- afm_bias, thresholds 20/80 (bad: 0-20, good: 80-100) - **only 4 domains in bad bin!** results below are just so we can contrast the two sets of results (modeling `t1` vs `t1-t0`) - see [[220322_151215 bad-good-count condition interaction|original analysis modeling just t1 instead of t1 minus t0]] - 133 domains in good bin - outcome y: t1-t0 difference (separately for bad and good bins; 2 rows per user) OLS model ```r # conditionC: -0.5, 0.5 (control, treatment) # qualityC: -0.5, 0.5 (bad, good) dt1[, t1_t0_diff := t1 - t0] feols(t1_t0_diff ~ conditionC * qualityC | block, dt1, vcov = ~ screen_name) # OLS model > feols(t1_t0_diff ~ conditionC * qualityC | block, dt1, vcov = ~ screen_name) OLS estimation, Dep. Var.: t1_t0_diff Observations: 65,776 Fixed-effects: block: 5,424 Standard-errors: Clustered (screen_name) Estimate Std. Error t value Pr(>|t|) conditionC -0.048363 0.028850 -1.67637 0.093676 . qualityC -0.978786 0.031359 -31.21207 < 2.2e-16 *** conditionC:qualityC -0.114435 0.062718 -1.82458 0.068073 . # reparameterized OLS models Estimate Std. Error t value Pr(>|t|) conditionC 0.008855 0.018332 0.483022 0.629083 # treatment effect for 4 bad domains conditionC -0.105580 0.057405 -1.83921 0.065894 . # treatment effect for 133 good domains ``` quasipoisson model - similar results as OLS model - re-scaled such that min `t1_t0_diff` value is 0 ```r # quasipoisson model > feglm(t1_t0_diff ~ conditionC * qualityC | block, dt1, family = "quasipoisson", vcov = ~ screen_name) GLM estimation, family = quasipoisson, Dep. Var.: t1_t0_diff Observations: 65,776 Fixed-effects: block: 5,424 Standard-errors: Clustered (screen_name) Estimate Std. Error t value Pr(>|t|) conditionC -0.002206 0.001308 -1.68663 0.091684 . qualityC -0.043507 0.001425 -30.53001 < 2.2e-16 *** conditionC:qualityC -0.005167 0.002850 -1.81292 0.069853 . ``` ![[s20220323_114254.png]] ![[s20220323_113538.png]]