Suggestions for testing health behavior theories: Implications for mediator analysis

Abstract

Dr. Neil Weinstein’s excellent article on ‘‘misleading tests of health behavior theories’’ (1) has implications for mediator analyses of ongoing health behaviors. Dr. Weinstein refers to four types of data that can be collected around an ‘‘event’’ that would ‘‘allow one to determine whether changes in perceptions predict subsequent changes in behavior’’ (p. 5). These data would include measurements of (a) perceptions before the event; (b) perceptions immediately following the event, but before behavior change is possible; (c) behavior before the event; and (d) behavior some time after the event, such that perceptions have had a chance to change behavior (1). The strength of this design lies in the temporal ordering of measured constructs in a way that is congruent with the hypothesized causal path. Temporal ordering of constructs has also been called for in the context of mediator analyses (2–4) but is rarely achieved in practice (5). A typical design involves measurement of the proposed mediator(s) and the target behavior immediately prior to the intervention at Time 1. Hypothesized mediators are then measured at Time 2 and behavior again at Time 3. This has the appearance of a temporal ordering of constructs; however, findings could reflect change in both the hypothesized mediator and the targeted behavior that occurred between Times 1 and 2, leaving one unsure as to what change occurred first. One solution is to measure change in the hypothesized mediator from Time 1 to Time 2, and change in the behavior from Time 2 to Time 3, though this may not yield expected results, as change in behavior often occurs early on in response to the intervention. Another solution is to consider the intervention as ‘‘the event’’ and follow the guidelines provided by Dr. Weinstein. For example, physical activity behavior and self-efficacy, a hypothesized mediator, could be measured immediately prior to a 1-hr physical activity counseling session. Self-efficacy could then be measured immediately following the session, with physical activity measured several weeks later. In such a design, it could not be argued that change in behavior preceded change in self-efficacy. A similar approach could be taken with an intervention of longer duration (e.g., 1–2 weeks) if behavior change is discouraged by the researcher until after the intervention is complete (e.g., a smoking cessation intervention in which the quit day is set two weeks after the start of the intervention). It is important to note that while such a design would ‘‘suggest a causal relation’’ (1, p. 5) between mediator and behavior, it would by no means prove such a causal relation. Although random assignment to the intervention followed by differential change in self-efficacy and physical activity behavior would suggest a causal effect of the intervention on both self-efficacy and behavior, it would be possible that the intervention caused both the hypothesized mediator and the behavior directly, and=or that the behavior was changed by the intervention through its effects on some other unmeasured construct. As noted by previous authors (e.g., 6), because perceptions cannot be directly manipulated, this type of evidence may be as close as we are able to come to supporting the hypothesized mediational path.