This is because my brain is still in policy evaluation mode, but I thought this was quite insightful and well expressed:
"The crucial difference in avoiding bias is not whether the assignments are random or nonrandom, but whether the investigator has knowledge of, and can model, the selection process"
- Wilbert van der Klaauw
That really strikes at the heart of people who scoff at non-experimental science. What's funny is that I think non-experimental science is a lot more wide-spread in the natural sciences than many social science skeptics understand, and the goal of those non-experimental scientists is very similar to our goals.
If you are ever interested in using regression discontinuity designs, click the link. Van der Klaauw's work is very clear at laying it all out. I had several papers at my side on the econometrics of it all too, but his paper provides the description of the method and an application - I think it's one of the best expositions on it out there.