Fig. 3: Content the algorithm promotes.
Average content shown to users in each feed setting: grey, chronological feed, red, algorithmic feed; 95% CIs around subsample means (mean ± 1.96 × s.e.) are reported. The unit of observation is X post. The number of observations for each feed setting is n = 268,532 for all variables, except for ‘Conservative’ among ‘Political content’, for which n = 129,790 X posts in each feed setting. All variables are indicator variables except Likes, Reposts and Comments, which are heavily right-skewed. For the algorithmic feed, the minimum, first quartile, median, third quartile and maximum are 0, 174, 933, 8,197 and 1,000,000 for Likes; 0, 9, 137, 1,363 and 316,700 for Reposts and 0, 5, 56, 384 and 247,200 for Comments. For the chronological feed, the minimum, first quartile, median, third quartile and maximum are 0, 4, 40, 342 and 1,700,000 for Likes; 0, 1, 8, 66 and 572,000 for Reposts and 0, 0, 3, 21 and 256,100 for Comments. Corresponding summary statistics are presented in Supplementary Information Table 1.8. On the left, for each outcome, \(\Delta \) (the estimated difference in content between the algorithmic and chronological feed from regressions with user and survey-wave fixed effects) and the corresponding P value are reported. For Likes, Reposts and Comments, \(\Delta \) is expressed in percent and estimated using a Poisson pseudo maximum likelihood model; for all other outcomes, \(\Delta \) is in percentage points and estimated using linear ordinary least squares. Because fixed effects are included, \(\Delta \) does not exactly match the raw ratio or difference in means across feed settings shown in the bar chart. See Supplementary Information section 2.6 for the specification, full regression results and robustness; all tests are two-sided with s.e. values clustered at the respondent level.
