Who is Mobilized to Vote by Short Text Messages? Evidence from a Nationwide Field Experiment with Young Voters

Voter mobilization through get-out-the-vote (GOTV) interventions is often successful in increasing voter turnout. However, the effects of mobilization campaigns may not be evenly distributed across the electorate and can potentially increase inequalities in political participation (Enos et al., 2014). As a result, GOTV interventions, even with the best of intentions, may aggravate uneven participation in voting and exacerbate existing disparities in descriptive and substantive representation that are biased against underprivileged citizens (Fowler, 2013; Harjunen et al., 2023).

The effects of GOTV interventions on voter turnout have been examined in numerous experimental studies that have covered various mobilization tactics, explored different means of communication (e.g., door-to-door canvassing, postcard mailings, phone calls and text message reminders) and investigated the potential behavioral mechanisms driving the impact of mobilization campaigns on voter turnout. In a standard randomized controlled trial (RCT), a researcher assigns individuals or households to receive a treatment before the elections and estimates the average treatment effect on voter turnout by comparing the differences in turnout between the treatment and control groups after the elections. Since the early GOTV experiments, randomized controlled trials have proven to be a feasible method to evaluate the effectiveness of GOTV appeals across different electoral contexts, enabling researchers to accumulate a large body of evidence about the impact of various mobilization interventions on voter turnout. However, there still exists limited knowledge how the effects of GOTV interventions may spill over from treated individuals to untreated individuals and how this transmission of voting decisions in social networks may affect the composition of the electorate.

This paper evaluates how GOTV appeals affect inequalities in voting, transmit from treated to untreated individuals within households, and how this transmission of voting decisions through family networks affects inequalities in voting. For this purpose, we conduct a large RCT and evaluate the effectiveness of non-partisan text message reminders on voter turnout in nationwide county elections in Finland. The target population of our experiment are young adults aged between 18 and 29 years, a population group with high levels of education and health in a global comparison, but persistently low turnout rates.¹

The main contribution of this article is the measurement of spillover effects and, more importantly, the examination of heterogeneous spillover effects by voting propensities. First, using electronic voter turnout records and rich individual-level administrative data on eligible voters, we estimate the effect of a large non-partisan text message-based GOTV campaign on social inequalities in voting. Second, using unique household IDs, we investigate how turnout decisions transmit between household members. Finally, we examine how large spillovers from treated individuals to untreated individuals may influence the effects of GOTV mobilization campaigns on the composition of the electorate.

Our estimation of the electoral effects builds upon access to administrative records that contain comprehensive information about individuals’ own and their household members’ demographic background and labor market outcomes. Thus, the linked dataset contains a barrage of potential individual-level characteristics that could be used to explore the heterogeneity of direct and spillover effects among the eligible voters. To keep our statistical analyses conceptually tractable and facilitate the interpretation of results, we organize the estimation of heterogeneous treatment effects into two categories. First, following the paper by Hirvonen et al. (2023), we examine the potential heterogeneity of treatment and spillover effects using data-driven estimation methods. This exploration of heterogeneous effects contributes to the literature on voter mobilization and inequalities in political participation. These results, however, may hold limited practical implications for the development of concrete targeting strategies to ameliorate the stubborn demographic gaps in political participation. Thus, we complement the analysis by estimating heterogeneous treatment effects in various (pre-registered) sub-samples that are created using a single observed characteristic at a time. Through these complementary analyses, we aim to advance the applied research literature that assesses the potential of enhancing the effectiveness GOTV interventions through selective targeting of interventions.

Our paper relates to several strands of literature. First, our study relates to the literature on social inequalities in political participation. A recent literature on the compositional effects of get-out-the-vote mobilization strategies suggests that many current mobilization strategies may widen existing social disparities in voting by predominantly mobilizing high-propensity voters instead of under-represented low-propensity voters (Enos et al., 2014). Simultaneously, there is evidence that the success and compositional effects of GOTV mobilization strategies may vary by the salience of the elections (Arceneaux & Nickerson, 2009). However, to date, there is relatively little evidence on the compositional effects of text message-based mobilization strategies. Our paper complements the existing literature on the compositional effects of GOTV mobilization strategies in several ways. First, and most importantly, we assess how accounting for spillover effects from treated individuals to untreated individuals modifies the effect of GOTV mobilization on the composition of the electorate. Second, we assess the compositional effects of GOTV mobilization in an electoral context where all eligible citizens are automatically registered to vote. Third, we assess the robustness of the prevailing empirical strategy in the relevant literature that estimates baseline voting propensities using within-sample covariates and interacts the predicted propensities to vote with the GOTV treatment indicator. To address the concern that the within-sample estimates of voting propensities may not predict turnout out-of-sample, we estimate citizens’ propensities to vote using machine learning techniques. Consistent with the existing Nordic GOTV interventions (Bhatti et al., 2018; Bergh et al., 2020; Bergh & Christensen, 2024), we find that our intervention mobilized mainly low-propensity voters. More uniquely, we find that the effects are the largest for the citizens belonging to low-propensity groups, but who have high personal interest in voting.

Second, our paper relates to the very few experimental studies on voter mobilization with an explicit objective to measure spillover effects. Prior to our work, Nickerson (2008), Sinclair et al. (2012) and Bhatti et al. (2017a) have investigated how the effects of different get-out-the-vote appeals may transmit from treated to untreated individuals and reported within household spillover effects varying from 30 to 60% of the direct effect. We find that over 100% of the direct treatment effect spilled over to the untreated household members. Moreover, our findings complement the existing literature on the measurement of spillover effects by estimating how spillovers interact with the eligible voters’ predicted propensity to vote and affect the composition of the electorate. We find that also the spillover effects reduce the existing inequalities in voting. These results stress that an inadequate analysis of the spillover effects may not only lead to severely underestimate the effectiveness of GOTV interventions in general, but also to overlook the effects of GOTV mobilizations on the composition of the electorate.

Our results related to the direct effect of the intervention on voter turnout are largely consistent with the existing Nordic literature that has systemically, although with varying magnitudes, documented the effectiveness of SMS reminders on voter mobilization (Bhatti et al., 2017a, b; Bergh et al., 2021; Naess, 2022; Bergh & Christensen, 2024). We find a statistically significant, about 0.9% point, direct average treatment effect in the probability of voting. The effect size of 0.9% points equals 3% increase compared to the control group average turnout of 30.9%. Our evidence stems from a relatively low-salience elections and may not readily generalize to high-salience elections (Malhotra et al., 2011; Green & Gerber, 2019; Bergh & Christensen, 2024; Mann & Haenschen, 2024).

Finally, our study builds on and contributes to the literature that has investigated the effectiveness of numerous voter mobilization strategies and different contents of campaign messages on voter turnout and choice (Green et al., 2013; Green & Gerber, 2019). We find the largest effects for neutral message types.

The paper proceeds as follows. In Sect. “Theoretical Background”, we discuss theoretical arguments regarding the potential effectiveness of the intervention. Section “Background and Context” describes the relevant electoral context. In Sect. “Experimental Design and Data”, we describe our data, experimental design and the sample. Section “Data Analysis” presents our empirical methods. Section “Results” presents the results. Section “Discussion” discusses the results before Sect. “Conclusions” concludes. The paper is accompanied by an online supplement.

The most prominent theory regarding the average treatment effects of text message reminders on turnout is the Noticeable Reminder Theory (Dale & Strauss, 2009; Malhotra et al., 2011). It’s premise is that registered voters tend to have an intention to vote, but may fail to do so, because of time constraints and lack of planning and attention, in which case only a simple nudge (Thaler & Sunstein, 2009) is needed to remind them on their intention to vote, and thus, mobilize them. According to Dale and Strauss (2009), text messages are likely to surpass the threshold of attention, because voters pay attention to their phones and text messages are difficult to ignore. However, in our context, all eligible voters are automatically registered, and thus, this theory does not perfectly apply. Nonetheless, a substantial share of eligible voters are likely to have an intention to vote and yet fail to do so.

The Receive-Accept-Sample theory (Zaller, 1992; Matland & Murray, 2012) can explain why different types of potential voters may be differently affected by the experiment. The theory implies that politically more aware citizens who are more likely to vote are receiving more political messages in general, and therefore, less likely to accept any individual message due to the flood of them. Less aware individuals receive fewer messages in general, and thus, are more likely to accept them. This would imply that the effect of an intervention is larger for the inattentive citizens who are less likely to vote. However, the standard assumption in GOTV experiment literature is that inattentive citizens with little interest in the election are unlikely to take note of such a small nudge as a text message. Thus, the effects are expected to be the strongest among the voter group in the middle of the distribution who pay some but not too much attention to politics. This argument is similar to the proposal by Arceneaux and Nickerson (2009) which suggests that citizens who are nearly indifferent between voting or not voting are the easiest to mobilize.

