidates) appears to hold under plurality voting, but close three-way races often arise under approval voting and Borda rule. Voters do not generally cast votes that match their poll responses, but they do so more often under plurality voting. Polls predict outcomes better under plurality voting and may serve as equilibrium selection signals. Voters usually cast votes that are consistent with some strategic equilibrium. By the end of an election series, most votes are consistent with a single equilibrium, although that equilibrium varies by group and rule.