- see complementary [[220316_143628 goodness multiverse interaction models|goodness multiverse interaction models]] # parameters - bad/goodness measures: fc, mbfc (factual, bias, mean(factual, bias), min(factual_bias)), afm (reliable, bias, mean(reliable, bias), min(reliable, bias)) (see [[different measures of domain quality]]) - goodness thresholds: $[5, 10, 15...90, 95]$ - values **below** threshold are set to $100$ (maximum badness) - values **above** threshold are scaled $[99...0]$ (decreasing badness) - outcomes: **sum** badness (range: $[0, \infty]$) or **mean** badness (range: $[0, 100]$) - for both outcomes, users with NA values are filled with $0$ - see [[220316_121141 impute after winsorizing with mean vs zero|imputing with mean vs zero differences]] - winsorize user summed values (both at t1 and t0): $[0.90, 0.95, 0.99, 1.00]$ (1.00 = no winsorizing) - mean badness not winsorized (since upper bound is 100, probably doesn't make sense to winsorize) ## bad/goodness threshold examples ![[dv_mbfc_badness_threshold70.png|600]] ![[dv_fc_badness_threshold15.png|600]] # results models - sum badness, quasipoisson model: `feglm(t1 ~ conditionC * t0LC | block, dt1, family = "quasipoisson", vcov = "HC1")` - mean badness, OLS: `feols(t1 ~ conditionC * t0LC | block, dt1, vcov = "HC1")` - `conditionC` $[-0.5, 0.5]$ - `t0` time0/pre-campaign badness - sum badness: `log(t0 + 1)` then mean-centered - mean badness: mean-centered - `block` as fixed effect summary - `condition` effect: significant for summed badness (but not mean badness) - coefs mostly negative (i.e., treatment group shared less bad stuff) - winsorizing matters: lower values (0.95, 0.90) associated with larger `condition` estimates and smaller SEs - fact-checker measure has larger SEs - `conditionC * t0LC` interaction: not much? only for fact-checker mean badness when thresholds <= 15 (see [[220315_113004 interaction results|here]] for interaction results/figures) ## condition estimate ### summed badness - x-axis: threshold - y-axis: quality measure - z-axis (color): coefficient/estimate value (darker = more negative) - top: estimate - bottom: standard error for the corresponding estimate - significant results at different p-cutoffs (.05, .01, .005) highlighted in different shapes/circles/colors (if p > .05, not highlighted) ![[interact_retweet-badness_conditionC_sum_winsorize1.png]] ![[interact_retweet-badness_conditionC_sum_winsorize0.99.png]] ![[interact_retweet-badness_conditionC_sum_winsorize0.95.png]] ![[interact_retweet-badness_conditionC_sum_winsorize0.9.png]] ### mean badness ![[interact_retweet-badness_conditionC_mean_winsorize1.png]] ## condition-t0 interaction estimate Not much. See [[220315_113004 interaction results|here]] for results/figures.