models
- fixed effect block & day
- compare cluster SEs on block vs user
- .85 winsorization seems to be the best in this dataset (most significant effects) and no winsorization is the worst (no significant results at all)
- **but initial analyses on the bigger 80k dataset suggest higher thresholds might be better in that dataset?** how much to winsorize might depend on dataset...
- clustering SE on user might be a little better than clustering on block, but not consistently the case
- effects are stronger than [[220417_110655 cluster block vs hc1|original analysis]] (without day fixed effects)
- there are significant effects even for fraction badness in this analysis, whereas original analyses there weren't any significant effects for [[220417_110655 cluster block vs hc1#fraction bad retweets|fraction badness]]
```r
# summed vs count model: quasipoisson
feglm(t1 ~ conditionC * t0LC | block + day, dt1, family = "quasipoisson", cluster = ~ block)
feglm(t1 ~ conditionC * t0LC | block + day, dt1, family = "quasipoisson", cluster = ~ user)
# fraction model: ols
feols(t1 ~ conditionC * t0LC | block + day, dt1, cluster = ~ block)
feols(t1 ~ conditionC * t0LC | block + day, dt1, cluster = ~ user)
```
## winsorize at 0.85 and 0.95
- [[220531_173501 winsorize 90 and 99|see here for winsorizing at 0.90 and 0.99]]
## summed badness
![[_sum_winsorize-0.85-clustse 3.png]]
![[_sum_winsorize-0.95-clustse 3.png]]
## count badness
![[_count_winsorize-0.85-clustse 1.png]]
![[_count_winsorize-0.95-clustse 3.png]]
## fraction badness
![[_frac_winsorize-0.85-clustse 3.png]]
![[_frac_winsorize-0.95-clustse 2.png]]