External validity (or generalizability) of experimental research findings is arguably one of the most contentious areas in both natural sciences and development economics. Vox points to this meta-analysis of cancer risks (or benefits) by Jonathan Schoenfeld and John Ioannidis from various food items which finds the results inconclusive on pretty much everything.
Eva Vivalt does a similar meta-analysis of nearly 600 impact evaluation studies that examined the results of 20 development interventions and finds similar inconclusiveness with most of these interventions.
Clearly, even after rigorous impact assessments in sanitized environments with strong (atleast, most likely, better than business-as-usual) monitoring, the results of many of these interventions are not definitive enough to be embraced as unqualified policy prescriptions. At least not significantly more compelling than they were ab-initio from logical analysis. Further, as I have blogged earlier, the outcomes are likely to be far less benign when these interventions are scaled up on a business-as-usual implementation environment.