Experimental Design Details
This field experiment consists of randomization of eligible voters into treatment groups, which receives Short Message Service (SMS) reminders about the upcoming elections, and into a control group. The target population of this study is eligible voters in Finnish 2025 county and municipal elections, whose registered mother tongue is either Finnish or Swedish (official languages of Finland), or one of the most common foreign languages in Finland (Albanian, Arabic, Bengali, Chinese, English, Estonian, German, Nepali, Farsi, Polish, Romanian, Russian, Sinhalese, Somali, Spanish, Tagalog, Thai, Turkish, Ukrainian, Urdu, Vietnamese), and are living in municipalities with an electronic voting registry, meaning that the outcome variable is available for these individuals. After combining information from the election registry and phone number availability from Statistics Finland (SF), we are left with 1,546,578 individuals with a cellular phone number. For Finnish and Swedish speakers we randomize these individuals into the treatment group (20% of the total sample), which will receive the reminder in Finnish and Swedish, and into the control group (80% of the total sample). For foreign language speakers we randomize individuals into three different groups (each 33% of the total sample); treatment where they receive reminder in Finnish and Swedish, treatment where they receive message in their own language and Finnish, and a control group. Treated individuals will receive a SMS remainder a day before the advance voting period and a day before the election day. Below are English versions of the messages.
1.Before the advance voting period
“Hi, county and municipal elections will be held on April 13. Advance voting period in Finland is April 2 - April 8. See vaalit.fi. Regards, Ministry of Justice”
2.Before the election day
“Hi, county and municipal elections will be held on April 13. See vaalit.fi. Regards, Ministry of Justice”
After the election the outcome variable from the electronic voting registry is merged into covariate data and treatment statuses from previous two experiments (Hirvonen, Salomo et al. 2022. "Text Message Mobilization of Young Voters." AEA RCT Registry. January 11. https://doi.org/10.1257/rct.8790-1.0 and Hirvonen, Salomo et al. 2023. "Randomized SMS reminders to vote for youth voters in parliamentary elections." AEA RCT Registry. March 21. https://doi.org/10.1257/rct.11105-1.0), and pseudonymised by SF. We will use the Linear Probability Model as our main model for the ease of the interpretation of the coefficients but will also conduct analysis using logit regression as a robustness check. We will test potential effects for treatments for foreign language speakers all languages pooled and each foreign language separately. In addition, we will conduct heterogeneity analysis with all foreign languages pooled with respect to length of residence in Finland, 1st vs 2nd and 3rd generation immigrants, democracy index and geographical grouping of the origin country.
For youth voters we will test persistence of the treatment effect from two previous experiments and dynamic effects i.e. difference between being treated only in this experiment vs. being treated in this experiment and previous experiments. We will study intra household spillover effects, both spillovers from children to parents and vice versa.
We will estimate potential heterogeneous effects by predicted voting propensity, predicted previous voting participation, geographical area (municipality type) and socioeconomic status (education and profession). In addition to that, we will employ the Honest Causal Forest machine learning algorithm in order to further explore possible heterogeneous treatment effects and to possibly find groups with negative treatment effects due to paternal backlash. For youth voters we replicate analysis from our 2022 and 2023 experiments.
We will predict a propensity to vote for every individual using the available administrative data utilising logit and elastic net models using control group individuals as the sample and voting in 2025 county and municipal elections as the dependent variable. Independent variables, measured before the treatment, for the prediction model are individuals’ gender, age, immigration background, logarithm of taxable income, educational background, SES background, eligibility to vote for the first time and municipality fixed effects. We will group the voting propensities by 25th, 25-75th, and top 25th percentiles in order to test possible heterogeneous treatment effects between different voting propensity groups. We will do this analysis for all voters, and then separately for youth and old voters, where for the former we will use parental covariates for the prediction model. We will also conduct an analysis where we will split the voting propensity groups by voting in past elections.