The opportunities and limits of microtargeting cross-pressured voters

Introduction

Modern American political campaigns invest substantial resources in their efforts to communicate with voters. Campaigns often selectively target individuals with tailored messaging based on the campaigns’ perceptions regarding the individual’s likelihood of voting in the election and their likelihood of supporting each candidate. Such messages are often used to persuade undecided voters and mobilize supporters (as is more common outside the US). Campaigns can then tailor the type (mobilization or persuasion) and content of the message (e.g., issues, values, and images) to individuals identified by the campaign for outreach. This individual-level contact strategy is often referred to as microtargeting, and its effectiveness is dependent upon the accuracy of the predictions regarding where each voter stands on the policy issues and their receptivity to the campaign message. Thus, in this paper, we evaluate both which types of issues and which types of registered voters campaigns are more successful at accurately identifying. We do so by comparing individuals’ self-reported policy positions from an original survey of registered voters (N = 397) to the modeled position available in a voter file from a leading American commercial vendor.

Microtargeting is predicated on the premise that the effect of a persuasive message will be stronger if sent to the right recipient. The right recipient, for political campaigns, is often one who is likely to swing from another party to theirs, a tendency strongly related to their policy positions (Hare and Kutsuris 2023). For this reason, campaigns often seek out voters who are cross-pressured—those who identify with one party but agree with policy position(s) taken by another party. These opposing forces render voters more susceptible to persuasive messages (Hillygus and Shields 2008; RePass 1971). Microtargeted messages can affect voting behavior when campaigns successfully deliver them to cross-pressured voters on the conflicted issue (Endres and Panagopoulos 2019).

To benefit from cross-pressured voters, political campaigns need to identify who those voters are. Some campaigns rely on microtargeting tools provided by social media platforms (Kruikemeier et al., 2022), which may often miss their intended audience (Vrielink et al., 2025). Others utilize datasets sold by commercial vendors, which use publicly available administrative data, such as voter files, combined with other publicly available information (e.g., hunting and fishing licenses), consumer records (e.g., magazine subscriptions), and data about their neighborhood and community from the Census. These data are then fed into proprietary algorithms to predict voters’ party identification and positions on policy issues (Hersh 2015). These predicted policy positions can then be used to target voters who are believed to be receptive to the campaign communication. The accuracy of microtargeted policy positions can range widely and vary by state of residence, as models are influenced by the types of information that states collect about voters, especially party registration and race (Endres 2016). Nevertheless, politicians who send voters policy-congruent messages can affect voting behavior (Endres 2020; Lavigne 2021). The returns to microtargeting cross-pressured voters are potentially large (Tappin et al. 2023), but they depend on accurately predicting voters’ policy positions.

Even if accurate, the value of microtargeted advertising remains unclear. First, online advertising does not appear to be a great boon to political candidates seeking greater recognition or favorability (Broockman and Green 2014), although it may make issues more salient (Vrielink et al., 2025). Messages sent by voters’ preferred party may make them feel more favorable toward the party, but messages sent based on the issue’s relevance to voters’ geographic area (termed “issue fit”) have no similar effect (Hirsch et al., 2024a). Second, people may dislike being microtargeted. Targeting based on personal information may appear manipulative (Hirsch et al., 2024a) and feel like an invasion of privacy (Gahn 2025). Third, research finds scant support for certain forms of heterogeneous treatment effects of campaign messages (Coppock et al. 2020; Hackenburg and Margetts 2024). Similarly, messages matched to voters’ social groups have been found to be no more effective than general messages (Hersh and Schaffner 2013). Still, innumerable ways of targeting voters exist that may yet yield persuasive payoffs. For instance, messages that match voters’ personalities (such as openness and extraversion) (Decker and Krämer 2023) or demographics (Tappin et al. 2023) can have significant persuasive effects. Simchon et al. (2024), using online behavior data to predict people’s degree of openness, employed a large language model to craft messages to appeal to those personality types. As political campaigns become more technologically savvy and incorporate the ever-growing availability of online user data, the feasibility and promise of personalized messages sent to individuals based on key characteristics will only grow.

