Many people have criticized media polls this cycle because the respondents included in (often tiny) samples of primary voters don’t reflect the people who will actually vote in a primary. We think the answer to this problem is two-fold: 1) Polling elections using a voter file, and 2) using turnout models to ensure that polling samples reflect the right mix of likely and unlikely voters for a specific election.
We had the chance to put this theory to the test in South Carolina. The two nights before the primary, we interviewed 935 South Carolina Republican primary voters, and then gave the results to the Huffington Post for publication after the vote. Not only were the results extremely accurate — getting Donald Trump’s margin right within 1% and seeing a razor-thin edge for Marco Rubio over Ted Cruz — but we learned a lot about the dynamics in South Carolina and moving forward that we’ve shared in a post on Medium.
Warning: We didn’t use a traditional likely voter model. Rather than only asking voters if they planned to vote, we looked at their actual behavior in the past, and built different scenarios accordingly, projecting who was likely to win and by how much at 2012’s primary turnout of around 603,000 all the way up to 800,000. We were also able to see who benefits from turnout going up or down.
— Patrick Ruffini at Echelon Insights