We can hypothesize about the spillover effects using four related theoretical arguments. First, according to the Social Occasion Theory (Dale & Strauss, 2009), if the text message creates communication within the household it creates a social occasion that may establish feelings of connection to voting. Second, social norms and conformity pressure may lead to similar behaviors within households (Zuckerman, 2005; Foos & de Rooij, 2017). Thus, if a treated family member is affected by the experiment, then other members may conform. Third, information transmitted within households is often of low-costs and citizens may pay more attention to information coming from close social network (Klofstad, 2007). The fourth rationale for spillover effects is that many voters choose to vote at a polling station on the Election Day instead of voting in advance and households often go to vote together. The theories two to four imply that spillover effects can arise only if there is a direct effect, and therefore, the heterogeneity of the spillover effects should be contingent on the heterogeneity of the main effect. However, the Social Occasion Theory allows for the possibility that there is a spillover effect even without the main effect as long as the treatment is discussed among the household members.

We conducted our RCT in Finnish nationwide county elections held on January 23, 2022. Counties are the mid-tier level of decision-making in Finland between the municipalities and the central government. They resulted from a recent large social and healthcare reform. Thus, the elections were the first of their kind in Finland. The elections were expected to be of low salience and interest. This expectation also turned out to be true as turnout in the elections was 47.5%, which is lower than in any parliamentary elections in the Finnish history.

The allocation of seats in the county elections is proportional to the votes following d’Hondt system of open party list proportional representation (PR) and identical to the Finnish parliamentary and municipal elections. Finland uses a very pure form of open-lists in the sense that personal vote is obligatory: each voter gives exactly one vote to one candidate. Parties are assigned seats based on the sum of its candidates’ personal votes and the seats within the party are assigned purely based on the personal votes. Moreover, candidates are almost always presented in alphabetical order in the ballot lists limiting parties ability to signal their preferences over the candidates. Overall, the open list electoral system in Finland is highly personalised, which may increase incentives for individual campaigning compared to several democracies with closed list PR or mixed electoral systems (von Schoultz & Strandberg, 2024).²

Voters are automatically registered in all elections in Finland. An electronic register of all eligible voters (voting register) is established based on the Population Information System on the 46th day before the election day (Jääskeläinen, 2020). All voters listed in the voting register receive a notice of their right to vote (polling card) no later than 24 days before the election day. The polling card indicates the date of the election, the period for advance voting, the locations of advance polling stations within the voter’s electoral district, the address of the voter’s election day polling station, and contact information of the electoral authorities. The polling stations have only an administrative role as the elections are held at-large in the whole county. A typical characteristic of the Finnish elections is that a relatively large share of voters cast their ballots at polling stations during the period for advance voting that begins 11 days before the election day and ends 5 days before the actual election day. In the 2022 county elections, 57% of individuals who voted used the advance period to cast their vote.

Prior to our study, text message-based mobilization experiments have been conducted in the US, Denmark and Norway. The Finnish electoral system and voter mobilization environment closely resembles the other Nordic countries. Turnout in Finnish local and regional elections is typically markedly higher than in the local US elections, but has been in many recent elections noticeably lower than in comparable Danish and Norwegian elections (Bhatti et al., 2017b; Bergh et al., 2021). In the 2021 municipal elections, the turnout of eligible voters was 55.1%. There are notable demographic inequalities in voting. Young adults aged from 18 to 29 years are markedly less likely to vote than the older age cohorts. Their turnout in the 2021 municipal elections was 36.6%. The gender gap among young voters in the 2021 municipal elections was 8% points as women had a turnout of 40.7% and men had a turnout of 32.7%.³

Voters’ access to information on party platforms and individual candidates is supported through wide-ranging public information campaigns and strong public media presence. Political campaigning and advertising is regulated by the Election and Data Protection Acts that restrict the use of personalized advertising using direct mailings, phone calls and text messages. To our knowledge, prior to this study, there has not been politically motivated or government sponsored non-partisan text-message campaigns to mobilize voters in Finland.

Using the RCT design and rich individual-level data on eligible voters and their household members, we conduct three different types of empirical analyses. The following section describes how these data analyses were carried out in our study. We note that we registered a pre-analysis plan (PAP) for our study by completing the registration fields in the American Economic Association’s RCT Registry before implementing the study (Banerjee et al., 2020). Our PAP serves as a concise record of our initial intentions about the sample size, key outcomes and planned statistical analyses. However, our PAP did not provide a detailed description of the statistical analyses included in this article and did not specify in advance how the results are presented. Consequently, the statistical analyses presented in this paper do not follow a detailed PAP and should be primarily treated as empirical analyses that were conducted without a pre-registered statistical analysis plan.