Prior work has examined the accuracy of vendors’ predictions of voters’ demographics, which may also influence contact decisions and message content or imagery (Hersh 2015; Morris and Miller 2025). Our focus is on policy positions. To contribute to our understanding of the potential and pitfalls of such policy-based microtargeting, we verify a set of predicted policy positions provided by a top US commercial vendor. Using an original 2017 survey (N = 397) of the individuals in the vendor’s dataset, we test four hypotheses and conduct exploratory analysis. First, we anticipate that cross-pressured voters are more difficult to predict, largely because they hold policy positions that are out of step with their party. Since party affiliation is likely a key contributor to policy-prediction models, such misalignment may contribute to more prediction errors (Dancey and Geoffrey 2013). Party-misaligned preference profiles are likely to exhibit more idiosyncrasy, making prediction harder. If party-misaligned preferences are unrelated to the other information available to vendors, such as age or the income or racial makeup of the area in which the voter lives, then making accurate predictions may be especially difficult.

Second, we expect the accuracy of predictions to be shaped by the information that states make public. Specifically, individuals who live in states that require party registration for voting in party primaries should have more accurate policy-position predictions (Endres 2016). Because policy positions are influenced by parties’ position-taking (Freeder et al. 2019; Lenz 2012) and vice versa (O’Brian 2024), the availability of party registration should augment the performance of the vendor’s predictive model for those voters.

Third, we expect strength of partisanship and interest in politics to help align voters’ policy positions to those of their parties (Freeder et al. 2019). Voters who identify more strongly with their party should be more likely to share their party’s views and should therefore be easier to predict. Similarly, voters who are more interested in politics should be more likely to receive their parties’ messages and follow their stances.

Mirroring previous work, we find the accuracy of vendors’ predicted policy positions varies considerably from issue to issue. Across the issues we verified, the average respondent’s positions were predicted with 60.6% accuracy. The accuracy of the vendor’s predictions is closely related to the degree of cross-pressure exhibited on the issue. At the policy-issue level, a regression of the proportion of accurately predicted respondents on the proportion of cross-pressured respondents is approximately −1.0, implying that a percentage point rise in cross-pressure corresponds to a percentage-point decrease in accuracy. At the individual-respondent level, we find higher accuracy for voters who live in states with closed primaries. We also find that voters who care more about politics—both in the form of stronger attachment to a political party and stated interest in politics—are easier to predict. Finally, we examine whether the vendor tends to forgo making predictions (most issues have some missing predictions) for more cross-pressured issues. We find no evidence that they do. Our findings highlight a key challenge in executing successful, efficient microtargeting. The voters most likely to be persuadable are more difficult to predict.

Data

We analyze a dataset of observations of the same people from two sources (N = 397). The first source is a commercial vendor used by political campaigns in the US. Their portion of the dataset contains predicted positions on a set of 25 items, most of which touch on specific policies, three ask about affect toward certain social groups, and two relate to more general ideological orientations. We focus our analyses on the 20 policy-specific items. All the policy predictions are binary and evaluative, usually taking values of either “oppose” or “support.” Some, such as the item on environmental regulation, take more descriptive values like “less regulation” and “more protection.” The only nonevaluative item pertains to belief in climate change, and respondents are either “believers” or “non-believers.” For some voters, the vendor did not provide predictions. We code these as missing. See Appendix Table A1 for outcome values and frequencies. The vendor also provided us with predictions for other respondent attributes, such as their income, number of children, and educational attainment, which we do not analyze, although others have (see, e.g., Morris and Miller 2025).

