Do the activities of American Atheists lead to an increased interest in atheism? In our latest blog post Ryan Cragun discusses his statistical analysis of correlations between American Atheist activity and Google searches for the term ‘atheist’.
Thanks to an invitation from David Silverman, President of American Atheists, to present some of my research at his organization’s annual conference, I was present when he gave his opening speech in March 2013 to just over 900 attendees. Silverman developed an argument in his speech that may be of interest to those who study the nonreligious, and which social scientific tools can be used locate in wider context.
Using a visual analysis of Google Trends data, he claimed American Atheists were responsible for the growing interest in atheism. He did this by examining the peaks in searches for the term “atheist” in Google’s search engine. The chart he posted was similar to this one:
He then overlaid onto the first trend line a similar one of searches for American Atheists, like the one below:
Mr. Silverman then proceeded to point out some of the spikes and indicated what American Atheist’s activities led to them, like the bus advertisement campaign and the American Atheist’s billboards about Christmas (see here and here). With the lines for the two terms/phrases following somewhat similar patterns, Mr. Silverman arrived at his conclusion – American Atheist activity was driving interest in atheism. Of course, he quickly noted that correlation is not causation, but he then asserted, ‘That’s a lot of correlation!’
Having a few more statistical tools at my disposal than Silverman may have, I wondered if a more rigorous analysis of the data was possible and whether it would provide any insight into causality. In order to establish causality one must meet three criteria. First, the cause must precede the effect in time. Second, the cause and effect have to be correlated. And third, you must be able to rule out alternative explanations.
Google Trends had weekly search data for both terms/phrases going back to 2007 (even further back for ‘atheist’ but not for ‘American Atheists’, which only had monthly data going back to 2004 and is shown in the figure above). There are, however, a couple of potential concerns with the data. To begin with, Google Trends doesn’t report the absolute number of searches but rather the relative number with the maximum during the specified time period set to 100. Thus, in the two charts above, the peak is 100, but that is the peak from January 2004 through April 2013. To get finer detail in search interest, I had to break the data up into separate years, which means each year has independent maximums. This is not a problem when correlating the data, so long as correlations are done separately for each year. But this approach does mask a trend that Silverman pointed out: with every passing year, interest in atheism and atheists is increasing, setting new baselines. That’s a good point, but does not cause problems for my year-by-year analysis.
While the figures above use month-level data (i.e., the relative interest in the term/phrase for the entire month), I wanted to use week-level data as a month is an enormous amount of time for interest in a breaking news story about atheism to spread in the twenty-first century 24-hour news cycle. Even week-level data may not allow for fine enough discrimination to determine causality as an event involving an atheist unassociated with American Atheists could, within a day, raise interest in American Atheists and vice versa. Even so, given the time span of interest, the finest level of detail available to me was week-level data.
To analyze the data, I created separate files by year and aligned the data for the term ‘atheist’ with the phrase ‘American Atheists’. I then correlated the relative interest. The correlations are shown in the first row of Table 1.
|Table 1. Correlations and Lagged Correlations for “atheist” and “American Atheists” from 2007-2013 (data source: Google Trends).|
|AA → atheist||0.08||0.18||-0.02||0.17||0.00||0.29||0.09|
|atheist → AA||0.09||0.45||-0.18||0.49||0.23||0.45||0.57|
Correlations measure the relationship between two variables. A correlation can range from -1.0 to +1.0. The higher the absolute value of a correlation (i.e., the closer it is to ±1), the stronger the relationship. A ±1 would indicate a perfect relationship, meaning that if you know the value on one variable you would know the value on the other variable. A correlation of zero means there is no relationship between the two variables. Positive correlations indicate that as values of one variable go up, values on the other go up as well, while negative correlations indicate that as the value of one variable goes up, the values of the other go down.
The correlations between ‘atheist’ and ‘American Atheists’ have varied pretty substantially from year-to-year, with 2009 displaying a particularly unusual pattern. That could be due to some leadership complications for American Atheists during that period, resulting in a relative lull in activity. But since then, the correlation has been in the moderate to high range, between .46 (2011) and .81 (first part of 2013).
That searches for the two terms are correlated meets one criteria of causality, but the other two remain to be established. Temporal causality must also be established. A nifty little statistical trick provides a possible answer on this front: data lagging. Lagging data allows one to test whether prior values of one variable predict the values of a second variable. In other words, it may be the case that an American Atheist activity or event – like a new billboard campaign – that takes place in Week 1 increases searches for the term ‘atheist’ in Week 2. By examining the relationship between lagged versions of variables, temporal ordering can occasionally be determined.
Given that there are just two variables, the Granger Causality Test is applicable. I ran a Granger Causality test on the time-series data for relative search frequency by week for each year. In no year was one variable’s lagged values a significant predictor of the other variable. In other words, causality cannot be statistically determined between whether searches for ‘American Atheists’ cause searches for ‘atheists’ or vice versa. A simple illustration of this is shown in Table 1, where I correlated a single lagged value for each search. Row two correlates the one week lag of ‘American Atheists’ on ‘atheist’ and row three does the inverse. In no case is the lagged correlation larger than the non-lagged correlation (this isn’t how the Granger Causality Test works, but it helps illustrate that the lagged values do not improve predictions). In short, it’s impossible from these data to determine causality.
My inability to determine causality does not mean there is no causal relationship between American Atheist activism and interest in atheism. From a logical and theoretical standpoint, it would make sense that there might be some causal relationship. And Google Trends helps to illustrate this by including headlines in some of the charts, like this one (which also happens to show that searches for ‘American Atheists’ – the red line – are much less common than searches for ‘atheist’ – the blue line):
That’s also the approach David Silverman used in his presentation. And, perhaps with day-level search frequency (generally not available for ‘American Atheists’) or even hour-level search frequency causality could be determined. But, as it stands, causality cannot be determined.
Additionally, it is possible that other events influence both search terms. For instance, when Richard Dawkins makes the news, for whatever reason, it’s likely that both the term ‘atheist’ and the phrase ‘American Atheists’ see increases in searches on Google’s search engine. Such spurious events likely account for some of the spikes in both search terms/phrases, but it is also likely the case that American Atheist activism increases searches for the other terms at times, too. And if that is the case, then that is greater evidence for correlation overall, as specific spikes may be causal, but the long term trend appears to be more correlational.
Ryan T. Cragun is a sociologist with a primary interest in religion, specifically Mormonism, the nonreligious, and secularization. His research has been published in a variety of journals, books, and magazines, including Journal for the Scientific Study of Religion, Sociology of Religion, Journal of Contemporary Religion, and Nova Religio. He is also the author of two books: Could I Vote for a Mormon for President? and What You Don’t Know About Religion (but should). When not working, he is typically spending time with his partner and son, tinkering with computers, cooking, or hiking. For more information, see: http://www.ryantcragun.com