Support, Turnout, Persuasion Scores: Three models that run a campaign

In brief

A campaign's score on you is really three different models. Turnout can be checked against the public record of who voted, so it is the one that can be graded person by person. Support can only be checked in aggregate, against returns in places and groups, because your ballot is secret. A persuasion score is an experiment's estimate of whether contacting you changes your behaviour, so it says nothing about how you would vote on your own. A high support score is therefore no evidence that targeting you would work, and a persuasion score is a claim about the campaign's action rather than a forecast about you.

How to use this

Ask a vendor which of the three scores you are buying, and what evidence backs each one. For a turnout score, ask how it was validated against the public record of who voted, and build that check into the plan. For a support score, accept that the only check is aggregate, against returns in places and groups, and refuse to read it as a claim about one person. For a persuasion score, insist on the campaign's own randomised test: nothing else grades it. When choosing whom to contact, target on estimated effects rather than predicted outcomes, and if you do use predicted outcomes, avoid those predicted least likely to respond, which is what made a student aid nudge backfire.

What the story is about

A campaign that says it has scored you has usually worked out three separate things. Nickerson and Rogers (2014) describe campaign analysts building individual predictions of whether someone will vote, whether someone supports a candidate or an issue, and whether someone's support would change on being contacted, which they phrase as the chance of "changing their support conditional on being targeted with specific campaign interventions". Those are the turnout, support and persuasion scores. Each answers a different question, and each is checked in a different way.

Turnout is the one score that can be marked against a person, because whether somebody voted is a public record. Rogers and Aida (2014) tested it across seven pre-election surveys, with the vote checked afterwards, in three elections and 29,403 people. They found that "actual voting is more accurately predicted by past voting (from voter file or recalled) than by self-predicted voting". Enamorado and Imai (2019) then linked surveys to a national voter file of over 180 million records and recovered turnout rates close to the real ones. Their conclusion was that "the bias of self-reported turnout originates primarily from overreporting rather than nonresponse". In plain terms, the error left over is mostly people saying they voted when they did not.

Nobody can check a support score against how one person voted, because the ballot is secret. The strongest available answer key is aggregate: large surveys combined with official election returns and district-level demographics, which is what Kuriwaki et al. (2024) do for every US congressional district. That tells a campaign whether a model gets the mix right in a place, not whether it got you right. The survey half of that check is thinner than it used to be. Tyler et al. (2026) show that who answers a survey depends on party, so the surveys feeding these scores are more selected than when the three-score taxonomy was written.

A persuasion score measures something different: not what you will do, but what difference contacting you would make. Hitsch, Misra and Zhang (2024) set out the distinction formally. Indirect methods predict how likely an outcome is and read the effect off that prediction. Direct methods predict the effect itself. In their catalogue-mailing application, run across two campaigns a year apart, every targeting policy they estimated beat targeting nobody or everybody, and the direct methods were worth substantially more money. That application is commercial mail, not a campaign. No public study has checked whether a real campaign's persuasion score picks out the people whose vote actually changes.

The labels that train the support and persuasion scores are survey answers, which makes them the models' softest evidence. Contacting people changes who answers. Bailey, Hopkins and Rogers (2016) ran a field experiment with 56,000 Wisconsin voters in the 2008 presidential election, testing canvassing, phone calls and mail. Canvassing "reduced responsiveness to a follow-up survey among infrequent voters", which the authors call a real behavioural response that could bias estimates of persuasion effects. Self-report is unreliable in patterned ways as well. Ansolabehere and Hersh (2012), the first fifty-state vote validation, found misreporting on other socially desirable items besides voting, party among them.

Athey, Keleher and Spiess (2025) ran a randomised field experiment nudging students to renew their financial aid. Nudging everyone improved early filing by 6.4 percentage points over a 37% base. Targeting students the model expected to be least likely to renew their aid, which the authors note is common in practice, made things worse; targeting intermediate predicted outcomes worked best. The setting is student aid rather than an election, so the reasoning transfers and the effect sizes do not.

Vendors sell all three scores and publish no accuracy figure for any of them. Data Trust's platform page says it aims to help clients win elections with the most accurate data possible and promises unmatched accuracy by sourcing up-to-date data. It then quantifies its inputs instead: 240 data elements, 14 Identification Fields. TargetSmart's data-enhancement page describes matching incomplete lists to its national voter and contact database and adding rich demographic, behavioural, turnout and predictive attributes for a complete record, again with no accuracy figure. Both pages are evidence of what is promised, not of how well the models work. No published study validates a commercial turnout, support or persuasion score against outcomes.

