In 2020, Deck was actively used by 1,142 campaigns and organizations to target the right voters, efficiently allocate resources, and track their progress through Election Day.
These campaigns and organizations used Deck to target over 80 million voters. We also gave several partners direct access to our underlying data, which they used to target over 20 million additional voters.
Most of the outreach powered by Deck focused on targeted digital ads, with an estimated reach of 65 million voters (targeted through the Deck web app and our underlying data). We estimate that users gained over 139,000 votes through this outreach, which mostly focused on persuasion.
The next biggest outreach channel powered by Deck was direct voter contact (such as phone calls, texts, and some door knocks). Through our VAN integration, Deck users reached 20 million targeted voters directly. We estimate that this netted 198,000 votes, mostly focused on turnout.
An additional 15 million voters were reached via direct mail programs, which focused on registration and voting by mail. We estimate that these mailers netted an additional 138,000 votes.
We see further evidence of our impact in our users’ outcomes. In districts with results from both 2018 and 2020, the Democratic candidate’s vote share was an average of 5.7% lower in 2020.
However, the vote share for Deck users was only 2.5% lower on average. Deck wasn’t the only factor behind this difference, but it’s worth noting that campaigns using Deck, on average, had less money and less media coverage than campaigns not using Deck. Raw data on this analysis can be found here.
In 2020, we experimented with adding polling data as an input in our forecasts. Our initial analysis shows that this made our forecasts less accurate. Still, it gives us an opportunity to compare public polling and polling commissioned by our users with our forecasts.
Among the 40 Deck users in 2020 whose campaigns had at least one poll (totaling 343 unique polls), Deck’s forecasts had a median absolute error of 3.3% while the polls had a median absolute error of 4.7%.
To assess the accuracy of our scores for individual voters, we worked with YouGov to compare our predictions in North Carolina with their survey responses.
We saw that our support scores had an extremely high area under the curve of 0.96, indicating that our scores’ rank ordering of support for Democratic candidates aligned with actual stated support for these candidates 96% of the time.
More details on this validation work can be found in our support model documentation.