increases in the cost of tobacco products, and bans on indoor smoking in public places—were supported by direct evidence from an RCT.
Because RCTs frequently are not possible for multicomponent interventions that deal with determinants at multiple levels (e.g., individual, family, health plan), evidence should be sought (or generated) from other designs, and some of the advantages of RCTs must be sacrificed. For example, randomized encouragement designs, in which individuals are randomized to be encouraged to take up an intervention, may allow for recruitment of a more representative sample since participants are involved in the choice of their treatment, but the level of certainty may be reduced. Cluster randomization trials may avoid the difficulties of recruiting individuals in complex settings (e.g., schoolrooms, workplaces, medical clinics, or even whole communities) while retaining randomized assignment of an intervention, but the cost or feasibility of identifying a sufficient number of groups can be prohibitive. Nonrandomized designs, with or without controls (quasi-experimental designs), may be the only feasible approach in situations where randomization is difficult or impossible. These include pre–post designs, interrupted time series, and regression discontinuity designs (elaborated in Appendix E). Such designs, when implemented well, can produce valuable evidence for effectiveness but require trade-offs with the level of certainty of causality and control of bias.
Two perspectives provide useful, complementary approaches on which researchers can draw to strengthen causal inferences, including approaches that do not involve experimentation. In the behavioral sciences, Campbell and colleagues (Campbell, 1957; Campbell and Stanley, 1966; Cook and Campbell, 1979; Shadish and Cook, 2009; Shadish et al., 2002) have developed a practical theory providing guidelines for ruling out confounders that may yield alternative explanations of research results. In the field of statistics, Rubin and colleagues (Holland, 1986; Rubin, 1974, 1978, 2005, 2008) have developed the potential outcomes perspective, which provides a deductive, mathematical approach based on making explicit, ideally verifiable assumptions. Shadish (2010), West and colleagues (2000), and West and Thoemmes (2010) offer full discussions and comparisons of these two perspectives.
Campbell and colleagues have considered the full range of experimental, quasi-experimental, and pre-experimental designs used by researchers in the behavioral sciences. They have identified an extensive list of threats to validity, which represent an accumulation of the various criticisms that have been made of research designs in the field of social sciences and are applicable to other fields as well (Campbell, 1988). Although Cook and Campbell (1979) describe four general types of threats to validity, the focus here is on potential confounders that undermine the level of certainty of