19.2 Assumptions
- instrument is relevant
- instrument is valid
Relevance
- Z needs to explain X
- if correlation is small: weak instrument
- estimate likely to jump around from sample to sample
- if small, variation matters
- We can actually check for relevance: look at relationship b/w X and Z
Validity
Instrument has no back doors
Any paths b/w Z and Y need to pass through instrument, or be closed
here: need to close Z –> B –> Y and Z –>A –> Y
validity very hard to argue, e.g. with rain: correlated to many things
choosing good instruments:
as-good-as-random: truly comes from outside the system, applied almost completely randomly (real randomization / Mendelian randomization)
you wouldn’t ever think to include it in your regression
randomness is more believable (e.g. RCT with imperfect compliance, RDD)
Monotonicity
- no defiers (see IV and treatment effects)