For average treatment effect on the treated (ATT), the estimates and SEs from individual studies are divided by proportion reach before meta-analyzing. Proportion reach for the four studies:
- 0.5600, 0.6180, 0.6429, 0.6254
- main text prior: `t(df = 3, mu = 0, sigma = 0.05)`
# 0) descriptives
* plot of proportion of active days pre-campaign by "wave"
* distribution of demrep (pooling over studies 1 3 4)
- [[230719_164713 descriptives]]
# 1) Main effect (and moderation by t0)
for thresholds in [40:5:80]:
* quasipoisson predicting W95 # bad link RTs using treated * scale(t0) for studies 1, 3, 4
-> frequentist meta-analysis of treated; and treated * scale(t0)
-> Treatment effect on the treated version (divide coeff and SE for each study by p(treated) then do meta)
- [[230718_135343 frequentist fixed-effects meta for studies 1 3 4]]
- [[230719_184036 bayesian fixed-effects meta for studies 1 3 4]]
For comparison to FB analyses:
* LMP predicting # badlinkRTs > 0 using treated x scale(t0) for studies 1, 3, 4
-> frequentist meta-analysis of treated; and treated x scale(t0)
- [[230719_183343 linear probability model - compare FB analyses - bayesian and frequentist]]
To assess robustness to winsorizing
* quasipoisson predicting W99 # bad link RTs using treated * scale(t0) for studies 1, 3, 4
-> frequentist meta-analysis of treated; and treated * scale(t0)
* quasipoisson predicting # bad links using treated * scale(t0) for studies 1, 3, 4
-> frequentist meta-analysis of treated; and treated * scale(t0)
- [[230719_183842 robustness to winsorizing-99]]
- [[230719_183914 robustness to no winsorizing]]
# 2) Moderation
for thresholds in [40:5:80]:
* quasipoisson predicting W95 # bad link RTs using treated * scale(t0) + treated * day# for studies 1, 3, 4, restricting to days 1-8 for all campaigns
-> frequentist meta-analysis of treated * day#
* quasipoisson predicting W95 # bad link RTs using treated* scale(t0) + treated * day# for studies 1, 3, 4
-> frequentist meta-analysis of treated * day #
- [[230727_120706 treatment effect moderation by day - days 1 to 8]]
- [[230728_10706 treatment effect moderation by day - all days]]
* quasipoisson predicting W95 # bad link RTs using treated * rep for studies 1, 3, 4
-> frequentist meta-analysis of treated * rep
* quasipoisson predicting W95 # bad link RTs using treated * scale(t0) + treated * rep for studies 1, 3, 4
-> frequentist meta-analysis of treated * rep
- [[230727_150706 treatment effect moderation by repub]]
- quasipoisson predicting W95 # bad link RTs using treated * political (or not) for studies 1, 3, 4
- [[230729_140939 treatment effect moderation by political or not]]
# 3) Effect on all RT
* quasipoisson predicting W95 # RTs using treated * scale(total # all RTs in t0) for studies 1, 3, 4
-> frequentist meta-analysis of treated
- [[230723_184513 total retweets - bayesian and frequentist]]
# 4) Low-activity users experiment
* quasipoisson predicting W95 # bad link RTs using treated * scale(t0) for study 2
* LMP predicting # badlinkRTs>0 using treated * scale(t0) for study 2
- [[230723_190621 study 2 count models and linear probability models]]
---
# other analyses
- heterogeneity of treatment effect across campaigns? not possible with fixed-effects models, since any variation between studies is due to sampling error alone
- [ ] check study2 t1 and t0; compare with other studies