Tweets That Matter: Evidence From a Randomized Field Experiment in Japan

Abstract

Although election campaigns are increasingly utilizing social media, only a few studies have investigated their effects experimentally. To fill this gap in the literature, we conducted a field experiment to examine the effects of a campaign that used Twitter during the 2013 House of Councillors election in Japan. The treatment was exposure to tweets from Tōru Hashimoto, the mayor of Osaka and co-leader of the Japan Restoration Party, who has the largest number of Twitter followers among Japanese politicians. Participants assigned to the treatment group followed Hashimoto and the two placebos, whereas those assigned to the control condition followed only the two placebos. They followed the politicians continuously for approximately one month. Pre- and posttreatment measures were collected using online surveys, and treatment compliance was continuously checked via Twitter application programming interface (API). Following Hashimoto on Twitter during the election campaign had a positive impact on feelings toward Hashimoto. This effect was not mediated by issue knowledge or the evaluation of Hashimoto’s personal traits, and no effects were observed on voting. These findings suggest that repeated exposure to a politician’s messages on Twitter may only result in a mere exposure effect, which nevertheless generates favorable overall attitudes about the politician.

Keywords:

• social media
• election campaign
• Twitter
• mere exposure effect
• field experiment

Notes

1. The average frequency of reading timeline was measured with a 9-point scale: (1) every day; (2) six days a week; (3) five days a week; (4) four days a week; (5) three days a week; (6) two days a week; (7) one day a week; (8) less than one day a week; and (9) never. The reading devices included both computers and mobile devices.

2. The style of reading tweets was measured with a bipolar 4-point scale ranging between (A) I read all the tweets on timeline and (B) I read only tweets that I have an interest in and skip others. Only the respondents who selected “close to A” or “rather close to A” participated in the experiment.

3. Participants were restricted to respondents who selected “I read all the tweets displayed on one timeline” in a multiple-answer item regarding the display of tweets.

4. It would take too much time to confirm whether participants followed their assigned politicians using Twitter API when more than 5,000 accounts are followed.

5. It is possible that the participants were following politicians other than Abe, Hosono, and Hashimoto. However, the effects of following other politicians are on average balanced across treatment and control groups because of random assignment, and thus it does not bias the estimation of the treatment effect.

6. This design cannot reveal the effect of following Abe and Hosono on attitudes toward Hashimoto and the JRP. To do so, it would be necessary to set up another control condition in which participants follow none of the three politicians and then compare the posttreatment measurements with the original control group. However, this is beyond the scope of the present study, and any effect that arises from following Abe and Hosono is balanced between the treatment and control groups.

7. Had there been no effect of history, the feeling thermometer scores for the treatment group would have increased because of the positive mere exposure effect. This argument is based on the idea that the effects of history and mere exposure work independently. The independence of the mere exposure effect from the effect of history is supported by the findings that the mere exposure effect is rooted in implicit memory (see, e.g., Seamon et al., 1995) and that implicit memory works independently of explicit cognitive processes (see, e.g., Halpern & O’Connor, 2000; Matlin, 1971; Moreland & Zajonc, 1977). Therefore, because the directions of the effects in the implicit and explicit processes were opposite in the present study, they canceled each other out in the treatment group.

8. We do not need to test whether knowledge of JRP issue positions is a mediator because there was no treatment effect on knowledge (Baron & Kenny, 1986).

9. The closest comparisons between the original field experiment and the replication are Model 3 in Table 1 and Model 1 in Table 2; both are CACE without covariates. With these two models for Hashimoto’s feeling thermometer scores, the joint null probability using Fisher’s method is 0.038, which is still significant. However, this test can only be used for reference because our replication was not a direct replication of the original experiment.

10. Note that this consistency could suggest either that the campaign messages were not effective, even using the forced exposure design, or that the reduced number of tweets in the replication created spurious consistency.

Additional information

Funding

This work was supported by JSPS KAKENHI Grant Number 25871051.

Notes on contributors

Tetsuro Kobayashi

Tetsuro Kobayashi is Associate Professor, National Institute of Informatics, Tokyo. Yu Ichifuji is Postdoctoral Researcher, Transdisciplinary Research Integration Center, Research Organization of Information and Systems.

Yu Ichifuji

Tetsuro Kobayashi is Associate Professor, National Institute of Informatics, Tokyo. Yu Ichifuji is Postdoctoral Researcher, Transdisciplinary Research Integration Center, Research Organization of Information and Systems.