The second set of variables comes from a survey we fielded between March 22 and April 11, 2017, on a random sample of 10,000 voters with email addresses from the vendor’s national voter file. We invited respondents by email, using graduated incentives to encourage participation, starting with no incentive, then offering entry in a random drawing for a $100 Amazon.com gift card, and last providing a $5 gift card. Ethics approval was received from Northeastern University’s IRB. While response rates in the single digits are common, our rate of 3.97% may raise some concerns about selection bias. In Appendix Table C2, we report sample characteristics next to population estimates of registered voters from the 2016 ANES. We find similarities on age, gender, race, and religiosity but differences in favor of more Democratic and politically engaged voters. In one regard, this sample limits our ability to generalize to all registered voters. In another, by over-representing the politically engaged, it may create more favorable conditions for predicting policy positions.

Our survey asked respondents to state their positions on the same 20 policies, matching our items’ response options to those used by the vendor. On most policy items, respondents chose from a four-point scale ranging from “strongly oppose” to “strongly support.” On a handful, matching the question wording to the vendor’s items required other response options, such as thinking whether the Common Core is a “good idea” or a “bad idea.” We coded predictions as correct if the response on our survey matched the prediction made by the vendor and incorrect if not. Policies without a prediction from the vendor were coded as missing.

Missing predictions are fairly common in the vendor’s data. Few respondents have predictions for all 20 issues; the median number of valid predictions per respondent is 9. To deal with this issue, our main individual-level dependent variable is the proportion of accurately predicted policy positions per respondent. Individual-level cross-pressure is the proportion of issues on which the respondent expressed a preference in misalignment with their party identification.

We also aggregate the data to the policy level, treating policies as observations, and calculate accuracy and cross-pressure for each. Accuracy is the proportion of correctly predicted positions for that policy, and cross-pressure is the proportion of respondents who report a cross-pressured view on that issue. We coded policy positions as cross-pressured by first determining which response options were associated with the Democratic and Republican parties. In most cases, the decision was straightforward. Pro-fracking positions were coded as Republican; pro-gay marriage positions were coded as Democratic, for example. Issues on which respondents’ stated position and party identity disagreed were coded as cross-pressured. Two issues, legalizing state casinos and placing labels on genetically modified foods, did not have obvious party positions, so these were excluded from the cross-pressure analyses. See Appendix B for our coding of each item.

We also use explanatory variables from our survey’s portion of the dataset. The main variables are party strength, which is the 7-point party identification variable folded at the midpoint so that higher values indicate stronger Republican or Democratic identity. We also control for age, race (a binary indicator for white), gender, education, party identification, political interest, religious attendance, and the type of party primary held in the respondent’s state. This is a binary variable coded 1 for states that hold closed primaries, in which voters must register with a party to vote in its primary, and 0 otherwise.

Results

Policy-level accuracy

First, we examine the accuracy of the predicted policy positions. Figure 1 plots the accuracy for each policy, with points’ shapes determined by tercile of cross-pressure. The median issue is predicted with 67.3% accuracy and the mean issue with 64.8%. Looking at the specific policies, highly salient and intensely partisan issues tend to be more accurately predicted. This would make sense if party identity were a prime contributor to the prediction model. Attitudes toward the Affordable Care Act (ACA) are most accurately predicted, an issue with near perfect party disagreement in Congress. Similarly, views on building a border wall and the Tea Party are well-predicted. Less salient issues like Citizens United and casinos, and less polarized issues like Social Security, are less well-predicted.OPEN IN VIEWER

Figure 1.

Proportion of accurate policy positions per policy.

In Figure 2, we plot the proportion of respondents cross-pressured on each issue. Values range from 5.3 to 36.4%. On the mean issue, 20.3% of respondents are cross-pressured. Here, we see Social Security and a pathway to citizenship for immigrants among the most cross-pressured issues. The ACA is among the least cross-pressured. To explore the relationship between the two variables, we plot accuracy against cross-pressure in Figure 3 and add a regression line. The relationship is pronounced and negative: the more cross-pressured the issue, the less accurate the predictions. Regressing accuracy on cross-pressure and a control for the number of valid predictions yields a highly significant negative coefficient estimate on cross-pressure: β = −1.08, SE = 0.303, t = −3.57, p = 0.002, and an adjusted R² of 0.36 (see Appendix Table D3). The slope coefficient of approximately −1.0 implies a percentage point increase in cross-pressure corresponds to a percentage-point decrease in accuracy. The model’s high adjusted R² suggests issues’ party-typicality explains deviations in the vendor’s accuracy in predicting them.OPEN IN VIEWER

Figure 2.

