models
- each user can have up to 8 rows/days of data (but fewer if they weren't active on all 8 days)
- see [[220518_170142 campaign 3 daily analysis and results|here]] for distribution of effects for each day
```r
# summed and count models
# SE cluster on block & day
feglm(t1 ~ conditionC * t0LC | block + day, dt1, family = "quasipoisson")
# HC1
feglm(t1 ~ conditionC * t0LC | block + day, dt1, family = "quasipoisson", vcov = "HC1")
# fraction models
# SE cluster on block & day
feols(t1 ~ conditionC * t0LC | block + day, dt1)
# HC1
feols(t1 ~ conditionC * t0LC | block + day, dt1, vcov = "HC1")
```
summary
- effects are much stronger than [[220417_110655 cluster block vs hc1|original analysis]]
- HC1 SE (cyan in figures below) is better than clustering on block **and** day (red below) (previously we only had 1 fixed effect [block] and clustered SEs only on that)
- clustering on block was better in original analysis
# summed badness
![[_sum_winsorize-0.95-clustse 2.png]]
![[_sum_winsorize-0.85-clustse 2.png]]
# count badness
![[_count_winsorize-0.85-clustse 3.png]]
![[_count_winsorize-0.95-clustse 1.png]]
# fraction badness
![[_frac_winsorize-0.95-clustse 3.png]]
![[_frac_winsorize-0.85-clustse 2.png]]