Modelling the interplay between news reporting and user engagement through random effects regressions with AR(1) disturbance.

Results refer to the model , where <i>i</i> is the news source identifier, <i>t</i> is time (days since January 1st, 2016), <i>y</i>_<i>it</i> is the user engagement on vaccine-related content from source <i>i</i> on day <i>t</i>,<i>x</i>_<i>i</i> is a vector of time-independent source-related variables (e.g., the factualness classification: Reliable or questionable), <i>z</i>_<i>it</i> are time-dependent factors (e.g., number, type and topic of news content), and <i>θ</i>_<i>i</i>,<i>ϵ</i>_<i>it</i> are error terms with auto-correlated component. Being the panel data unbalanced and <i>T</i>><i>N</i>, we choose a model with autocorrelated disturbances. Robust estimators of variance were also estimated, and the results are confirmed [<a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0316258#pone.0316258.ref067" target="_blank">67</a>,<a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0316258#pone.0316258.ref068" target="_blank">68</a>]. Model I results show that content conveying pro-vax stance decreases user engagement (elasticity -0.03%) while anti-vax stance increases it (elasticity 0.5%), especially when it is proposed by questionable sources (Model II, coefficient of the interactive variable 1.34%). Conversely, when questionable sources convey pro-vax stance, engagement drops by 0.3%. If we focus on topics (Models III and IV), “safety concerns” is the one that elicits the greatest reactions (0.4%), especially when content comes from questionable sources (maximum elasticity 1.1%). Engagement drops instead when questionable sources cover the topic of effectiveness of vaccination. (DOCX)