The estimate given in the title suggests a much higher number of American atheists than estimates from other studies relying on self-report (e.g., “Are you an atheist?”). Those self-report estimates range between 3% and 11% (the authors of the paper below define “atheists” as “people who disbelieve or lack belief in the existence of a god or gods”). The higher number from the present study could reflect errors or biases in how the authors derived their estimates, or (and I think this is a bit more likely) the fact that people’s atheism was estimated indirectly rather than by self-report.
Note that the paper, which you can get free by clicking on the link) is on Arχiv, so it hasn’t yet gone through peer review. Nevertheless, the results are interesting and it’s well worth reading. (It’s not overly long.)
I’ll try to be brief. The authors estimated the frequency of atheists among Americans by surveying people using two YouGov samples of 2000 people each. They also did their estimates using Bayesian techniques: seeing what proportion of atheists in the public was most likely to yield the survey results. The composite result was, as I said, 26%.
How did they indirectly estimate the proportion of atheists? They used a clever technique in which people were asked to list how many statements were either true or not true about them, with one list adding an atheist belief and the other missing it. The difference in the number of statements that people agreed or disagreed with in the two lists can be used to estimate the proportion of atheists. They also had a control question that you’ll see in on the second list, which should yield a 0% estimate of people who think that 2 + 2 is more than 13. The authors say that this indirect method of estimation has been showing in other studies to be revealing in that it gives a higher percentage than self-report, but only for socially sensitive traits which people don’t want to disclose in a direct self report.
The authors also asked people directly if they were atheists, so they have an estimate of self-report.
This first group of questions yielded a Bayesian estimate of 32% atheists, with 95% confidence intervals of 11% and 54%, while the self-report (first question) yielded only 17%. You can see that the added question is in the third column and the participants aren’t supposed to say which statements aren’t true of them, but merely give the number. The difference between the totals in column 2 and 3 can be used to give a Bayesian estimate of the proportion of people who do NOT “believe in God”:
Some confirmation of the technique’s validity comes from analyzing the data from those who self-report being atheists in column 1. The Bayesian data from columns 2 and 3 give an estimated proportion of atheism of 100% of these people, so the self-report among those brave enough to disclose their atheism matches the indirect estimate.
Sample II used the same method, but couching atheism as a positive rather than a negative answer (i.e., you have to note whether “not believing in God” is true of you). There’s a control question about math in the third column.
This report yielded an estimate of atheism (comparison of first versus second column) of 20%, with confidence intervals of 6% and 35%.
The lower estimate in Sample II versus Sample I may, as the authors note, be attributed to the fact that in the second sample you have to note (indirectly) that atheism is “true of me”, which is more similar to a self report. And indeed, the 20% Bayesian estimate here is close to the self-report estimate of 17%.
Overall, combining both studies gave a Bayesian estimate of the proportion of atheists in America of 26%, with confidence intervals of 13% and 39%. The authors add that it is 99% certain that more than 11% of Americans are atheists (the Gallup poll estimate) and 93% certain that the proportion of atheists is higher than 17% (their self-report estimate). That means that about a third of atheists won’t disclose their nonbelief when asked directly. At any rate, the higher estimates from this study than in direct-question surveys suggests that there are far more atheists in America than believed: perhaps more than 80 million.
One weakness of the study is that the control question, which should show 0% of people rejecting the statement “I do not believe that 2 + 2 is less than 13”, actually gave an estimate of 34%. The authors note this, showing that they are careful about the data:
Without a doubt, this is our most damning result (cf. Vazire, 2016). It may reflect any combination of genuine innumeracy, incomprehension of an oddly phrased item, participant inattentiveness or jesting, sampling error, or a genuine flaw in the unmatched count technique. Fortunately, we were also able to assess validity in a second way. In Sample II, the unmatched count to generated an atheist prevalence estimate of almost exactly 100% among self-described atheists, but only 13% among all other religious identifications. It is unlikely that a genuinely invalid method would track self-reported atheism this precisely. Across two assessment attempts our validity evidence was a mixed bag. This perhaps suggests that future researchers should attempt to—as we were able in Sample II but not Sample I—include diagnostic self-reports alongside the unmatched count to assess validity. And, as the present estimates are only as strong as the method that generated them, they should be treated with some caution. In our view—given heavy social pressures to be or appear religious—the 11% atheism prevalence estimates derived solely from telephone self-reports is probably untenable. Does this imply that our most credible estimate of 26% should be uncritically accepted instead? Of course not. The present two nationally representative samples merely provide additional estimates using a different technique, and our model suggests a wide range of relatively credible estimates. We hope that future work using a variety of direct and indirect measures will provide satisfactory convergence across methods, and the present estimates are merely an initial indirect measurement data point to be considered in this ongoing scientific effort.
Finally, here’s a table breaking down atheism (both self-report and indirect estimates) by sex, politics, age, and education. We see that the prevalence of atheism isn’t that disparate among any groups except “political affiliation”, but follows the familiar pattern of more atheists among males than females; more atheism among more highly rather than less educated people; more atheism among Democrats than among Republicans (note the 0% indirect estimate for Republicans!); and no difference between Millennials and baby boomers. Self-report is always lower than indirect estimates except among Republicans, which is a mystery. (The last column gives the probability that the indirect estimates are higher than the self-reporting estimates.)
The authors discuss the wider implications on the last page of their paper, noting that we can’t extend these estimates to the rest of the world because the degree of underreporting in the direct-question technique (the only one used) may vary, with opprobrium less in countries like Norway and Denmark, and greater in countries like Saudi Arabia. But the authors do speculate that there may be around two billion atheists worldwide, which makes the number of atheists higher than the number of Muslims (1.8 billion), but less than the number of Christians (2.4 billion) And their final paragraph gives the implications for the social acceptance of atheists:
Finally, the present results may have considerable societal implications. Preliminary research suggests that learning about how common atheists actually are reduces distrust of atheists (Gervais, 2011). Thus, obtaining accurate atheist prevalence estimates may help promote trust and tolerance of atheists—potentially 80+ million people in the USA and well over a billion worldwide.
Join the club! I’m referring to you, Andrew Sullivan!
And for dessert, you can have this new op-ed by David Leonhardt about the demonization of atheists in America (h/t: Greg Mayer):
h/t: Ginger K





