- see also [[220305_142915 fact-checker threshold analysis|similar analysis for fact-checker domains]] - follow-up to [[220310_150631 mbfc threshold 80|mbfc threshold 80 analysis]] ## Threshold linear regressions (control for blocking) For each badness value (`b: [0...10000]`; x-axis), we fit a model (`y ~ condition[-0.5/0.5] * sum_t0LC | block`) where we assign the user's outcome `y = 100` if the user's summed badness during the campaign is above `b` (otherwise `y = 0`). - x-axis: each threshold value (one model per value) - y-axis: regression coefficients estimate (95% CI) - panels: different coefficients in the model ```r # winsorize data # time1 sum is log(time1 + 1) and mean-centered (hence LC) # for each badness thresold b in 0 to 10000 dt1[, y := ifelse(sum_t1 > b, 100, 0)] m_temp <- feols(y ~ conditionC * sum_t0LC | block, dt1[domain_type == "overall"]) ``` - first panel (**condition coefficient**): When `b` is between 0 and 1000, condition coefficient is **negative**: Being in the treatment (vs. control) group makes it less likely for users to be belong to the "lower-quality" (`y = 100`) group. ![[dv_mbfc_thresholds_interact 3.png|800]]