TO
FROM
DATE
RE

Interested Parties
Global Strategy Group on behalf of Democratic Governors Association
May 17, 2019
Overview of the DGA Late Movement Model

This memo outlines the DGA’s work over the last three years to develop a Late Movement Model to better
understand and predict late shifts in the electorate leading up to election day. In other words, a model that
helps us predict where undecided voters are going to end up, and which voters are likely to defect from their
stated choice in pre-election surveys. The model allows the DGA to adjust pre-election survey toplines to get
a more accurate assessment of how the race is likely to look on Election Day based on projected late
movement. It also allows the DGA to better target individuals who have a higher chance of defecting or
breaking at the end of a campaign. Below, we review how this project came to be, how we have continued to
build out and develop this model every cycle, and our most recent results.
Origination
This project was born out of the surprising results of the 2015 governor’s race in Kentucky. That year, the
polling done by the Jack Conway campaign and the DGA IE, as well as public polling and, as we found out later,
internal Republican polling all showed Jack Conway with a small but stable lead with about a week to go before
Election Day. As we know, Matt Bevin ended up winning the race by about 9 points.
In order to understand this apparent industry-wide polling miss, we quickly rushed into the field with a postelection panel survey where we re-contacted respondents of the campaign and IE’s pre-election polls and
asked how they voted. This survey showed that the polling wasn’t inaccurate, but rather that there had been
significant movement in the last week to 10 days with pre-election undecideds breaking overwhelmingly for
Bevin and a significant number of Conway voters also defecting to the Republican.
2016: Rural/Red States
In 2016, we took the first steps in a multi-year investigation to better understand late movement. Across four
states – Missouri, Indiana, West Virginia, and Montana – we conducted pre- and post-election polling to
understand what voter file data and survey questions, when asked before the election, were likely to predict
late movement among undecideds or late defection from either the Democratic or Republican (or third party)
candidates. For example, as shown below, whether a voter had ever considered voting for another candidate
was an important predictor of defection. Based on these results, we then built a statistical model to predict
the likelihood of each respondent in a survey to defect from their stated pre-election candidate and the likely
direction that late deciders would break based on voter file variables and responses to questions on preelection surveys.

 2017: Virginia
In 2017, we deployed this model, using it to adjust pre-election survey toplines in Virginia to provide an
estimate of the likely vote on Election Day. The model proved to be effective, taking pre-elections survey data
at three different points during the campaign that showed Ralph Northam with leads of anywhere from two
to five points and predicting that late movement would boost that margin up to a 7 to 11-point range – very
close to the final actual results of a 9-point victory.

DGA Late Movement Model 2016-2018

2

 2018: A Diverse Mix
In 2018, we then updated the model with our data collected in Virginia in 2017 and deployed it across several
races. It again proved to be a value add, adjusting polling toplines to predict Election Day margins that ended
up being closer to the actual results in states like Georgia, Nevada, Ohio, Oregon, Maine and New Mexico (for
instance, in Georgia, polling in September and early October showed Stacey Abrams with a 6-point lead, but
when the model adjusted these toplines, it predicted just a 1-point lead, much closer to the eventual margin
of -1). Unlike in 2017, we did not engage in any pre-election polling as part of this specific project, rather the
topline numbers represent the final topline numbers from each campaign.

The graph below demonstrates how the shift occurred in Oregon, where the model predicted that late breakers
would inflate Kate Brown’s margin from 1 to 4 points (close to her actual 6-point win). Our post-election survey
in the state showed that Brown held on to her supporters better than Republican Knute Buehler would and
that she won slightly more voters who were undecided or indicated that they would vote 3 rd-party on preelection surveys.

DGA Late Movement Model 2016-2018

3