Health Research and Policy


DATE: April 25, 2013
TIME: 1:15 - 3:00 pm
LOCATION: Medical School Office Building, Rm x303
TITLE: Pooling and Smoothing Election Polls with a Multivariate Dynamic Linear Model
SPEAKER: Simon Jackman
Professor, Departments of Polical Science and (by courtesy) Statistics, Stanford

During the 2012 U.S. presidential election campaign I developed a poll-averaging model that produced daily estimates of voting intentions national and state levels, published by Using over 1200 published polls, my model correctly predicted the election outcome in every state and Obama's 332 vote Electoral College tally. I elaborate the various elements of the model: (1) reliance on historical election returns; (2) corrections for house effects; (3) imposing a covariance structure between states and national levels; (4) a multivariate dynamic linear model for state-level voting intentions over the campaign. I report estimates of key parameters of the model (e.g., house effects, the day-to-day rate of change parameter), details as to the forecasting performance of the model and sensitivity to various model assumptions. Collectively, the polling industry underestimated Obama's two-party vote share by about half a percentage point; I examine the sources of this systematic, collective bias in 2012 election polling. Since the model produces estimates of trajectories of voting intentions in every state, I also assess the extent to which "set pieces" of the campaign (the end of the Republican nominating process, the nominating conventions, the debates) and exogenous events (e.g., Hurricane Sandy) appears to have moved voting intentions, and variation across states in the magnitude of responses to these events.

Suggested readings:
West and Harrison. 1997. Bayesian Forecasting and Dynamic Linear Models, Springer.

Simon Jackman. 2009. Bayesian Analysis for the Social Sciences. Wiley. Chapter 9.

Drew Linzer. 2013. Dynamic Bayesian Forecasting of Presidential Elections in the States. Vol. 108. DOI: 10.1080/01621459.2012.737735

Footer Links: