"Third, we use the results of our new survey to estimate the share of Americans who saw and believed..."

Third, we use the results of our new survey to estimate the share of Americans who saw and believed each of a set of 14 fake news headlines. … To address survey misreporting, we also include a set of placebo fake news headlines – untrue headlines that we invented and that never actually circulated. This approach mirrors the use of placebo drugs as controls in clinical trials. Consistent with a similar survey carried out the same week as ours (Silverman and Singer-Vine 2016), about 15 percent of U.S. adults report that they recall seeing the average fake news headline. About 8 percent report seeing and believing it. However, these numbers are statistically identical for our placebo headlines, suggesting that the raw responses could overstate true exposure by an order of magnitude.

Using the difference between actual and placebo stories as a measure of true recall, we estimate that 1.2 percent of people recall seeing the average story. Projecting these per-article exposure rates to the universe of fake news in our database under the assumption that exposure is proportional to Facebook shares, our point estimate suggests that the average voting-age American saw and re- membered about 0.92 pro-Trump fake stories and 0.23 pro-Clinton fake stories in the run-up to the election. Our confidence intervals rule out that the average voting-age American saw, remembered, and believed more than 0.71 pro-Trump fake stories and 0.18 pro-Clinton fake stories.

Comparing the magnitudes of the different coefficients in column 1 suggests that fake news exposure might be ideologically segregated: Republicans are more likely than independents, and independents in turn more likely than Democrats, to report seeing pro-Trump headlines, although for pro-Clinton headlines, the differences are less stark. In column 2, however, we see very similar results for Placebo headlines, and column 3 shows that five of the six coefficients do not differ for Fake relative to Placebo. In the context of our model, we interpret these results similarly to the results of table 4: differences across people in recalled exposure seem to be primarily driven by differences in perceived plausibility, and less by differences in true exposure. There may still be differences in true exposure, but this would need to be documented with web browsing data instead of our survey recall measures.

Are there indeed differences in believing rates that might generate these differences in false recall? Columns 4 and 5 indeed show dramatic differences: Republicans are four to eight times as likely as Democrats to report believing pro-Trump headlines, and Democrats are 50 to 100 percent more likely than Republicans to believe pro-Clinton headlines. Appendix table 1 repeats these regressions in the subsample of social media users with ideologically segregated networks. The relative ordering of coefficients is similar, but the magnitudes are considerably larger: social media users with segregated networks are, as we saw above, more likely to report seeing and believing fake news, and relatively more likely to report seeing and believing fake news that favors their candidates.