Early Vote Score Release Notes

Definition

In states where voting by mail and/or early in-person voting rates are high, the score represents the probability that an individual (who votes) will cast a ballot more than 1 week prior to election day, rather than less than 1 week prior to election day. In states where sufficiently good data on vote timing is unavailable, the score represents the probability that an individual (who votes) casts a ballot prior to election day, rather than on election day.

Score Type By State

The Early Vote Score is meant to predict how likely an individual who votes will cast their ballot early,  with higher scores indicating greater likelihood of voting early. However, because early voting laws and data availability vary significantly across states, the specific score threshold used to define ‘early’ differs, depending on geography:

• Timing Definition Used -  probability of voting more than 1 week before election day (among voters): AR, CA, CT, CO, DE, FL, GA, HI, IA, ID, IL, IN, LA, MD, ME, MI, MN, MT, NC, ND, NE, NJ, NM, OH, OR, PA, SC, TX, UT, VA, VT, WA, WI, WY
• Early Vote Binary Definition Used — probability of voting before election day, rather than on election day (among voters): AK, AZ, DC, KS, KY, MA, MO, NH, NV, NY, OK, RI, SD, TN, WV

Note also that 2 states (Alabama and Mississippi) do not have early voting, so no early vote score is available for registered voters in these states.

Note: The Early Vote Score represents the probability of voting early conditional on voting. It predicts when someone votes, not whether they vote. Thus, it should always be paired with an appropriate turnout score — such as the Midterm General Turnout Score — to identify individuals who are both likely to vote and likely to vote early.

Technical Details

Models were trained at the state level using  vote history from the 2022 midterm elections as labels. Each state model is a CatBoost Classifier, and was trained using a sample of 1 million 2022 midterm voters (or the entire universe of 2022 midterm voters in states with smaller populations). Features are drawn from the Atlas by Murmuration dataset, and capture past voting behavior with respect to turnout, vote method, and vote timing.

Models were evaluated on a hold-out test set comprising approximately 5.7 million voters across 39 states with state-specific models. Among individuals with scores in the top 20% of the Early Vote Score, 70% actually voted early in the prior midterm, making them 109% more likely to have voted early than the average voter. At the other end of the distribution, individuals in the bottom 20% of the Early Vote Score are 70% less likely than average to have voted early in the most recent midterm. The overall (population-weighted) AUC across all state models was 0.84, indicating the model's strong ability to discriminate between those who did and did not vote early.

Use Cases

The Early Vote Score is designed to help campaigns and organizations optimize the timing and channel of voter contact. Because the score predicts when a voter votes — not whether they vote — it should always be used in combination with an appropriate turnout score.

The score can be used at both ends of the distribution:

• Persuasion and Early Outreach (high end of score): Voters at the high end of the Early Vote Score are the most likely to cast their ballots weeks before election day. Reaching these voters early - before their ballot is already cast - is critical for persuasion programs, candidate introductions, and issue education. Early vote and mail programs are most cost-effective when targeted to voters who will actually use them.
• Election-Day and Late GOTV (low end of score): Voters at the low end of the Early Vote Score are the most likely to vote on or very close to election day. Targeting this group for election-day GOTV operations, poll monitoring support, and late-cycle direct contact ensures resources are concentrated on voters who have not yet voted and can still be reached.

As a result, the score is useful throughout the entire voter contact calendar — not just in the early vote window. For example, suppose a partner is running a state legislative race that is held concurrently with a midterm election, and wants to optimize their voter contact program across the full election cycle:

• Early mail and canvassing universe (6+ weeks out): High Midterm General Turnout Score [70–100] + High Early Vote Score [65–100] — identifies likely voters who typically cast ballots weeks before election day, ensuring early outreach reaches people before they've already voted.
• Late GOTV and election-day operations universe: High Midterm General Turnout Score [70–100] + Low Early Vote Score [0–35] — identifies likely voters who typically wait until election day, allowing the campaign to concentrate late-cycle resources where they will matter most.

Targeting Table

Note: Because the Early Vote Score measures different underlying behaviors across state groups (>1 week early in timing states; before election day in binary states), the score distributions vary across states. The tables below are separated by state group and should be used as a starting point. Adjust thresholds based on your state's early voting landscape and your desired universe size.

States with Timing Definition Used (Voted More than 1 week prior to Election Day):

AR, CA, CT, CO, DE, FL, GA, HI, IA, ID, IL, IN, LA, MD, ME, MI, MN, MT, NC, ND, NE, NJ, NM, OH, OR, PA, SC, TX, UT, VA, VT, WA, WI, WY

To target the top...	Early Voters - set the Minimum scores at	Non-early voters - set the Minimum scores at
10%	81	2
20%	58	4
30%	43	8
40%	32	13
50%	21	21
60%	13	32
70%	8	43
80%	4	58
90%	2	81
100%	0	100

States with Binary Definition Used (Voted Prior to Election Day):

AK, AZ, DC, KS, KY, MA, MO, NH, NV, NY, OK, RI, SD, TN, WV

To target the top...	Early Voters - set the Minimum scores at	Non-early voters - set the Minimum scores at
10%	84	7
20%	51	10
30%	33	13
40%	25	16
50%	20	20
60%	16	25
70%	13	33
80%	10	51
90%	7	84
100%	0	100