Reading this post by Andrew Gelman got me thinking: how do we really know if a variable is an instrument? Demonstrating that a variable is an instrument often seems to be a matter of telling a compelling clinical story but leaves one with strong and unverifiable assumptions.
This is a real issue with the Physician Preference instrument in pharmacoepidemiology where it has the potential to be either a major advance or a blind alley. But how would one know for sure? I suspect that it is more likely the former than the latter but sorting this out with any level of certainty isn't easy.
Any insights out there?
Post a Comment