Increasing turnout with a text message: evidence from a large campaign from the government

ABSTRACT

This paper analyzes the effect of a voter mobilization campaign to increase electoral turnout at the population level. The government sent a text message with an encouragement to vote to all 200,000 voters in Bergen, Norway. The timing and the order of the text message was randomized, and I am able to measure the immediate treatment effect by exploiting a unique data set containing the exact timing of each Norwegian vote. I estimate that the text message increases electoral turnout by between 1 and 2.2 percentage points.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 See e.g. Verba, Nie, and Kim (1978), Wolfinger and Rosenstone (1980), Hicks and Swank (1992), Hill, Leighley, and Hinton-Andersson (1995), Lijphart (1997), Mahler (2008), Fowler (2013).

2 This large literature is summarized by Green, McGrath, and Aronow (2013), Green and Gerber (2015), Gerber and Green (2017).

3 Notable exceptions include Gerber, Green, and Larimer (2008) finding a treatment effect of 8 percent by threatening to publish voting records.

4 An organization called Vote.org also conducts GOTV studies using text messages, and their findings are available on their website https://www.vote.org/research/.

5 Bergh, Christensen, and Matland (2020) also show that letters are effective for mobilizing immigrants to vote.

6 Another possibility to understand the population-level effect could have been to send the text message to a large majority, e.g. 90% of percent, of voters. But given that the effect is contagious within households and social networks, such a design can be problematic if the voters in the control group are affected by the treatment.

7 In all Norwegian cities, early voting is higher close to the election day. To separate the time trend from the effect of the text message, I also analyze changes in voting in Bergen relative to other cities and changes in voting within Bergen.

8 I also show that the set of comparison units may be changed without affecting the results.

9 The data is available at ssb.no/statistikkbanken and valg.no.

10 These districts vary in size. Hordaland, where Bergen belongs, elected 16 out of 169 MPs in 2017, while the neighboring district (Sogn og Fjordane) elected only 4 MPs (see, e.g. Fiva and Smith 2017). To clarify the somewhat confusing notation, the term electoral district here refers to the larger electoral district, while districts in other sections refer to boroughs within Bergen.

11 The following website provides further official information about the Norwegian electoral system: https://www.regjeringen.no/en/topics/elections-and-democracy/den-norskevalgordningen/the-norwegian-electoral-system/id456636/

12 See Smith Jervelund and De Montgomery (2020) for a discussion of the validity and value of Nordic register data.

13 There are various differences to minimize, but I use a regression-based method that is used as default in the Synth package developed for Stata. This package can be found at https://web.stanford.edu/jhain/software.html.

14 I will show that this assumption is satisfied in this setting, in particular when I extend the set of control units.

15 The closest city, Haugesund, is more than three hours away from Bergen by car. The possibility of spillover effects through electronic communication cannot be excluded, but this is arguably not a major issue.

16 I searched the media database available through Nasjonalbiblioteket for the Norwegian word for electoral turnout, which is valgdeltakelse (also including the alternative spelling valgdeltagelse). I then counted the number of matches in the period from 1 August 2017 to 8 September 2017 in newspapers located in Bergen and other parts of the country.

17 See, e.g. Wolfinger and Rosenstone (1980) and Blais (2006) for discussions of which variables that affect electoral turnout.

18 In Norway in 2013, electoral turnout was higher in the 3 largest cities than in the other 16 cities in this sample.

19 The pool of 19 cities were chosen using an optimization algorithm, but it is also important to get a close pre-treatment fit. This raises a dilemma, as I want to the stay true to the pre-specified setup, but also want to get a close pre-treatment fit. I choose to report the results using both specifications.

20 The weights for turnout in the synthesized version of Bergen are Trondheim (0.476), Tønsberg (0.253) and Oslo (0.274) with 19 cities in the donor pool. With 23 cities, the weights are given by Trondheim (0.338), Bærum (0.362) and Haugesund (0.3).

21 Using the rdrobust pacakage for Stata available from https://sites.google.com/site/rdpackages/rdrobust. RD allows for choices of bandwidth and local polynomials. I use the default options in the rdrobust package when I estimate the treatment effect. I only use observations from few days around the treatment, as Hausman and Rapson (2018) argue that using observations distant in time from the treatment may lead to biased effects.

22 The average voter is exposed for the treatment effect for 7.25 h and Bergen has 199918 voters. If 72 voters are mobilized each 15 min, the number of mobilized voters per hour is 288, which leads to a treatment effect of $((7.25\ast 288)/199218)\approx 1.0\text{\%}$. This underestimates the true effect if voters are mobilized also on subsequent days, but may overestimate the effect if the effect decreases later in the day of the text message.

23 There are 139,946 voters in Bergen that have not cast their votes at the time of the text message, which leads to a treatment effect of $((0.0027\ast 7.25\ast 139946)/199918)\approx 1.4\text{\%}$.

24 I construct the two groups to be deliberately balanced on certain variables, so I randomly draw a larger number of other possible groups that satisfy balance on these variables, and then I analyze the balance for these hypothetical realizations on other variables that I do not use for constructing balance.

25 The text messages to the first group were sent out uniformly through this hour, the voters in Group 1 were, on average, exposed to the text messages for half of this time. The hourly estimated effect is then 196 votes for Group 1, and Group 1 consists of 98457 voters. Extrapolating to the rest of the day leads to a treatment effect of $((196\ast 7.25)/98457)\approx 1.4\text{\%}$.

26 The correlation is strongly negative; if a large share of voters chooses to vote before the election day, there are few potential voters left to mobilize.

27 The estimated number of mobilized voters lies between 0.01∗199,918 and 0.022∗199,918, while the cost of the campaign is given by $0.08∗199,918. The total cost of the campaign divided by the number of mobilized voters lies in the interval between $(\mathrm{\$}0.08/0.022)\approx \mathrm{\$}3.6\mathrm{and}(\mathrm{\$}0.08/0.01)\approx \mathrm{\$}8$.