RCTs Flashcards
(9 cards)
What are some implementation problems with RCTs?
1) Non-compliance: Actual treatment is often endogenously chosen by potential participants -> ITT
2) Attrition: If attrition is
correlated with treatment, we may have biased estimates despite randomization.
3) Experimental Effects: Hawthorne Effect, John Henry Effect, Survey Effects
What does Beta_0 mean in the following DD regression?
Y_it = β_0 + β_DDi + β_PostPost_t + β_DD(D_i × Post_t) + U_it
Mean outcome in control areas before the intervention
What does Beta_post mean in the following DD regression?
Y_it = β_0 + β_DDi + β_PostPost_t + β_DD(D_i × Post_t) + U_it
Change in the mean outcome in control areas
What does Beta_D mean in the following DD regression?
Y_it = β_0 + β_DDi + β_PostPost_t + β_DD(D_i × Post_t) + U_it
Difference in mean outcome between treatment and controls before the intervention
- we expect this to be zero if actually randomized
What does β_Post + β_DD mean in the following DD regression?
Y_it = β_0 + β_DDi + β_PostPost_t + β_DD(D_i × Post_t) + U_it
Change in the mean outcome in treatment areas
What does β_DD mean in the following DD regression?
Y_it = β_0 + β_DDi + β_PostPost_t + β_DD(D_i × Post_t) + U_it
Difference in the mean change between treatment and controls (the DD!)
What is the key assumption of DD?
common trends between treated and control areas in the absence of the treatment.
Why collect data at baseline in an RCT?
1) Balance check (i.e. testing the equality of the means for treated, non-treated groups pre-treatment)
2) Information
3) Including controls - check for lack of balance, and can improve precision of estimates
4) Enforce balance using stratification
5) Heterogeneity analysis
What happens to standard errors if you do standard OLS for a clustered dataset?
You will massively understate the variance, leading to lower standard errors and mistaken rejection of null
