This study demonstrated that TPB can be used as a tool for theory-based process evaluations with the aim of investigating possible causal mechanisms in KT intervention studies when the intervention is hypothesised to be mediated by the constructs of TPB. There were differences in intention, attitude, and subjective norm to FSH and ferritin test requesting, suggesting that the intervention may have enhanced attitudes and subjective norms resulting in higher intention and subsequent behaviour changes. Indeed, mediational analysis was highly suggestive that the differences in test requesting behaviour between trial groups were mediated through intention. There were high intentions, subjective norms, and attitudes for HPS requesting, suggesting that there may have been a psychological ceiling effect resulting in the observed lack of effect on test requesting behaviour in the trial.
This study had several strengths. First, the main trial demonstrated strong intervention effects (behaviour change), so provided an ideal platform to investigate why change did or did not occur. In particular, the randomisation element provided the opportunity to robustly investigate whether intentions mediated the trial result. Second, use of a well-established psychological model (TPB) enabled the psychological constructs to be clearly defined. Third, the derived measures of psychological constructs were sensitive to group allocation, suggesting that the constructs were identifying real differences. Finally, the TPB survey was returned completed from 42 of the 43 practices, suggesting that the results were generalisable.
The mediational analysis suggested that intentions to request an FSH or ferritin test were part of the causal pathway in the trial, i.e., the observed trial reduction in test requesting was partially mediated by a change in intentions. In our experience, formal mediational analyses have been rarely used to investigate the causal factors in KT randomised trials, and we suggest investigators should make more use of theory-based process evaluations.
Given that responses were received from several primary care doctors within a practice, we were able to demonstrate that there was clustering of psychological constructs within practices. Behavioural intentions and attitudes to test requesting had intra-cluster correlations greater than 0.1. This clustering provided some empirical evidence that social or organisational factors within practices may influence test-requesting behaviour. The clustering also needs to be considered from a statistical power perspective when conducting such surveys in the future. The effects of clustering are that precision is reduced and confidence intervals are wider than if clustering were not present. The surveys therefore need a larger sample size to attain the level of precision that investigators are interested in .
Whilst nearly all practices (42/43) were represented in the final survey, only 56% of the primary care doctors within those practices responded. This response rate from individual primary care doctors is very similar to that of other surveys of health professionals . We cannot however be sure that the responders' views are representative of the practice, but the response rates were the same across the trial intervention groups suggesting that the results were not biased. Further, use of different measures of aggregated practice intention acted as a form of sensitivity analysis on the influence of different aggregation methods on the study results .
In this study, our behavioural outcome was practice level test requesting. Ideally, to operationalise TPB model faithfully, the outcome would be individual practitioner-level requesting. A multi-level model analysis could then be used to account for any clustering of behaviour or behavioural predictors by practice. However, it was not possible to obtain data on individual primary care doctors' requesting patterns from the administrative data systems. The implication for statistical analysis was that some measure of practice-level psychological cognitions had to be derived. An obvious summary measure is the mean cognition of the primary care doctors within each practice . Using the mean cognition, intentions predicted about 8% of the variability in FSH and ferritin testing. Because the psychological measures were generally high with little variability, an alternative summary measure (the minimum) was considered. The minimum predicted about 12% of the variability in FSH and ferritin testing. The observed lack of relationship between HPS testing behaviour and intentions to request was likely due to the ceiling effect in intentions, but could also be due to the insensitivity in the behavioural measure. That is, whilst the intervention (and therefore the scenario description in the questionnaire) targeted requesting of the tests in specific clinical circumstances, the information system cannot distinguish between specific clinical circumstances (e.g., for HPS repeat testing after eradication therapy, the measure of behaviour was all HPS test requests because the information system cannot distinguish between initial tests, repeat-tests, and does not identify the reasons for the request). Therefore, our dependent variable may not exactly match the context of the intervention and scenario. Our findings and future investigations of causal mechanism would be strengthened by individual, context specific, measures of behaviour. We would recommend that researchers consider conducting a sensitivity analysis on any summary measure of psychological cognitions when attempting to describe a group level behaviour 
No formal statistical power calculation was performed for the survey. The confidence intervals for the testing of the constructs that were predictive of intentions (Table 3) and the models predicting behaviour-using intentions (Table 4) demonstrated the study was adequately powered to detect important effects. For the results of the mediational analysis the study was underpowered. This was due mainly to only 50% of the original study practices taking part in the survey. This meant that the original study findings on test-requesting behaviour could not be replicated with the same precision (though the magnitude of effects were similar). We would recommend that future studies of mediational factors in KT trials conduct a formal sample size to ensure adequate power for the theory based process evaluation.
We investigated behavioural predictors at one time point after initiation of the intervention i.e., the survey was conducted after the study interventions had been delivered for 12 months. In this example, the difference in constructs scores between intervention and control practices were large and provided evidence of changes in construct. Future process evaluations may be augmented by the addition of pre-intervention measures of behavioural predictors. Furthermore, the results of this study provide some evidence that TPB could be used to design an intervention. The ceiling effect on the intention to request HPS tests suggests that an intervention targeting a primary care doctors' intention to request the test would likely fail. In the context of the trial reported here, this would have suggested that feedback and reminders might not have been effective interventions to use and that was indeed the trial finding.
The aims of process evaluation alongside randomised trials of complex interventions are numerous (e.g., fidelity of implementation; mechanisms, mediators, and the process of change; acceptability) and often encompass a range of methods [17–19]. There are few RCTs of professional behaviour change strategies that utilise theory to investigate the process of change . Whilst TPB seems to be the most commonly applied social cognition model for investigating health professional behaviour, few studies have attempted to predict clinical-related behaviour . The results of this process evaluation utilising theory, re-enforces that TPB seem an appropriate theory to predict health professional behaviour change , and that it may offer useful insight into the processes underlying change (trial effects) in KT trials .