Does social media matter for elections?

Is campaigning on social media sites like Twitter and Facebook overrated? Did politicians and their teams fall victim to a buzzword?

The short answer is perhaps, because it seems that social media presence does not have a very large effect on election outcomes. So it is entirely possible that some politicians are overinvesting in their social media campaigns.

This post is based on a recent working paper by Cameron, Barrett and Stewardson (2013) and pertains to the 2011 New Zealand General Elections. But of course the findings are likely to be valid anywhere in the developed world.

The authors look at the number of friends/followers candidates have on Facebook and Twitter, and test whether these can help predict election results. They run an OLS regression with voter share as the dependent variable. The results for Facebook are as follows:

ols

There is not much difference if Facebook friends are measured two months, one month or one week before Election Day. Clearly, the effect of Facebook friends is small. Since their number is measured in thousands, if you can get a thousand more Facebook friends on Election Day ceteris paribus, your voter share is predicted to increase by 1.4 percentage points.

When using the change in the number of friends from a set date to the Election Day (not pictured here), the coefficient is higher somewhat, but significance is lost. The largest significant coefficient is that of the change in the number of friends over one month prior to Election Day. If change over said period is 1000 friends, a candidate’s predicted voter share increases by 1.6 percentage points.

Clearly the control variables can explain most of the variation in voting share. However, the effect of Facebook friends is significant albeit low. I used an F-test to compare Model (2) to Model (3) and indeed the increase in R^2 due to the addition of the friends variable appears to be significant (p = 0.021993).

The authors also estimate a probit model where the dependent variable is whether a candidate won or not. Here are the results for Facebook:

probit

Changes in friends are not significant anymore, and the number of friends is only significant at 10%. The number of Facebook friends thus seems to have little predicting power when it comes to the probability of victory.

A little analysis of my own based on these tables yielded the following graph:

prob_of_victory

(the graph is based on the assumption that there is no intercept term in the model; the authors do not say anything about this but their table does not specify an intercept term)

First note, that the predictive power of an additional thousand Facebook friends is virtually zero for all initial levels of friends for those who are either incumbent individuals, are members of incumbent parties or both. However, for those who are neither, the number of Facebook friends can have considerable predictive power.

Using the above graph one would predict that if there are two candidates, neither of whom are incumbent or are members of the incumbent party, then if Candidate A has 1000 friends on Facebook and Candidate B has 2000, then Candidate B is almost 17 percentage points more likely to win on Election Day according to the model. Clearly, the predictive power of Facebook friends is so high because the model in this scenario cannot differentiate between the two candidates in any other way.

Of course, it is a very rare situation that neither candidate is incumbent and/or a member of the incumbent party. Furthermore, once the candidates have 5000+ friends, the predictive power of Facebook friends is effectively zero even for this group.

Both the OLS and probit results are similar for Twitter. For details see the paper.

What does this all tell us? There are two general messages, one of which will be at least moderately surprising to you depending on what your expectations were a priori:

  1. Social media data can help us predict the outcomes of elections. The coefficients of Facebook friends and Twitter followers are both significant in the OLS model. Surprising if you thought social media had no effect.
  2. The predictive power of social media data is rather limited. In other words, the coefficients are low and insignificant (not in the statistical sense) relative to the control variables. Surprising if you thought social media can give us profound insights into political races.

An immediate corollary of the above two points is that social media data may be most useful for predicting closely contested races where even one percentage point can make a difference.

Finally, note that one may expect that in the future the coefficients of Facebook friends and Twitter followers will grow as more voters will use the internet both in general and to express their political views.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s