The primary purpose of this study was to examine the potential mediators of behavior change in a worksite-distributed PA intervention . Potential mediators included constructs of the Theory of Planned Behavior, Protection Motivation Theory, the Transtheoretical Model, and Social Cognitive Theory allowing for a broad assortment of social-cognitive constructs. The results suggest a general failure of these constructs to account for mediation, and several potential reasons for this null effect are discussed below. We believe the results of this study will inform future research efforts.
Of the 14 constructs and 42 tests (including initial change, delayed change, and overall change) only two positive results were identified (i.e., overall change in pros, initial change in experiential processes), and these were of very small effect size . Adjusting the alpha to account for the multiple tests, eliminate any mediation effect in this study.
Thus, it is clear that the intervention did not change these mediators in any substantive way. Low fidelity observed in our current interventions has also been found among many other trials (see Lewis et al. , Lubans et al, , and Rhodes and Pfaeffi  for reviews). In a recent study  somewhat similar to ours also reported that a lifestyle physical activity intervention in women effectively increased physical activity, but none of the proposed psychosocial constructs showed a mediating effect. In a recent review of 28 primary prevention PA trials, Rhodes and Pfaeffli  found that over half of the trials failed to change the proposed mediators of the intervention.
Clearly, continued innovation to increase the power of interventions is needed to bring about change in social-cognitive variables. This problem poses less of a challenge to current theories as it does to interventions; nevertheless, it does suggest that social-cognitive constructs may be difficult to change following interventions. Interestingly, in our study, several of the changes in social-cognitive constructs were related to subsequent changes in PA independent of the intervention arm. Thus, there was some evidence of validity for these mediators despite the intervention's failure to change these constructs. However, there are practical limitations to changing these constructs that also need consideration. For example, many of the constructs demonstrated very high mean values . This suggests a ceiling/threshold for the possibility of change. Fishbein et al.  highlight the importance of demonstrating 'room for change' in psychological constructs as an important step in interventions. There are also other potential methodological artefacts such as measurement times used (too long or too short to capture changes), and attenuation of measurement error. Another explanation for the observed small changes may be explained by response shift theory . Response shift theory refers to a change in one's self-evaluation and/or internal standard (e.g., attitude, perceived control) that result from a change in a measured variable (e.g., PA behavior). An individual for example, may have high perceptions of control before engaging in PA. However, when the individual initiates the behavior, perceptions of control may diminish as barriers to PA and other related constraints may increasingly inhibit the ability to perform PA. Although to date response shift theory has not been evaluated in the social-cognitive domain, this theory may provide one explanation for the observed positive behavioral changes (i.e., PA) while cognitive variables remained relatively stable.
In our study, there was an intervention effect on behavior, despite the intervention's general inability to change the examined psychological mediators. A failure to show mediation by social-cognitive variables despite behavior changes, is common [8, 36–39]; indeed most intervention effects on behavior have not been completely accounted for by the proposed mediators [8, 34]. Further, a recent study  that reviewed mediators of dietary behavior change observed that interventions were relatively unsuccessful in changing mediators. The most powerful and consistent mediator of behavior change interventions in past research has been the behavioral processes of change [8, 34], however, this construct was ineffective in the current trial.
Similar measurement (attenuation, response shifts) and method (e.g., time-frame) artefacts as highlighted above may account for this result. Still, the finding is suggestive that our current psychological constructs may be inadequate to account for behavioral changes. This also reflects correlational research where even our best prediction models typically explain less than 30% of the variance in PA . An expansion of current models to include additional relevant constructs seems prudent yet most of our current theoretical conceptions share considerable redundancy [7, 15, 26] and thus it is difficult to make suggestions based on the extant literature. Perhaps more sophisticated tests such as an examination of moderators such as personality and environmental characteristics (see Rhodes and Smith ) or more specific measurement domains (e.g., belief level vs. aggregation) would be helpful (see Vallance et al. ). It is also important to note that the behavior change effect in this study was small (2% variance). While this is typical of worksite interventions  and behavioral interventions more generally  the precision needed from the purported mediators to account for this very small effect may be difficult. However, it is important that physical activity interventions perform and report mediation analyses even if mediation is not established.
Finally, we must acknowledge the study limitations which include the use of self-report measures to assess physical activity, and the low level of participant retention throughout the 12 month study period. This should be taken into account in the interpretation of our findings.
In summary, an examination of mediators of PA behavior change using four leading social cognition theories was unable to account for the effect of the intervention on behavior in a worksite intervention. The results highlight the importance of reporting mediation results in randomized controlled trials whether the findings are null or positive. Future studies may wish to focus on more detailed dose-response issues between mediators and behavior, the inclusion of moderators that could affect individual change, or different mediator constructs at higher levels of measurement specificity. Overall, however, the results show a similar pattern to prior research where interventions did not result in a change in mediation of behavior. Continued work on innovative and more powerful intervention approaches to PA is timely.