Proportion of cross-pressured policy positions per policy.

OPEN IN VIEWER

Figure 3.

Cross-pressured issues are less accurately predicted.

Respondent-level accuracy

Next, we turn from analyzing issues to analyzing respondents. Campaigns may wish to target not only issues on which many voters are cross-pressured but also tailor issue messages to specific voters identified for persuasion or mobilization (Panagopoulos 2016, 2020). Some voters, 21.4% in our sample, are not cross-pressured on any of the issues we asked about, but over half are cross-pressured on 3 or more. We expect the more cross-pressured voters are, the less accurately their positions will be predicted.

Counteracting this, we expect the positions of voters who are more involved in politics and about whom states collect more information to be more accurately predicted. We estimate a regression of the proportion of correctly predicted positions and present the results in Table 1. ¹

OPEN IN VIEWER

Table 1.

Correlates of accurate and missing policy predictions.

	Accuracy	Missingness
Cross-pressured	−0.273 (0.135)*	0.042 (0.105)
Strength of party ID	−0.000 (0.020)	−0.037 (0.013)**
Political interest	0.057 (0.025)*	−0.019 (0.020)
Party ID (7-pt)	−0.001 (0.009)	−0.003 (0.007)
Age	0.000 (0.001)	−0.003 (0.001)***
Female	0.083 (0.032)**	−0.030 (0.026)
White	−0.090 (0.039)*	−0.019 (0.031)
Education	−0.004 (0.014)	−0.023 (0.011)*
Religiosity	−0.007 (0.010)	0.013 (0.008)
Closed-primary state	0.055 (0.032)+	0.031 (0.026)
Adj. R2	0.072	0.088
N	306	355

Note: Estimates of linear regression models of the proportion of respondents’ (1) correctly predicted policy positions and (2) missing predictions. Accuracy model omits independents. Standard errors in parentheses.

^***p < 0.001; ^**p < 0.01; ^*p < 0.05; ⁺p < 0.1.

First, we find that the more cross-pressured the respondents, the fewer correct predictions made about them. Just as with the issue-level correlation, this relationship is strong: going from the least (cross-pressured on 0 issues) to the most cross-pressured (on 70.8% of issues) is associated with a drop in accuracy of 19.4 percentage points. Some individual attributes, however, are associated with higher predictive accuracy, such as interest in politics. Neither party identity nor strength of party identity appears to be. And as expected, the positions of respondents living in states that collect party registration are more accurately predicted, by 5.5 percentage points.

Some unexpected associations surface as well; most notably, predictions for nonwhite and female respondents are more accurate. Their higher accuracy may be driven by specific, perhaps group-related, issues, but that does appear to be the case. In Appendix A.1, we present accuracy plots for men and women and white and nonwhite respondents side-by-side that show the differences occur across the range of issues. Nor are the differences explained by more consistent positions among women or nonwhite respondents, as women and men (M_Women = 0.172, M_Men = 0.196, t = 1.50, p = 0.133) and nonwhites and whites exhibit similar levels of cross-pressure (M_Nonwhites = 0.192, M_Whites = 0.180, t = 0.700, p = 0.488).

