9: Randomised Control Trials (RCTs)

Question 1

actual vs natural experiments

Answer 1

RCTs vs IV, DiD, RD, etc.

Question 2

basic logic of RCT

Answer 2

kills the covariance between the explanatory variable of interest and anything else that might determine the outcome

randomly selecting treatment and control groups, then comparing average outcomes after treatment
- randomisation eliminates selection bias

Question 3

what does randomisation imply?

Answer 3

ATE = ATT

elimination of selection bias
- conditional expectations are the same as unconditional expectations

Question 4

ATE

Answer 4

E[Yi(1) - Yi(0)]

Question 5

important caveats when evaluating RCTs

Answer 5

sample attrition

hawthorne effects

spillover on control group

small scale effects vs. scaling up

external validity

less than full compliance and ITT

Question 6

sample attrition

Answer 6

bias in the estimate of the causal effect occurs if intervention leads to a differential drop out of the estimation sample between treatment and control
- sample selection bias

e.g. cash transfers to households in rural areas but those who use the money most productively move to the city

Question 7

hawthorne effects

Answer 7

bias created if subjects change their behaviour since they know they are studied as part of an experiment
- bigger problem if it occurs in treatment rather than control

Question 8

spillovers on the control group

Answer 8

control group is no longer a reference group in which nothing changed

e.g. cash transfers that increase overall price levels

Question 9

small scale vs large scale effects

Answer 9

general equilibrium effects can change outcomes when scaling up
- unknown effects that lead to different outcomes since everything is connected

e.g. cash transfers that increase house prices, food prices, wages, trade flows, etc. when scaled up