The design was a predictive study with predictor measures (theory-based cognitions) measured by a single postal questionnaire survey during the 12 month period to which the behavioural data related. Two interim outcome measures of stated intention and behavioural simulation were collected at the same time as the predictor measures. Behavioural data was collected from routinely available prescribing data.

Study participants were a random sample of GPs from Scotland selected from a list of all Scottish general practitioners by a statistician using a list of random sampling numbers.

Theoretically derived measures were developed following the protocols of Ajzen 23, Bandura 2627, Connor and Sparks 33, Moss-Morris 34, and Francis et al. 35. The cognition questions were developed from initial interviews with 14 GPs in Scotland who took part in a semi-structured interview of up to 40 minutes, as recommended for the theory of planned behaviour. The interviews use standard elicitation methods and covered the views and experiences about managing patients with an URTI. Responses were coded into belief domains (behavioural, normative, control) which were then used, in conjunction with the literature, to create the questions measuring constructs. Five knowledge questions were developed by the study team based on issues for which there was good evidence. Appendix 1 provides a summary of the predictor measures used in this study (see also 24); the instrument is available as Additional File 1. Unless otherwise stated, all questions were rated on a seven-point scale from Strongly Disagree to Strongly Agree. We aimed to include at least three questions per psychological construct.

The three outcome measures were significantly correlated with each other: for Behaviour and Behavioural Simulation, the Pearson r statistic was 0.17 (p = 0.013); similarly for Behaviour and Behavioural Intention it was 0.19 (p = 0.004); and for Behavioural Simulation and Behavioural Intention, it was 0.44 (p < 0.001).

The mean (SD) number of prescriptions issued was 57 (31) per 100 patients. The results of the correlation analyses are shown in Table 1. TPB attitudes, intention and perceived behavioural control, SCT risk perception, self-efficacy, action planning, OLT anticipated consequences, evidence of habitual behaviour and CS-SRM cause (chance/bad luck) significantly predicted the use of antibiotics to treat URTIs. For the Stage Model, the 167 GPs who endorsed that they had 'already changed my management of URTIs to try to avoid the use of antibiotics' issued a mean (SD) of 54 (30) prescriptions per 100 patients versus 66 (29) for the 52 GPs who endorsed any other response (mean difference (95%CI) = 11.8 (21.1 to 2.5), p = 0.014).

The results of the theory level analyses are shown in Table 1. The TPB explained 3% of the variance in behaviour, SCT explained 5%, and OLT explained 6%.

In the cross theory analysis, only evidence of habitual behaviour (OLT) was retained in the regression model, explaining 6% of the variance in the number of antibiotic prescription issued (Table 3).

Forty-five respondents for whom we also had behavioural data indicated that prescriptions had frequently been written/printed by someone else. Their mean (SD) number of prescriptions issued was 70 (31) per 100 patients compared to 55 (30) for respondents who answered anything else (p = 0.006). When the analyses were repeated excluding these respondents, there were no differences from the overall analysis.

In response to the five clinical scenarios, the respondents indicated that they would prescribe for a mean (SD) of 1.6 (1.2) cases. The median number of prescriptions issued was one with a range of zero to five. From Table 2, the constructs which predicted behavioural simulation (i.e., what GPs said they would do in response to the specific clinical scenarios) were: TPB attitudes, perceived behavioural control and intention; SCT risk perception, outcome expectancies, and self-efficacy; action planning; OLT anticipated consequences and evidence of habitual behaviour; CS-SRM time (acute/chronic), control (by treatment), cause (chance/bad luck); and knowledge.

The results of the theory level analyses are shown in Table 2. The TPB explained 31% of the variance in behavioural simulation, SCT explained 26%, II explained 6%, OLT explained 24%, and knowledge explained 4.5%. For the Stage Model, the 182 GPs who endorsed that they had 'already decided to change my management of URTIs to try to avoid the use of antibiotics' made a mean (SD) of 1.4 (1.1) decisions to prescribe versus 2.4 (1.3) for the 64 GPs who endorsed any other response (mean difference (95%CI) = -1.0 (1.2 to -0.7), p < 0.001).

In the cross theory analysis, perceived behavioural control (TPB), evidence of habitual behaviour (OLT), CS-SRM cause (chance/bad luck), and intention were retained in the regression model, together explaining 36% of the variance in the scenario score (Table 3).

