# fact-checker summed badness - threshold: 30 (for domains where original fc rating is > 30, badness = 0) - x: original fact-checker ratings (good to bad) - y: transformed fact-checker ratings (higher = worse) - good domains (original ratings > 30 have new badness scores [y-axis] of 0) - see also [[220310_101008 fact-checker summed badness - threshold 70|results when threshold is 70 instead]], [[220310_094605 iffy misinfome threshold - reverse-coded|same analyses but for iffy misinfome domains]] ![[dv_fc_badness.png|900]] # descriptives 32888 obs 17 cols | | NUnique| PercentMissing| Mean| SD| Min| Median| Max| Histogram| |:-------------------|-------:|--------------:|-------:|-------:|-------:|-------:|--------:|----------:| |block | 5424| 0| 2897.31| 1577.02| 0| 2973.00| 5423| ▅▅▅▅▆▆▆▆▇▇| |weight | 24| 0| 2.00| 0.22| 1.67| 2.00| 2.50| ▂▁ ▇ ▁▁▁| |fc_sum_t0_retweet | 2619| 0| 241.85| 693.65| 0.00| 0.00| 40279.17| ▇| |fc_sum_t1_retweet | 2214| 0| 203.21| 731.21| 0.00| 0.00| 44559.55| ▇| |fc_count_t0_retweet | 220| 0| 11.46| 22.45| 0| 3.00| 522| ▇| |fc_count_t1_retweet | 273| 0| 10.18| 25.48| 0| 2.00| 652| ▇| |conditionC | 2| 0| 0.00| 0.50| -0.50| -0.50| 0.50| ▇▇| |fc_sum_t0_retweetC | 2548| 0| 0.00| 693.65| -241.85| -241.85| 40037.32| ▇| After winsorizing (99th percentile) | | NUnique| PercentMissing| Mean| SD| Min| Median| Max| Histogram| |:----------|-------:|--------------:|-------:|-------:|-------:|-------:|--------:|----------:| |block | 5424| 0| 2897.31| 1577.02| 0| 2973.00| 5423| ▅▅▅▅▆▆▆▆▇▇| |weight | 21| 0| 2.00| 0.22| 1.67| 2.00| 2.50| ▂▁ ▇ ▁▁▁| |sum_t0 | 1873| 0| 221.54| 489.47| 0.00| 0.00| 3024.28| ▇▁| |sum_t1 | 1557| 0| 178.19| 441.99| 0.00| 0.00| 2784.64| ▇▁| |count_t0 | 108| 0| 11.01| 19.25| 0.00| 3.00| 106.13| ▇▁▁| |count_t1 | 116| 0| 9.44| 19.18| 0.00| 2.00| 115.00| ▇▁| |conditionC | 2| 0| 0.00| 0.50| -0.50| -0.50| 0.50| ▇▇| |sum_t0C | 2197| 0| 0.00| 693.65| -241.85| -241.85| 40037.32| ▇| |count_t0L | 220| 0| 1.56| 1.34| 0.00| 1.39| 6.26| ▇▅▄▃▃▂▁| |count_t0LC | 220| 0| 0.00| 1.34| -1.56| -0.18| 4.70| ▇▅▄▃▃▂▁| |count_t1L | 273| 0| 1.35| 1.34| 0.00| 1.10| 6.48| ▇▄▃▃▂▁▁| |sum_t0L | 1873| 0| 2.50| 2.89| 0.00| 0.00| 8.01| ▇▂▁▁▁▁| |sum_t0LC | 1873| 0| 0.00| 2.89| -2.50| -2.50| 5.52| ▇▂▁▁▁▁| |count_t1LC | 273| 0| 0.00| 1.34| -1.35| -0.25| 5.14| ▇▄▃▃▂▁▁| # models ```r > m <- feglm(sum_t1 ~ conditionC * sum_t0LC | block, dt1[domain_type == "overall"], family = "quasipoisson", vcov = "HC1") NOTE: 1,017 fixed-effects (5,658 observations) removed because of only 0 outcomes. > m GLM estimation, family = quasipoisson, Dep. Var.: sum_t1 Observations: 27,230 Fixed-effects: block: 4,407 Standard-errors: Heteroskedasticity-robust Estimate Std. Error t value Pr(>|t|) conditionC 0.076275 0.079198 0.963082 0.335517 # treatment group shared more bad content? sum_t0LC 0.203863 0.012149 16.779655 < 2.2e-16 *** conditionC:sum_t0LC -0.034560 0.021002 -1.645578 0.099864 . --- Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 # winsorize time0 and time1 summed badness, 99th percentile dt1[, sum_t0 := Winsorize(sum_t0, probs = c(0, 0.99))] dt1[, sum_t1 := Winsorize(sum_t1, probs = c(0, 0.99))] > m <- feglm(sum_t1 ~ conditionC * sum_t0LC | block, dt1[domain_type == "overall"], family = "quasipoisson", vcov = "HC1") NOTE: 1,017 fixed-effects (5,658 observations) removed because of only 0 outcomes. > m GLM estimation, family = quasipoisson, Dep. Var.: sum_t1 Observations: 27,230 Fixed-effects: block: 4,407 Standard-errors: Heteroskedasticity-robust Estimate Std. Error t value Pr(>|t|) conditionC -0.012194 0.046037 -0.264864 0.79112 # treatment group shared less bad content? sum_t0LC 0.232626 0.008021 29.002608 < 2.2e-16 *** conditionC:sum_t0LC -0.002530 0.011939 -0.211888 0.83220 --- Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 ``` # user CDFs Only showing time1 summed badness up to 5000 (the max value is actually 44559.55). - x: user's summed badness - y: proportion (bottom: proportion difference) More users in the treatment condition shared less bad stuff (badness < 500). But also more users in the treatment condition shared more bad stuff (badness > 1200)? Backfire? ![[dv_fc_cdf_summedbadness 2.png|800]] Winsorize 99th percentile ![[dv_fc_cdf_summedbadness_winsorize.png|800]]