causal RD Flashcards
What is the key identification assumption of RD? Gary
The continuous function of the running variable is correctly specified around the cutoff.
What bias does RD remove? Gary
Selection bias, although other biases may remain. Attrition, history, demoralization, diffusion of treatment.
Why does RD remove selection bias? Gary
The determinants of treatment status are “soaked up” by the assignment variable, breaking any channels with treatment and error due to selection.
Is instrumentation a problem with RD? Gary
Not really since what matters is that the forcing variable is responsible for assignment to treatment, we don’t care if what’s measured has construct validity, only that measurement determines assignment.
What are Gary’s requirements of RD?
Assignment to treatment is quantitative and continuous; no other changes at cut off; treatment occurred; no hidden treatments ( no discontinuities other than the target cut); plausible theory of action; effects are robust for various bandwidths; baseline equivalence of covariates
What are the reporting expectations of RD? Gary
graphical analysis of outcome variable and forcing variable is displayed; evidence that proper functional form was used; slopes on either side of cut must be allowed to vary; report number of units above and below cutoff
Under what circumstances do we add covariates to an RD model? What changes should the covariates induce? What do these changes mean for the analysis? Gary
Add covariates to increase precision. Adding covariates should only alter the standard errors (unless it’s something like a FE), if there were no baseline differences. If there are baseline differences then adding covariates only controls for the average effect of the covariate for both groups, but if the average effect for each group differs then IT’S IMPOSSIBLE TO DISENTANGLE HETEROGENEITY IN COVARIATES FROM TREATMENT EFFECTS!
Who are the godfathers of RD? Who said to use pre-tests in RDD?
Imbens, Lemieux 2008. Wing, Cook, 2013.