This paper employs four different estimation methods to detect heterogeneous treatment effects. First, we split the sample in three vote propensity groups using a logit model. Second, we reproduce the same procedure using an elastic net model. We find that predicting the voting propensities using a logit model led to the highest point estimate for the low propensity group. In addition, we find that predicting the voting propensities using the elastic net method led to the highest point estimate for the marginal propensity group. Both of these two complementary methods yielded for the high propensity voters estimates that were close to zero and not statistically significant. The differences arise from using municipality fixed effects, which improve voting prediction.¹⁹ Around 83% of the sample remains in the same vote propensity group regardless of which of the two prediction models is used. The difference between the prediction models is due to the LASSO regularization part of the elastic net dropping some of the municipality fixed effects by setting the corresponding fixed effect coefficients equal to zero. Without the municipality fixed effects, there are practically no differences between these two methods as over 99.99% of observations are assigned to the same vote propensity groups.

A mere comparison of the point estimates might not give the full picture about relative effects as the number of individuals differ by vote propensity groups, and direct effect and spillover samples are not of equal size. Thus, we define the number of activated voters as: \(\beta _{1}(Treatment) \times N_{treated\_individuals}\) in each group. Online Appendix (Table A18) presents the number of activated voters by vote propensity groups for direct effect and spillover effect samples, and by age for the spillover sample using the logit estimates. When we account only for the direct effect, the experiment resulted in an increase of around 150 voters in the low propensity group and 177 voters in the group of marginal voters. As the heterogeneity estimates are remarkably similar to those from direct effect sample with logit vote propensity groups and elastic net vote propensity groups, it means that voting inequality was not just reduced by the direct effect but also among individuals subject to the spillovers.²⁰ In terms of activated voters, the experiment resulted in 116 more votes from the low propensity and 177 votes from the marginal voters spillover groups. Overall, accounting for the spillover effects more than doubles the total number of activated voters from 268 to 557 voters. Additionally, spillovers have an implication on how the voting gap between the target population of youth voters and older voters evolved. When also the within household spillovers are taken into account, the experiment activated 90 more under 30-year-old voters compared to over 29-year-old voters. This is a sizable relative reduction from the net of 268 activated youth voters when spillovers are ignored. However, it should be noted that the experiment differs from a policy where every youth voter is reached by SMS, as for that policy there is no possibility for spillovers to under 30-year-old voters who did not receive a voting reminder. Additionally, we are not able to determine the precise effect for the voting gap between youth and old voters in relative terms, as the spillover group is not a random sample from the general population of over 29-year-old eligible voters.

Third, we utilize a machine learning method called Honest Causal Forest to detect heterogeneous treatment effects. While this algorithm has nice properties in terms of reducing the freedom of a researcher when specifying a model, in our case, we are unable to draw conclusions from the results as estimates obtained from the model are very imprecise. This is unsurprising as even though the number of CATE groups is the same as the number of vote propensity groups, the effective number of observations is lower as only part of the sample is used for the estimation due to sample splitting and folding.

Lastly, we use sub-group analysis where we split the sample by a set of covariates. At first glance, the higher estimate for the group that had voted in 2021 municipality elections could be seen as contradictory to our vote propensity group results coming from low and mid propensity voters. How can we reconcile these contrasting conclusions about the effects of our intervention on inequality in voting? In Online Appendix (Table A17), we split the three voting propensity groups by the past voting variable and provide effectiveness estimates for all the resulting six groups. The highest point estimates of 6.2 p.p. and 3.0 p.p. (both statistically different from zero at 1% error level) are, respectively, for the low propensity and marginal voters groups who also voted in 2021. The control group mean is remarkably higher for these groups compared to the individuals belonging to the same vote propensity group but who did not vote in 2021. Also these two latter groups have estimates which are positive and statistically significant, albeit smaller. However, for high propensity voters we do not find statistically significant estimates for either those who voted in 2021 nor did not vote in 2021. These results indicate that among the low and mid propensity groups the treatment effect is largely coming from individuals who tend to vote. Therefore, the effects are coming from citizens who clearly have intention to vote, but whose characteristics would indicate that they are unlikely to vote. This is consistent with the Noticeable Reminder Theory (Dale & Strauss, 2009) that the effect comes from citizens who have an intention to vote, but may fail to do so, because of lack of attention for example, in which case only a simple nudge is enough to remind them of their intention to vote. It is also consistent with the The Receive-Accept-Sample Theory (Zaller, 1992) as citizens belonging to this social group may not receive too many messages, and thus, accepting the message is particularly likely for the voters who have high individual interest in voting, but live in low interest environment. Taken together, this combination of high interest but low exposure could make nudging more efficient.

