This big paper by Angus Deaton on “instruments of development” is really good. In short,

• Lots of people are using instruments that are unlikely to satisfy the exogeneity assumption required for identification.
• Randomized controlled trials are not nearly as informative as we think, and when we push trial results to be more informative, we’re back in the world of instruments and econometrics.
• Economic theory needs to make a lot of progress too – in most cases, we need theory to do decent empirics.