Google Searches and Turnout: A Cautionary Tale

by Erik Voeten on October 25, 2012 · 3 comments

in Campaigns and elections

Jesse Richman’s post below shows that compared to past election periods, Google Trends data indicate that searches for “vote” are generally good indicators for actual turnout and that searches in the current election period lag behind past elections. I find the first observation exceptionally interesting and believable. I think we should be cautious with the second.

Google Trends provides an indicator on a scale of 0 to 100. The value 100 will be assigned to the week in which the term was most searched for in Google’s sample compared to other search behavior.

The key problem is that the composition of internet users has changed dramatically since 2004. The figure below illustrates this. Searches for terms such as “university,” “science,” and “law” have gone down compared to overall search behavior in Google. If we were to interpret Google Trends data as indicators of societal interest, we would conclude that society is only about half as interested in these topics in 2012 compared to 2004. A more probable explanation is that the population of internet users in 2004 was more biased towards those interested in these topics than it is today (week 0 starts in 2004). Similarly, the population of internet users in 2004 was likely more biased towards those interested in politics.

One quick fix would be to create an index for these three searches and look at searches for “vote” as a proportion of that index. When you do that (see below) 2012 does not look so different from other years. Indeed search volume last week was nearly identical to that in 2004 and 2008 the same number of weeks before the election.

“Law,” “Science”, and “University” are pretty good candidates for search terms that may correct for changes in the composition of web users. There is little reason to think that underlying societal interest in these terms has changed much. Yet, this is not a perfect solution. There surely should be more search terms in such an index (which I would gather if I had more than the half hour it took to do this post). Ultimately, I just don’t think we can use Google Trends data for these long intertemporal comparisons in a very precise way. This takes nothing away from the importance of turnout for this election or the ability of search data to provide meaningful insights into where turnout may be highest.

{ 3 comments }

Pat October 25, 2012 at 5:16 pm

You could try something like “threesome,” “kinky” or “porn.” It’d be interesting to see if searches for sex-related terms have changed over that period of time.

Justin Murphy October 27, 2012 at 6:25 pm

I took another look at this and some related data, and after looking at what’s going on under the hood of “vote” searching, it actually looks like, if anything, more people might be preparing to vote than in previous elections… I’ve also made the data and R code available if anyone is interested in pursuing this.

http://justinmurphy.wordpress.com/2012/10/27/predicting-turnout-from-google-search-data/

Jesse January 9, 2013 at 3:55 pm

I appreciate the comments and potential corrections here… but now that the post-election turnout rate calculations are more or less complete, I think I can say with reasonable confidence that search frequency as a turnout indicator held up quite well.

It looks like the search volume was (on the whole) an indicator of lower turnout. According to http://elections.gmu.edu/Turnout_2012G.html turnout VEP turnout for the United States as a whole was down. On average across the states I examined in this piece (DE, ND, and WY were excluded because they lacked sufficient weekly search volumes) turnout was down by 3.3 percent and search frequency was down by 2.02 points on Google’s 0 to 100 frequency scale. As I suggested was likely to be the case in my original posting, people searched for “vote” less, and they also voted less.

Comments on this entry are closed.

Previous post:

Next post: