During the 2016 U.S. presidential election, digital ads aimed at preventing people from voting primarily targeted specific demographics. People who saw these unreleased political ads were less likely to vote than those who didn’t see them. The study, published in PNAS, presents real-world data linking personalized social media messaging and offline voting behavior.
Political campaigns have a history of attempting to demobilize selected segments of the population. This practice is known as voter suppression. This includes targeted strategies aimed at dissuading or preventing opposing population groups from voting.
Historically, voter suppression manifested itself through physical intimidation and harsh localized regulations. In previous eras, tactics included regulatory measures such as poll taxes, strict ID laws, and intentionally confusing information about voting places. These targeted efforts are now increasingly moving into the digital realm. Modern platforms work on customized feed algorithms to ensure messages reach specific individuals.
Advertisers use microtargeting to reach these specific audiences online. They rely on vast amounts of data about users’ interests, geographic location, and demographic background. By packaging this data into consumer categories, social media companies allow political groups to deliver customized messages to very narrow segments of the public.
A subsequent government report revealed in 2016 that Russian operatives were buying ads on the platform using historical search terms related to the African American civil rights movement to find targeted users. Many of these digital strategies operate in regulatory blind spots. This message is frequently sent by private campaigns that do not file financial reports with traditional tax authorities or federal election agencies. These sponsors remain anonymous, allowing misleading election content to spread unchecked across social networks.
It’s very difficult to accurately measure who saw a particular ad and track whether they voted. Most previous studies relied on computer simulations or asked people to self-report their voting history, which can be inaccurate. Young Mi Kim, a media researcher at the University of Wisconsin-Madison, recognized this gap in research.
She collaborated with Ross Dahlke, Hevin Song, and Richard Heinrich to design an observational study to measure direct exposure to anonymous negative campaign ads. The team wanted to know exactly who was receiving these messages. They also sought to assess whether visual exposure was tied to actual turnout at the ballot box.
To monitor ad exposure, researchers asked thousands of volunteers to install a custom digital tracking application. Tracking programs functioned similarly to traditional ad blockers. Instead of blocking promotional content, the program cataloged each ad and its associated data on a secure research server. During the six weeks leading up to the 2016 election, the application recorded all political ads that appeared on participants’ social media feeds.
A major challenge when studying the effects of social media is to account for user choices, often referred to as self-selection bias. When individuals browse free posts on social networks, they actively choose which accounts to follow and interact with. This mechanism makes it difficult to separate existing political beliefs from the influence of new information.
Digital ads work differently because they are delivered solely based on algorithmic targeting and not user subscriptions. People encounter promotional messages simply because a sponsor has paid to put them on their feed. By analyzing these forced exposures, researchers were able to remove self-selection from the equation and add validity to measuring electoral influence.
The researchers also asked participants to complete a questionnaire about their political leanings and demographic background. After the election, the team partnered with an outside data company to link these profile surveys and advertising logs with localized official voting records. This allowed researchers to confirm whether a person actually voted without relying on a person’s memory.
Kim and his colleagues scoured the ads they collected and identified specific forms of voter suppression messages. They looked for content that encouraged election boycotts or promoted third-party candidates primarily aimed at splitting votes. For central statistical analysis, the team isolated tens of thousands of messages sponsored by anonymous entities.
Researchers identified common themes utilized by anonymous sponsors. Campaigns often spread deceptive information about how voting works, including using text messages and social media posts to tell users they can vote from home. These tactics built directly on historic efforts to reduce voter turnout while aligning with modern digital consumption habits.
The research team documented a very specific distribution pattern for these ads. Nonwhite voters living in counties with large racial minority populations within battleground states received a disproportionate amount of negative voting messages. The data showed that these specific demographic and geographic groups were intensively targeted compared to white voters living in less competitive districts.
To estimate the impact on voting behavior, the researchers used a statistical adjustment technique known as entropy balancing. This method creates groups of exposed and unexposed people with closely matched characteristics. By pairing individuals with identical age, income, education, and political ideology, the researchers were able to compare differences in their final voting habits. Since the exposure took place before the election, the timeline ensures that the ad preceded voting behavior.
Across the sample population, exposure to voter suppression advertising was associated with lower turnout. On average, people who saw the ad turned out to vote about 2 percentage points lower than those who didn’t see the message. Several battleground states in 2016 were decided by less than 1 percentage point, and even small shifts in voter participation could change the final election outcome.
The researchers noted that turnout declined even more significantly among the specific groups most heavily tracked by the targeting algorithm. Nonwhite voters living in minority population centers within battleground states experienced the largest drop in turnout after exposure. The targeted subpopulation experienced an approximately 14 percent decrease in voting compared to the subpopulation that did not receive negative election messages. This indicates that the ads had distinct and varied effects depending on the demographic profile of the matched audience.
The researchers tested their data against multiple control groups to validate their findings. The researchers compared the targeted audience to voters who received general political messages and voters who did not see any political ads. The pattern of suppressed turnout was consistent across different groups. The researchers also noted that people exposed to positive political ads had a small increase in overall turnout, highlighting the unique depressing effects of suppressive messages.
This study relies entirely on observational data rather than actively manipulated and randomized experiments. Although the researchers used matching techniques to account for confounding variables such as income and political ideology, theoretically unknown factors could still influence the results. For example, an individual’s community environment can influence their decision to go to the polls on election day. Therefore, the team advises caution when assuming a direct causal relationship between digital advertising and individuals’ voting decisions.
This result is also unique to the political landscape of the 2016 presidential election, as the digital advertising landscape and social media moderation policies continually change with each election cycle. Future observational studies could focus on other election periods to build a more comprehensive understanding of how customized online messages influence local voting habits. The study, “Targeted digital voter suppression efforts are likely to reduce turnout,” was authored by Young Mi Kim, Ross Dahlke, Hyebin Song, and Richard Heinrich.
