Extended Data Fig. 5: Engagement and Political Attitudes. ITT Estimates by Initial Feed Setting: Re-weighted to Address Selection into Initial Feed.
Replication of main-text Fig. 2 with subsample of respondents initially on the chronological feed re-weighted to mimic the subsample initially on the algorithmic feed, and vice versa. Intention-to-treat (ITT) effect estimates in SD. Left panel: ITT estimates of switching the algorithm on. Right panel: ITT estimates of switching the algorithm off. “Chrono” and “Algorithm” stand for the chronological and algorithmic feeds, respectively. For each outcome, the results of two specifications are reported. In blue: the unconditional estimates with robust standard errors, controlling only for the initial feed setting and, where applicable, pre-treatment outcome levels. In orange: the conditional estimates, controlling for pre-treatment covariates using Generalized Random Forests (GRFs). 90% and 95% CIs are reported. Numerical effect sizes and p-values printed on the figure correspond to the conditional estimates (all tests are two-sided). The unit of observation is respondent. From top to bottom, sample sizes are n = 4965, n = 3337, n = 4965, n = 4965, n = 4596, n = 4596, and n = 4850. Methods describe the tests. SI Table S2.16 reports the exact numerical point estimates, standard errors, confidence intervals and sample sizes for every specification. All outcomes are standardized.