Discussion and conclusion

The era of highly tailored and personalized messaging is only beginning, both in political campaigning (Trish 2018; Witzleb and Paterson 2021) and political science (Velez and Liu 2025). Research has shown that identifying individuals who would be most receptive to certain messages—microtargeting—can be more effective than a generic but popular message (Matz et al., 2024; Simchon et al. 2024 but see Hersh and Schaffner 2013; Hackenburg and Margetts 2024). Although more sophisticated microtargeting does not necessarily yield substantial improvements (Tappin et al. 2023), arguments that match individuals’ values or preferences can be persuasive (Joyal-Desmarais et al., 2022). Issue-based cross-pressure is a theory-driven form of microtargeting that may yield results and attract more political practitioners. Its success depends, however, on the ability to identify who those cross-pressured voters are.

In part because they are atypical partisans, cross-pressured voters appear to have policy positions that are more difficult to predict. Based on our analysis of data from a nationally reputable political vendor, the more cross-pressured voters are on an issue, the lower the chances that predictions on that issue are accurate. Similarly, the more cross-pressured a voter is, the more inaccurate the predictions made about them are. Because cross-pressured voters tend to be less interested in politics (Hillygus and Shields 2008), they likely leave behind less behavioral data, such as by voting in primaries and contributing to campaigns, on which to base predictions. While vendors generally have access to the same types of information used to make predictions, it is certainly possible that there is variability across firms in their accuracy both overall and across policy issues. A related limitation of our study is its reliance on voters with a valid email in the vendor’s file. However, it is likely even less is known about registered voters without an email address in the file.

While cross-pressure has a negative effect on predictability, gender and race seem to have a positive one. The generalizability and explanation for these findings are beyond the scope of this paper; however, the greater accuracy across issues suggests vendors’ predictive models may be better at predicting the positions of women and nonwhites. This relates to an important limitation of these results, which are the source and time from which the data derive. These data were collected in collaboration with a single, albeit major, commercial vendor. And they were collected in 2017, before the advent of large language models, other advances in machine learning, and larger troves of voters’ online behavior data available for use in vendors’ prediction models. Lacking a theoretical explanation for why accuracy might differ across demographic groups, we speculate that prediction models from later periods may not show the same patterns. Ever more sophisticated prediction models may, and most likely will, mitigate errors. How much they do so is for future research.

The same may be said of predictions of cross-pressured voters’ policy positions. Based on our 2017 data, we estimated a prediction gap of 19.4 percentage points between the most and least cross-pressured voters. It is possible that improvements in machine learning techniques and cheaper computing power, coupled with richer voter file data, may have significantly shrunk the prediction gap (e.g., Hare and Kutsuris 2023). Studies exploiting more recent voter file data, however, still find missing predictions on variables of interest (Morris and Miller 2025). We speculate that, due to their dissimilarity to partisans and lower interest in politics, cross-pressured voters exhibit greater idiosyncrasy and remain harder to predict (Lauderdale et al., 2018). We note that improvements to predictions could also arise from changes in the parties’ positions, issues’ salience, and other factors. Nevertheless, future research is well-positioned to examine the persistence of the prediction gap.

The difficulties in predicting the persuadable speak to some debates over the democratic implications of microtargeting and open up additional avenues for future research. Scholars of data-driven campaigning have raised concerns about the potential for microtargeted messaging to manipulate opinion (e.g., Kefford et al., 2023). Our findings suggest targeting the right people with the right message is easier said than done. In today’s increasingly polarized electorate, there are likely even fewer opportunities for targeting cross-pressured voters. Moreover, mis-targeting voters may not only be wasteful, it may have unexpected consequences (Hersh and Schaffner 2013; Hirsch et al., 2024b; but see Hernandez et al. 2025). Sending a correctly targeted message after a mistaken one, for instance, may come across as disingenuous or manipulative. Sending many messages may also begin to appear manipulative (Gahn 2025). It is important to note also that cross-pressured voters may be targeted in order to be demobilized. Predictive models may be able to identify voters who are cross-pressured on some issues but do not know it (Kefford et al., 2023). If they learn that their preferences are out of step with their preferred candidate or party, they may feel less inclined to turn out (e.g., Endres 2020; Fowler and Margolis 2014). As microtargeting and personalized messages proliferate, these are possibilities to be cognizant of.

Supplemental material