Multilevel Regression and Post-Stratification for Pre-Election Polling: Strategies and Lessons from the 2016 Brexit Referendum and US Presidential Election Seminar

We (1) estimate a joint distribution of demographics over the eligible voting population and electoral geography using census and electoral history data, (2) estimate turnout patterns conditional on those demographics using high quality post-election surveys from previous elections, and (3) estimate vote choice conditional on those demographics using current data on vote intention from YouGov's online panel. This strategy, while computationally intensive, performed well for the Brexit referendum, not only in estimating the national result (50.6% estimated vs 51.9% actual), but also the results in the local authority reporting areas (RMSE of 4.1 percentage points over the 380 estimated reporting areas). A similar evaluation of the US Presidential Election model ( https://today.yougov.com/us-election/ ) will be presented as well.