As DDHQ prepares for an epic 2020 election season we are continuing to innovate how we deliver election data to our clients and the general public. For the 2019 general elections tonight we’re taking another stab at the heart of election night confusion by providing turnout projections in MS, KY, and select Virginia state senate and house of delegates districts as votes are being counted.
Election night watching used to be straightforward- you wait to see how the votes are looking as they’re counted, and a winner is declared when there are no outstanding votes left that could flip the count. Over time, campaigns and data outfits have innovated on this process to make it as efficient as possible and find out who will win first, relying on modeling and segmentation to know from an early point who is overperforming across counties in a way where their opponent is unlikely to catch up as the night goes on.
So to bring some of that level of analysis to the public, DDHQ is continuing to iterate our Vote Model worksheet (debuted on our previous election night live streams on Twitch, Youtube, Twitter) by providing 3 modeled turnout scenarios for Mississippi, Kentucky and Virginia in which to provide a more accurate analysis on who is winning and why. This Vote Model worksheet is something I have been using since first working in campaign election night war rooms back in 2006 and generally is a tool used by all good data campaign operatives (every operative has their own style/version of one).
Our Vote Model sheets will be going live at 6:00pm EST for Kentucky, Mississippi, and select Virginia Legislative Districts. In Mississippi and Kentucky, we are using turnout models to provide the scenarios and produce the vote goals. In Virginia, we are using the past election specifically as the meter for overperformance or underperformance since we are not tracking statewides.
Some quick facts:
- They are Google Docs and thus open to the public for review and download
- The results populating the sheet are coming directly from the DDHQ API so you can be sure they are the fastest and most accurate
- These are projections and designed to give the reader insight into who is winning (obvious), where they are winning (kind of obvious) and what it will take for them to win (the cool innovative thing)
- The modeled turnout projections we’re done with the data science team from 0ptimus and the brief methodological review is detailed below (our turnout modeling is designed for use by political campaigns and not for journalistic purposes so we’ll continue to refine and modify as we get into 2020 — expect deeper dives into how we do it later on)
- Look to see these turnout models incorporated into our API and client feeds going into the 2020 cycle
- Our turnout modeling uses past like-elections to predict likelihood to turn out for the 2019 general at the individual level. We roll back ages on training to fully account for age dynamics since the last election, and use the most up-to-date L2 voter file to make sure we are capturing new voters.
- We run our modeling using sk-learn pipelines, running classification tree and regression versions of the models and then selecting the best model. Our models regularly achieve 85%+ accuracy in predicting the individuals who turn out to vote in a given general election.
- Our turnout models for KY and MS feed directly into our benchmarks by predicting the county composition of votes in a low, mid, and high turnout scenario. Based on the likelihood for its residents to vote, some counties see their share of the electorate slightly grow or shrink as we move from low turnout to high turnout.
- Our middle scenario is the base prediction for turnout when aggregating turnout scores for all registered voters. The low and high scenarios are adjusted versions of the same, which act as the extreme bounds for realistic turnout in the contest.
- Our vote goals act as statewide and county-by-county “benchmarks” for what each candidate would be expected to receive in a razor-thin victory. If a candidate is outperforming these vote goals by 5% statewide, we would expect them to win by roughly 5%.
- These vote goals are developed by using data from recent elections and adjusting those elections to 50/50 contests. We used a mixture of state contests and the 2016 presidential election to account for both nationalized and local patterns. Where possible, we used elections in which the current candidates last ran (e.g. Hood in the 2015 AG race).
- Since this is a simple approach, there is an unmeasured level of error in these estimates. They are meant to give a guide as you watch election night to gauge who has an advantage given the results thus far, and how big that advantage is.