A turnout score can be audited person by person, because the public record of who voted is the thing it is trying to match. A support score can only be checked in aggregate, against returns in places and groups, because your vote is secret. A persuasion score is an experiment's estimate of whether contact would change your behaviour, which makes it a claim about a campaign's action rather than a forecast about you. That is why a high support score is no evidence that targeting you would work.

So what

So the practical conclusion is about what the campaign does with the number. Turnout targeting gets graded by the record eventually, and that check can be built in from the start. Persuasion targeting gets graded by nothing unless the campaign runs its own test. A campaign that treats the three scores as one accuracy story is paying for a claim nobody has checked, and the test worth applying is whether the choice serves the decision being made.

For political parties

The turnout machinery is politically neutral. The strategy built on it is not, and Catalist's What Happened in 2024 shows why. It is the vendor's own post-election analysis, graded from public vote history and precinct results, rather than a peer-reviewed study. Irregular voters, defined there as those who missed at least one of the last four general elections, moved 5 points away from the Democratic candidate between 2020 and 2024. And the window rolls forward: in 2016, a super voter was defined as someone who had voted in 2010 and 2014, while in 2024, a super voter was defined as someone who had voted in 2018 and 2022. An unchanged score describes a different electorate every four years.

The persuasion score is the one a party cannot check, which makes the question of what it tracks unusually important. The score is trained on who responds to contact. The labels are people's own responses; contact changes who responds; and who responds follows party lines. Taken together, those three findings support an argument that the score partly measures who answers surveys rather than who is persuadable. This is an argument, not a measurement. Nobody has published a test of it, so a party should treat it as a question to put to a vendor rather than as an answer.

For government

An audit would start from a baseline: how often a model's guess at a voter's party turns out right, file by file. The nearest published version of that is Pew Research Center's 2018 chapter on political data in voter files. That leaves the disclosure an audit would start from not readily available.

Case studies

Athey, Keleher and Spiess (2025) worked with more than 53,000 college students, sent a randomised nudge to renew student aid, and measured the outcome for everyone in the study. They then compared two rules for choosing whom to nudge: targeting on predicted outcomes, against targeting on estimated effects. Targeting on estimated effects won. The setting is student financial aid, not an election. That is why this article leans on it: the mathematics of choosing on the estimated effect transfers to elections, and the effect sizes do not.

References

Ansolabehere, S. and Hersh, E. (2012) 'Validation: what big data reveal about survey misreporting and the real electorate', Political Analysis, 20(4), pp. 437–459. Available at: https://doi.org/10.1093/pan/mps023 (Accessed: 10 September 2026).

Athey, S., Keleher, N. and Spiess, J. (2025) 'Machine learning who to nudge: causal vs predictive targeting in a field experiment on student financial aid renewal', Journal of Econometrics, 249, 105945. Available at: https://doi.org/10.1016/j.jeconom.2024.105945 (Accessed: 10 September 2026).

Bailey, M.A., Hopkins, D.J. and Rogers, T. (2016) 'Unresponsive and unpersuaded: the unintended consequences of a voter persuasion effort', Political Behavior, 38(3), pp. 713–746. Available at: https://doi.org/10.1007/s11109-016-9338-8 (Accessed: 10 September 2026).

Enamorado, T. and Imai, K. (2019) 'Validating self-reported turnout by linking public opinion surveys with administrative records', Public Opinion Quarterly, 83(4), pp. 723–748. Available at: https://doi.org/10.1093/poq/nfz051 (Accessed: 10 September 2026).

Hitsch, G.J., Misra, S. and Zhang, W.W. (2024) 'Heterogeneous treatment effects and optimal targeting policy evaluation', Quantitative Marketing and Economics, 22(2), pp. 115–168. Available at: https://doi.org/10.1007/s11129-023-09278-5 (Accessed: 10 September 2026).

Kuriwaki, S. et al. (2024) 'The geography of racially polarized voting: calibrating surveys at the district level', American Political Science Review, 118(2), pp. 922–939. Available at: https://doi.org/10.1017/s0003055423000436 (Accessed: 10 September 2026).

Nickerson, D.W. and Rogers, T. (2014) 'Political campaigns and big data', Journal of Economic Perspectives, 28(2), pp. 51–74. Available at: https://doi.org/10.1257/jep.28.2.51 (Accessed: 10 September 2026).

Rogers, T. and Aida, M. (2014) 'Vote self-prediction hardly predicts who will vote, and is (misleadingly) unbiased', American Politics Research, 42(3), pp. 503–528. Available at: https://doi.org/10.1177/1532673x13496453 (Accessed: 10 September 2026).

Tyler, M. et al. (2026) 'Why are surveys struggling to estimate vote shares?', American Journal of Political Science. Advance online publication. Available at: https://doi.org/10.1111/ajps.70051 (Accessed: 10 September 2026).