Quantifying the Potential Persuasive Returns to Political Microtargeting

Ben M. Tappin; Chloe Wittenberg; Luke Hewitt; Adam Berinsky; David Rand

doi:10.1073/pnas.2216261120

In brief

Using randomized survey experiments on U.S. policy issues, the authors test whether political microtargeting—using machine learning to match campaign messages to individual audiences—confers persuasive advantages over simpler strategies. In policy-focused campaigns (influencing opinion on a single issue), microtargeting outperformed alternative strategies by ~70%, but this advantage required only simple demographic targeting and was driven primarily by one policy issue. In multi-issue campaigns, the advantage was much more limited, suggesting context-dependent returns.

Contributions

Direct empirical test of microtargeting's persuasive advantage: Most prior work on political persuasion either lacks direct comparison of microtargeting to alternatives or examines very small message samples (≤5), limiting conclusions about treatment effect heterogeneity.
Large-scale message experiments: Calibration phase uses 26 distinct messages (immigration) and 10 messages (UBI); experimental phase randomizes 5,284 U.S. adults across three messaging strategies.
Simple targeting nearly as effective as complex models: No evidence that targeting by multiple covariates yields larger persuasive gains than single-covariate targeting; basic demographic/ideological targeting (e.g., by party alone) achieves ~70% of microtargeting's advantage.
Context dependence: Advantage substantial in policy-centered campaigns but limited in multi-issue issue advocacy, suggesting applicability varies by campaign goal.

Method

Two-phase study design:

Calibration Phase: The authors analyze existing survey-experimental data on persuasive messages (U.S. Citizenship Act, UBI, and 36 diverse policy issues). For each dataset, they use generalized random forests to estimate which messages are most persuasive for different demographic/psychological subgroups. They compare three strategies:

Microtargeting: Assign each individual the message predicted to be most persuasive to them, based on their demographics, ideology, and moral foundations.
Single-Best-Message (SBM): Show all individuals the message with the largest average treatment effect in the calibration sample.
Naïve: Show individuals a randomly selected message, with no message pretesting.

Experimental Phase: New survey (n=5,284, quota-matched to U.S. demographics) in which respondents are randomized to control or one of the three treatment conditions for each of three policy modules (U.S. Citizenship Act, UBI, multi-issue). Within microtargeting, respondents are sub-randomized to messages targeted by 1 covariate, 2–3 covariates, or 4+ covariates to test whether complexity increases effectiveness. Outcomes measured via validated policy-support items.

Results

Study 1 (Policy-Centered):

On immigration: Microtargeting = 5.17 pp; SBM = 2.41 pp; Naïve = −0.96 pp
On UBI: Microtargeting = 7.16 pp; SBM = 5.03 pp; Naïve = 5.70 pp
Combined (precision-weighted): Microtargeting = 5.96 pp; SBM = 3.48 pp; Naïve = 1.79 pp
Microtargeting ~70% larger than SBM (p=0.004) and >200% larger than Naive (p<0.001)
Key limitation: No evidence that targeting by more covariates increases persuasive impact; single-covariate targeting nearly as effective

Study 2 (Multi-Issue):

Microtargeting = 12.30 pp (vs. control showing no message)
SBM = 15.42 pp; Naïve = 6.19 pp
Microtargeting directionally (but not significantly) smaller than SBM (p=0.11)
Advantage over naive persists (p=0.008) but strategy-by-issue interaction implies finding which issue to target offers less benefit than in Study 1

Notes

Strengths:

Pre-registered design with rigorous two-phase structure (calibration on existing data, out-of-sample validation on new experiment)
Large message sample and substantial n, enabling reliable detection of heterogeneity
Clear policy relevance: provides empirical grounding for ongoing regulatory debates around data-driven political advertising
Transparent about assumptions: microtargeting strategy is plausible but may exceed current campaign sophistication

Limitations:

Survey rather than field experiment: respondents engage with content they might ignore in real-world settings; real ads achieve only ~50% view-through rates
Limited issue diversity (two low-salience issues in Study 1); generalizability to high-salience or contested issues unclear
Assumes high-quality individual-level demographic/psychological data; real campaigns often lack such precision
Cost-effectiveness not assessed: SBM may deliver greater persuasion per dollar spent even if microtargeting has larger absolute effects

Open questions:

Why does Study 2 (multi-issue) show a different pattern? Authors hypothesize people are most persuadable on low-information issues; messages targeted to such issues may not vary much by audience, reducing targeting benefit.
Do the findings replicate for high-salience, contested issues (e.g., candidate preferences during campaigns)?
How do social-media platform algorithms affect microtargeting returns in practice?

Bearing on misinformation: While the paper focuses on legitimate political advertising, the findings inform understanding of how tailored messaging could be weaponized for propaganda or disinformation. The finding that simple demographic targeting nearly matches complex data-driven targeting suggests that even low-resource actors could potentially exploit heterogeneity.