,
3 tweets,
4 min read
Read on Twitter
@EpiEllie @causalinf @EpiDancer @MariaGlymour @autoregress @causalinf , to be clear, do you mean that IV is out of favor relative to other quasi-experimental methods like RD (I agree), or do you mean IV is out of favor relative to observational "back-door" type approaches?
@EpiEllie @causalinf @EpiDancer @MariaGlymour @autoregress And @EpiEllie, of course if you picked a random IV, confounders would be equally a concern as with "X". But the whole point of finding a quasi-experiment is to locate some portion of the variation in X that is well-understood so confounders can be credibly assessed.
@EpiEllie @causalinf @EpiDancer @MariaGlymour @autoregress Applying the BDC without a natural experiment amounts to pretending that you understand *all* the variation in X (generally impossible) as opposed to understanding a small part of the variation in "X" (hard, but doable with sufficient care in many contexts).
