Trained causal forests on **pilot campaign** data on covariates below (outcome being modelled is daily **count badness**). Used the trained forests to predict treatment effects tau for users in **main campaign**. Ran analyses below on only users with predicted negative treatment effects. ```r # outcome being trained on: count badness # t1 and t0 values were NOT winsorized # covariates expose, t0, n_retweet, friends_count, followers_count, favourites_count, statuses_count, friend_follow_ratio, days_since_create, name_len, name_upper_pct, quality_mean, quality_sum, quality_count, quality_count_yhat, quality_mean_yhat, quality_sum_yhat ``` # included all 13 days of campaign ![[count-winsorize-0.95-clustse.png]] ![[frac-winsorize-0.95-clustse.png]] ![[sum-winsorize-0.95-clustse.png]] ## models included only day fixed effect (removed block fixed effect) Perhaps blocking might be less of an issue since we're already conditioning on users with high predicted treatment effects, so we might not need it as a fixed effect. Results don't change much. ![[count-winsorize-0.95-clustse 7.png]] ![[frac-winsorize-0.95-clustse 6.png]] ![[sum-winsorize-0.95-clustse 8.png]] # included only last 7 days of campaign Exclude first 6 days, since we focused only on final week in the **pilot campaign**. ![[count-winsorize-0.95-clustse 2.png]] ![[frac-winsorize-0.95-clustse 2.png]] ![[sum-winsorize-0.95-clustse 2.png]]