Regression Discontinuity

Question 1

what data are we looking at?

Answer 1

focusing on groups close to the thresholds that are almost identical observationally

if you can account for trends on both sides, you have a smooth relationship and don’t expect to see big jumps anywhere
- if there’s a big jump, leading explanation is the program

Question 2

when can an RD be used?

Answer 2

whenever program/treatment assignment has a sharp discontinuity
- rapid change in the likelihood of assignment with respect to an underlying smooth variable

something we carry out when we don’t have randomisation
- with GSP awarded above a certain threshold, the winners and non-winners are clearly different

Question 3

RD limitations

Answer 3

effects can be estimated only locally at the threshold
- might care about people at other points but we can only look at the cutoff
- contrast with RCT since you can estimate the ATE for the whole population and the effect would be local to the population you randomise in

a lot of data required at the threshold

certain rules can be manipulated, leading to non-random “sorting” at the threshold
- many instances where people might be able to manipulate things or cheat to qualify for something
- can’t compare people right above the threshold to those right below in this sense

Question 4

advantages of RDs

Answer 4

only needs a single cross-section of data
- no need for repeated observations over time (which you need with a DD)

does not require setting up a randomised experiment in advance
- politically and logistically difficult