With the range of possible scores for intention of 3 – 21, the mean (SD) intention score was 6.5 (2.5); the median intention score was 6 with a range of 3 to 14. The constructs which predicted behavioural intention were: TPB attitudes, perceived behavioural control; SCT risk perception, outcome expectancy, self-efficacy; OLT anticipated consequences, evidence of habitual behaviour; CS-SRM time (cyclical), control (by treatment and by doctor), consequences, coherence; and knowledge (Table 2).

The results of the theory level analyses are shown in Table 2. The TPB explained 30% of the variance in behavioural intention, SCT explained 29%, CS-SRM explained 27%, II explained 9%, OLT explained 43%, and knowledge and attitudes together explained 22%. For the Stage Model, the 188 GPs who endorsed that they had 'already decided to change my management of URTIs to try to avoid the use of antibiotics' had a mean (SD) intention score of 6 (2.3) versus 7.8 (2.6) for the 66 GPs who endorsed any other response (mean difference (95%CI) = -1.8 (2.5 to -1.3), p < 0.001).

In the cross theory analysis, OLT evidence of habitual behaviour, TPB attitudes, risk perception, CS-SRM control by doctor, TPB perceived behavioural control, and CS-SRM control by treatment were retained in the regression model, together explaining 49% of the variance in intention (Table 3).

The management of URTI is a frequent behaviour, and our measure of self-reported habitual behaviour consistently predicted our outcome measures. Looking across our three outcome measures, there are also suggestions that issues of perceived control, risk perception, and attitudes may also be important.

The theories individually explained a significant proportion of the variance in our dependent variables, but the aggregated analysis suggested that they were measuring similar phenomena within their own individual structures. Our measure of habit was consistently identified as important, a finding that was supported by the result of the Stage Model analysis (albeit analysed as only two stages) which suggested that many GPs had already decided to prescribe fewer antibiotics. Because encouraging the implementation of any evidence-based practice commonly entails various methods of increasing knowledge, knowledge was included as a predictive construct in this study. The knowledge measure included questions about both how and why antibiotics might be used in the management of URTIs. The number of questions answered correctly was not related to the number of antibiotic prescriptions issued but was related to the behavioural simulation and intention scores. However, knowledge did not enter into any of the three stepwise regressions, indicating that other constructs are consistently more important and suggesting that behaviour change strategies aimed at changing knowledge alone are unlikely to be successful in this clinical area.

The stepwise regression analyses revealed that the main construct driving GPs' management of URTI was habit with additional influence from control, attitudes, and risk perception. Taken together, the results suggest that GPs have considered this frequently performed behaviour and operate in a predominantly habitual manner backed up by beliefs that support their habit.

This is a correlational study, so the causative aspects of the theories remain untested in this population; but it is promising for the utility of applying psychological theory to changing clinical behaviour that the constructs are acting as the theories expect. These results suggest that an intervention that specifically targets these elements should have the greatest likelihood of success in influencing the implementation of this evidence-based practice.

We used a range of theories and models in both this and another component 25 of our larger study 24. However, across the two studies of different behaviours (URTIs, taking dental radiographs) and different clinicians (GPs, dentists), different constructs predicted different proportions of the variance in the intention and behaviour. This raises the question of what would be an optimum core set of measures if the aim was to cover most behaviours and clinical groups. Given our current limited understanding, this would have to be the subject of both studies replicating this one and further work examining different combinations of theories and models.

Operationalising the constructs with theoretical purity was a challenge. The preliminary study revealed that it was difficult to ask clinicians about their control over prescribing antibiotics because they believed that, even if they felt there were barriers to performing the behaviour, ultimately they had total control because they wrote the prescription. In the final questionnaire, this meant some questions had to be worded in terms of not doing the behaviour. There was some concern that not prescribing antibiotics may represent a range of alternative behaviours rather than being just a negative reflection of prescribing antibiotics.

A number of the models (OLT, II, CS-SRM) have not previously been operationalised in this way. OLT and II have usually being used as intervention methods to change behaviour. However, they both have been able to predict behavioural simulation, and OLT predicted intention and behaviour. The CS-SRM did not predict significant variance in behaviour or behavioural simulation, but the model did explain 27% of the variance in intention, a similar proportion to both TPB and SCM. The model has previously been used mainly to refer to an individual's perceptions of their clinical condition; we used it to measure a clinician's perception of the condition in general. We had difficulty operationalising this model, and further work is needed to explore how best the model can be applied to clinician's behaviour in respect of their patients.

