Getting Personal? The Impact of Social Media on Preferential Voting

The little work that has been done on social media or web 2.0 focuses on majoritarian democracies or the Irish (personalized) STV-system, and it remains to be seen whether these findings also hold in a list-PR setting (Gibson and McAllister 2011; Rackaway 2007; Sudulich and Wall 2010).

As these variables are only included as controls, they will not be translated into hypotheses.

A preference vote is a vote to indicate a preference for (an) individual candidate(s) on a party list that includes multiple candidates.

We use the word ‘follower’ in a broad sense, not to refer merely to Twitter followers. As used here it means all people being social media ‘friends’ or ‘followers’ of a candidate.

To be clear, the label ‘interaction effect’ refers to the statistical type of effect, and not to the effect of interaction between candidates and their followers.

This causal mechanism is in line with the first mechanism where we contend that the number of messages is important in creating a connection with voters and increasing the probability of being remembered.

The electoral quota equals 1/150 or 0.67 % of all votes cast.

Candidates who crossed the preference threshold move to the front of the line, but if a party has no seats, these candidates do not conquer a seat. The candidate with the highest number of preference votes gets served first, as long as they have passed the threshold. For all other candidates the list position is decisive.

Given that the Dutch electoral system is very proportional and de facto only has one district, the number of electable positions on the list is fairly high. As a result, most popular candidates would get elected based on their position on the list anyway.

In 2010, the Netherlands had some 16.8 million inhabitants.

The following parties received at least one seat: VVD, PVDA, Partij voor de Vrijheid, CDA, SP, D66, GroenLinks, ChristenUnie, SGP, Partij voor de Dieren (www.​kiesraad.​nl).

List-pullers are the first candidates on the list. In total, 84% of all votes cast on the candidates in our data was cast for the 10 list-pullers.

Only two out of 10 parties had a woman as list-puller.

The following newspapers were included in our measurement: AD/Algemeen Dagblad; Algemeen Dagblad; Boerderij Vandaag; Dag; Dagblad De Pers; De Volkskrant; Het Financiële Dagblad; Het Parool; Metro (NL); Nederlands Dagblad; NRC.NEXT; NRC Handelsblad; Reformatorisch Dagblad; Spits; Trouw.

The intercoder reliability is 97.3 %, a value that can be considered very high, which is not surprising given the availability of explicit and fairly objective coding criteria. The disagreements mainly arose regarding ‘opinion leaders’ and ‘Presidents of highly visible NGOs’. The final decisions were taken by consensus.

The fairly low numbers are consistent with findings in other countries (e.g. Gibson and McAllister 2011; Hoff 2010).

The incumbency variable is mostly used in American research and is said to have a substantial positive effect. However, in a proportional electoral system with multi-member districts, the incumbents are routinely placed higher on the list. Hence most of the incumbency effect is probably absorbed by the latter variable. A second cluster of variables that most likely absorbs a substantial part of the effect consists of being well known and the number of newspaper articles a candidate featured in the year before the campaign. Moreover, a party-list proportional representation system produces a lot of lesser-known incumbents as they have no local constituency, but are still elected. Hence, it should not come as a surprise that the effect of the incumbency variables is very limited or absent. The same holds for having a personal website: many candidates (174), including many of the lower-ranked ones, had a website, and we control for list placing and a host of other variables such as media exposure.

It needs to be stressed though that using the average offers the best case scenario. Many candidates had far lower numbers of followers, as the average is biased towards the upper end (cf. Fig. 2).

An explanation for this relationship not being statistically significant may be that the number of Hyves followers has a very high skewness. Indeed, 69 % of all Hyves followers are accounted for by just two candidates; for Twitter this percentage is only 26 %.

This mirrors the logic of election research where citizens younger than 18 are not included as they cannot vote. Indeed, including candidates who do not have a social media account would artificially inflate the number of candidates with low scores on social media usage as having no account automatically means a ‘zero’ score on usage. Our dataset still includes candidates with zeroes on one or the two usage variables. Such candidates have an account, but do not use it at all. Since direct effects of social media are found, including artificial zeroes on the usage variables potentially biases the results in favor of our hypotheses, which is a second reason to exclude the candidates who do not have an account from the interaction models.

To test the robustness of these models we also performed the analysis for Twitter and Hyves together (integrated Models 3a and b), with all the candidates that used at least one of the two platforms. That model shows results that are similar to the ones reported here: both interaction effects are positive and statistically significant (at least p < 0.01). The coefficient for Twitter remains 11, for Hyves it decreases a bit to 784 (compared to 1,151 here). These additional analyses support our conclusions. In addition, we reran the interaction models on the full sample of 493 candidates as well, both for Twitter and Hyves separately and for both combined. All four interaction coefficients in the three models are statistically significant (p < 0.01). For both Twitter and Hyves we find interaction effects similar to the one reported here. The B-coefficient for Twitter is 9 or 10 (here 11) and for Hyves 964 and 1,268 (here 1,151). Again these robustness checks strengthen our conclusions.

The campaign started on 27 April and lasted until 8 June 2010.

As 49 % of the candidates who have a Hyves account also have a Twitter account, some of the impact of Twitter use may well be included in the Hyves effect. Another reason to be cautious is the fact that only 47 candidates updated their profile, so the data are very skewed. Moreover, it can be questioned whether the interaction effect is as gradual as presented here. For the usage of Twitter we were able to partly test this by running the model with a dummified ordinal variable for the number of tweets send, and interaction terms of each of those dummies with the number of followers. This model is not presented here because it is far more complex and yields roughly similar results: according to our data the effect is relatively stable and it is not the case that the major difference is found between candidates who update their account and candidates who do not. We distinguished groups based on the average number of tweets per day: 0-1; 1-2; 2-3; 3–5; 5–10; >10. The model on all 493 candidates showed a more or less linear model, or with some creativity two thresholds between three groups could be derived from the model: (1) candidates with no tweets up to 1; (2) candidates tweeting 1 to 5 times a day; and (3) candidates who tweeted at least 5 times a day. A model on the 168 Twitter users only showed the most difference between candidates not tweeting, candidates tweeting up to once a day, and candidates tweeting more regularly. However, the standardized coefficients of that model also suggested a gradual effect. For Hyves, we could not carry out such additional analyses. The nature of the platform is less directed to publishing small updates and only 47 candidates placed one or more short updates, in our data set: 21 placed one update, 6 two, and the rest between 3 and 32. The groups would become too small to perform the interaction analyses. All in all, these additional tests suggest that the advantage of having more followers in combination with posting more updates is cumulative and not just a difference between candidates who Tweet and those who do not.

Based on these coefficients, one can assess the overall impact of social media on the election outcome. It turns out the effect was very limited: social media made the difference for just two candidates who would not have been elected otherwise, namely Sabine Uitslag (15,933 preference votes) and Pia Dijkstra (15,705 preference votes).

Bond et al. (2012) show that such multiplicative effects can amount to significant numbers. In their study on political mobilization in the U.S. they show that the ‘spread of messages’-effect was small for each individual voter, but given large numbers of potential voters, the estimated overall effect adds up quickly.

Moreover, as grassroots campaigning is virtually nonexistent in the Netherlands, the country is an excellent case to study the impact of social media in the absence of grassroots campaigning.