Fig. 1

Traditional vs. Modern Generalization. A traditional approach to generalizing take effects observed in a study and seeks to apply those effects in a larger population external to the study (Panel A). If the underlying units of interest (whether patient, providers or health care units) is well characterized, the sampling probabilities into the study are known, typical threats to validity are adequately managed (e.g., measurement error), and the study itself does not create an artifactual environment, findings can be used to infer (with statistical uncertainty) about effects in the external population and contexts. Implementation research often assumes, in contrast, there is a meaningful diversity of contexts in the real world. This implies that the effects of any implementation strategy will differ across those contexts. Instead of identifying a single effect that applies in all contexts, the field may need to seek effects in one context in a way that enables inferring in other external contexts (Panel B), even when the effects will differ. We seek an approach to generalizing such that a study (Panel B) in one of the three contexts (Context A) can be used to infer about effects in other contexts such as Context B (where the strategy improves outcomes by threefold) or Context C (where the strategy has no effect). Is that possible?