This paper presents new evidence about the spillover and compositional effects of GOTV interventions. We contribute new evidence to the discussion on inequalities in political participation. Moreover, using an RCT design and data-driven estimation techniques, we provide new insights to the debate on the transmission of spillover effects in voting. We obtain several empirical findings. First, we find that receiving a text message reminder before the Finnish country elections in 2022 mobilized low-propensity and marginal voters and reduced, or at least did not exacerbate, existing social inequalities in voting within our target sample, 18 to 29-year-old voters. Second, the effects are dominantly coming from citizens belonging to low vote propensity social strata, but who have high personal intention to vote. Third, we document remarkably large spillover effects in voting behavior, suggesting that the behavior of adult children with voting rights may influence especially their parents’ turnout decisions. Fourth, the spillover effects reduced inequalities among the group affected by the spillovers. Fifth, we find suggestive evidence that the most simplified phrasing of messages merely reminding recipients about the approaching elections was more effective than the messages appealing to expressive and instrumental motivations to vote. One potential explanation for this finding is aversion to paternalism among the citizens.

Our paper is complimentary to studies that have previously examined the compositional effects of GOTV interventions and transmission of voting behavior in social networks. We document that implicitly assuming zero or little spillovers among social ties may underestimate the true effectiveness of voter mobilization interventions. Moreover, our results show that the average estimates of spillover effects may mask considerable heterogeneity in the peer-to-peer transmission of voting behavior, suggesting that studies that do not account for spillover effects may misreport the compositional effects of GOTV interventions. Our observation that the spillovers occur mainly within households and run from young adults to their parents speaks to the potential trickle-up effects in voting between the generations and may help understanding the role of trickle-up effects in political socialization (Jennings & Niemi, 1981; Linimon & Joslyn, 2002; Dahlgaard, 2018).

Our results suggest that the previously observed compositional effects of GOTV interventions that have widened the disparities in participation do not readily generalize to our context. In fact, we observe that SMS reminders are effective at mobilizing low-propensity voters and their household members. This observation is consistent with the existing Nordic studies reporting that GOTV interventions may mostly mobilize low-propensity voters (Bhatti et al., 2018; Bergh et al., 2020; Bergh & Christensen, 2024). However, unlike the previous European experiments, our experiment is conducted in a relatively low-salience elections, suggesting that the discrepancy between the findings from the US-based experiments (Enos et al., 2014) and European experiments is not explained solely by the differences in the average turnout in the electorate. One potential explanation for the difference is that Nordic countries have high levels of trust in government and public institutions. This mechanism should be evaluated in further studies. Overall, our results hold promise that impersonal but inclusive means of communication, like text messages, may not only successfully raise aggregate voter turnout but also encourage less likely voters to turn out their vote.

Our results stem from an electoral context that is characterized by notably high trust in public administration. Moreover, it is worthy of highlighting that our intervention was developed in collaboration with the Finnish electoral authority and we sent the messages on behalf of the Ministry of Justice. Thus, our project can be classified as a state-led intervention that could produce different results in other countries and contexts with different levels of trust in public administration. Relatedly, there is accumulating evidence that text messaging by public authorities may lead to larger effects in voter registration than messages by advocacy organizations (Cheng-Matsuno et al., 2023). Overall, it is important to recognize that the results from our state-led intervention conduced in a Nordic country may not readily generalize to institutional contexts with different demographic composition of voters and mobilization campaigns organized by non-governmental organizations.

More generally, our paper advances the literature that examines how different sub-populations respond to a given treatment and assesses the potential of enhancing the effectiveness of behavioral interventions through selective targeting of existing interventions. We believe that the blend of RCT designs, comprehensive individual-level administrative datasets and suitable high-resolution predictive methods constitute a promising approach to enhance the effectiveness of interventions aiming to motivate behavioral change. Moreover, we observe that the combination of prediction models and univariate sample splits may help form a more detailed picture about the activated voters.

Our results raise new directions for future research. Our results hold promise that text message-based interventions may not only successfully raise turnout but also reduce inequalities in voting. A natural step towards better understanding the promises and limits of GOTV interventions as a tool to ameliorate demographic gaps in political participation is to study hard-to-reach populations who may be beyond the reach of conventional GOTV interventions, but are accessible through their mobile phones. Attempts to address the minuscule political participation in certain hard-to-reach populations, like young immigrants, may also substantially benefit from the tailoring of treatment designs (e.g., use of their native language) to these specific subgroups.