We have identified a numerical error in this paper. In particular, a coding error caused two estimates to be reported as numerically identical, when they were quite similar but not identical. This affects the Results section and Figure 2.

Specifically, the first paragraph of the Results section should instead include:
To increase precision, we include in the regression past voting behavior in 2010
(encoded as two dummy variables, voted and abstained, with unknown status as the
baseline category). This yields an estimated treatment effect of +0.23​% (95% CI +0.01​%
to +0.46​%) that is unlikely to be due to chance (p = 0.0404​). We also estimated a
post-stratified model that included separate average treatment effects for those who
voted, abstained, or had unknown behavior in the 2010 election and combined these,
weighting on the number of individuals in each stratum. This procedure provides
additional robustness due to the different sized experimental groups and potential for
subgroup heterogeneity [24]. These estimates yielded nearly identical average treatment
effect of +0.24% (95% CI +0.03% to +0.44%, p = 0.0266).

Changes are shown in bold. Thus, the two estimates are distinct, but still "nearly identical". The estimates round to the same five-decimal proportion, but to different four- or six-decimal proportions.

Here is what we reported for both in the main text and the plot:
estimate: 0.002355
95% CI: [0.0003, 0.0044]
p-value: 0.0266

Here is the corrected version for the regression adjusted estimate of direct effects:
estimate: 0.002348
95% CI: [0.0001, 0.0046]
p-value: 0.0404


This also affects Figure 2. The second bar from the left ("banner + feed conditioned") is incorrectly identical with the third bar. We have created a corrected version of Figure 2. Only the standard error bars are visibly different. The corrected figure is available at https://doi.org/10.6084/m...

Thus, the main conclusions of the paper remain unchanged. All of the subsequent calculations (e.g., estimates of total attributable effects) used our preferred estimate (i.e. from the stratified analysis) and thus are already correct.

The code in the replication archive on Dataverse at https://doi.org/10.7910/D... continues to produce the correct results.

An additional typographical error exists in the sentence:
Fig 3 also shows our attempt to discern whether the joint treatment effect was driven by
seeing the banner or seeing stories in News Feed about other people who had clicked
on the “I voted” button.
Here "Fig 3" should be corrected to "Fig 2":
Fig​ ​2​ also shows our attempt to discern whether the joint treatment effect was driven by
seeing the banner or seeing stories in News Feed about other people who had clicked
on the “I voted” button.

Our apologies for this error.

Jason J. Jones
on behalf of the authors