- [[230222_171403 compare campaigns|compare covariates across campaigns]]
# interact treatment effect with extra covariate `t0propSC` (proportion active days before campaign)
Pre-campaign proportion of days with at least 1 tweet/retweet/quote/reply.
The `t0prop` variable is very very skewed and hard to make it normal - applied `sqrt` transform and then median-centered it (`t0propSC`). The skewness makes the `conditionC` simple effect harder to interpret (since estimates depend on how we center the covariates).
See `conditionC:t0propSC` effects below. Mixed evidence.
- exp1 and exp2: negative interaction?? (treatment worked better for users with higher t0 activity)
- exp3 and exp4: positive interaction?? (treatment was less effective for users with higher t0 activity)
## exp1 pilot
![[_count_winsorize-0.95-clustse 9.png]]
![[_sum_winsorize-0.95-clustse 9.png]]
## exp2 deepstate
![[count-winsorize-0.95-clustse 15.png]]
![[frac-winsorize-0.95-clustse 13.png]]
![[sum-winsorize-0.95-clustse 16.png]]
## exp3 trucker
![[count-winsorize-0.95-clustse 17.png]]
![[frac-winsorize-0.95-clustse 15.png]]
![[sum-winsorize-0.95-clustse 18.png]]
## exp4 deepstate
![[count-winsorize-0.95-clustse 18.png]]
![[frac-winsorize-0.95-clustse 16.png]]
![[sum-winsorize-0.95-clustse 19.png]]