How to Poll Runoff Elections

We present a polling strategy to predict and analyze runoff elections using the 2017 French presidential race as an empirical case. We employ balanced probability sampling based on auxiliary information from past elections to identify a small sample of as few as 20 out of 65,000 polling stations that reflects the national electorate well. We then survey the voters’ candidate evaluations in first-round exit polls at select polling stations. We post-stratify the voter sample to first-round election returns while imputing missing candidate evaluations to emulate campaign learning. Finally, we redistribute the votes for eliminated competitors according to their supporters’ lower-order preferences. The approach yields individual-level predictions which are consistent with standard theories of voting behavior. Our aggregate-level predictions outperform other polls. The strategy is less vulnerable to polling errors, and it is a fraction of the cost compared to common exit polls. Finally, the data produced offer analytic potential beyond its original use for election forecasting.