Record
How the score worksWhat controlled tests found

617 campaign ads, each tested against a control group

Swayable archive · 132 randomized experiments, 2018 and 2020
FavorabilityVote choiceVote choice (yes/no)
+1.05 ptsAverage ad
0.79 ptsSpread between ads
78%Ads above zero
0%Of a new study explained by style
Average ad
+1.1
One spread above
+1.8

Points of vote choice.

0%of the next study explained, fitted on all the others

So every script is a test to run, and the chart below is a prior, not a rule.

helps unclear hurtsVolunteer
Helps · 2
  1. Plain-folks appealVolunteer+0.79 ±0.5
  2. Explicit ask to vote forVolunteer+0.59 ±0.4
Unclear · 26
  1. Veteran as messenger+0.46 ±0.7
  2. Anger+0.35 ±0.4
  3. Contrast+0.31 ±0.4
  4. Negative testimonial+0.23 ±0.3
  5. Somber+0.20 ±0.4
  6. High production valueVolunteer+0.18 ±0.4
  7. Everyday people as messengersVolunteer+0.15 ±0.3
  8. Amateur, phone-shot styleVolunteer+0.14 ±0.3
  9. Personal storyVolunteer+0.10 ±0.4
  10. Politician as messenger+0.06 ±0.4
  11. Pride+0.04 ±0.5
  12. Enthusiasm+0.01 ±0.3
  13. Explains factsVolunteer0.00 ±0.4
  14. Policy facts0.00 ±0.3
  15. Offers a fact the viewer likely hasn't heardVolunteer-0.01 ±0.3
  16. Positive testimonialVolunteer-0.01 ±0.3
  17. Fear-0.01 ±0.6
  18. Positive tone-0.03 ±0.3
  19. Sounds scripted-0.08 ±0.3
  20. Advocates an opinion-0.08 ±0.3
  21. Cites a source for its facts-0.10 ±0.3
  22. Negative tone-0.12 ±0.3
  23. Humor-0.36 ±0.6
  24. Celebrity as messenger-0.42 ±0.9
  25. Animation-0.59 ±0.7
  26. Text on screenVolunteer-0.69 ±0.7

Points of vote choice, ads with the feature against ads without, with the 95% interval. Random-effects meta-regression, one feature at a time. Hewitt, Broockman, Coppock, Tappin, Slezak, Coffman, Lubin, Hamidian. How Experiments Help Campaigns Persuade Voters. APSR 2024. Dataverse doi:10.7910/DVN/LBPSSV (CC0).