One of the main strengths of this study is that the primary outcome was behaviour. The inclusion of the self-reported secondary outcomes of behavioural intention and simulation made it possible to examine the relationship between these three measures. This is important because behaviour is usually more difficult (and expensive) to measure than either of these proxy measures. By virtue of their significant correlation, the results suggest that self-reported measures have the potential to proxy behavioural data when testing an intervention prior to implementation in a service-level trial. However, although the two proxy measures (intention and simulation) were moderately correlated, the correlation between either and behaviour was weak. It is possible that the proxy measures are poor predictors of behaviour, though it is important to remember that the models we have used are focussing on modifiable behaviour. This cannot be quantified in our predictive study design but will only ever be a small proportion of behaviour. However, it is also important to consider the validity of our behaviour measure.

There is a stepwise decrease in the proportion of variance explained as we move from explaining intention to behavioural simulation to behaviour, with the models that we used explaining up to 49%, 36% and 6% of the variance respectively (Table 3). In a meta analysis of TPB studies in the general population, Armitage and Conner 38 reported TPB explaining 31% of the variance in self-reported behaviour and 20% in observed behaviour. Our data explaining up to 34% of the variance in behavioural simulation is very similar to Armitage and Connor's figure for self-reported behaviour, while our explaining up to 6% of the variance in behaviour is lower than their figure of 20%. In a parallel study using identical methods, we have been able to explain 16% of the variance in general dental practitioners' use of dental radiographs 25. This suggests that our operationalisation of the models was good, but that either the models do not work for this behaviour in GPs or there are problems with our measure of behaviour, or both. A systematic review 39 found only 10 studies exploring the relationship between intention and behaviour in healthcare professionals, but these reported explaining a similar proportion of the variance in observed behaviour to the studies in Armitage and Connor's review 38. This suggests that the problem is with our measure of behaviour.

There could be two problems – prescribing data may not be a good proxy for the behaviour as we asked about it in our questionnaires, or there may be biases within prescribing data. We have already identified the potential problem with using antibiotic prescription as a proxy for the management of patients with URTIs without prescribing antibiotics. Not only may 'prescribing' not be the reverse of 'not prescribing', 'not prescribing' may represent a number of alternate behaviours. Prescribing data was chosen because it was available from routine data sources, and was therefore inexpensive to collect. Antibiotic prescribing was chosen because it was more likely that a prescription for an acute illness (as opposed to a chronic illness managed through a repeat prescribing system) would be attributed to the GP who issued it. Scotland was chosen because the most commonly used computer system was likely to ascribe an antibiotic prescription to the issuing doctor. Despite this, we know that there are errors in the attribution of prescriptions to doctors, with 45 respondents reporting that prescriptions had frequently been written in their name by someone else. Finally, our standardisation by the number of patients registered with the GP assumes that each doctor has the same presentation rate. We sought to minimise variation in this by measuring over a 12-month period, but it is possible that this was still a problem. In future studies of this kind it will be important to invest more in the measurement of the behavioural data.

Our final response rate was not high compared to what would be expected for a postal questionnaire survey. Cummings et al. reported that up to 1995, response rates of surveys of healthcare professionals remained constant at approximately 60% 40. Our previous study using similar questionnaires to investigate antibiotic prescribing had a response rate of 68% 22. Kaner et al. reported doctors describing day to day work pressures and lack of perceived salience as reasons for not completing questionnaires 41. Since these three studies, day-to-day work pressures in UK NHS primary care have continued to rise, and our operationalisation of multiple models resulted in a long questionnaire asking seemingly repetitive questions. Additionally, our request to access behavioural data deterred 31 respondents who returned a completed questionnaire; it may have deterred a larger group from even completing a questionnaire.

Although we cannot make direct comparisons, our respondents appear well-matched with the overall population of Scottish GPs on gender, age and prescribing rates but came from larger practices and were more likely to be trainers. From publicly available data (see ISD Scotland 42) for 2003 and 2004, demographic data for all Scottish GPs were: 55% male, mean age 44.7 years (assuming qualification at age 23, this gives 21.7 years qualified), average practice size of 5,089, and 10% were trainers. Mean national rates of antibiotic prescribing in 2004 (having made, where possible, similar exclusions to those made in this study) was 65 prescriptions per 100 patients. Therefore, while we should be cautious about generalising from our respondents to the population of Scottish GPs, this is less of an issue at this exploratory stage of using these methods. Our aim was not to generate data that was representative but to receive our pre-specified number of responses from a population who had a range of behaviour, reported a range of behavioural simulation and intention, and who reported a range of cognitions. The study achieved this aim.