Applying psychological theories to evidence-based clinical practice: identifying factors predictive of lumbar spine x-ray for low back pain in UK primary care practice

Theoretically-derived measures were developed following standard operationalisation protocols wherever possible [21, 29‐33]. The cognition questions were developed from semi-structured interviews with 18 primary care physicians in Scotland and north England that lasted up to 60 minutes. The interviews use standard elicitation methods and covered physicians' views and experiences about managing patients with low back pain. Responses were used to create the questions measuring constructs. Five knowledge questions were developed by the study team based on issues for which there was good evidence. Table 1 provides a summary of the predictor measures used in this study (see also [28]); the instrument is available as Additional File 1. Unless otherwise stated, all questions were rated on a 7-point scale from 'strongly disagree' to 'strongly agree.' We aimed to include at least three questions per psychological construct.

Participants were mailed an invitation pack (letter of invitation, questionnaire consisting of psychological and demographic measures, a form requesting consent to allow the research team to access the respondent's referral data, a study newsletter, and a reply paid envelope) by research staff. Initially, 700 primary care physicians were surveyed between July and mid-August 2003. Due to a low initial response rate, a further sample of 400 primary care physicians were surveyed between October and December 2003. Two postal reminders were sent to non-responders at two and four weeks. Behavioural data were collected over a one-year period, from approximately six months before to six months after the assessment of cognitions.

The target sample size of 200 was based on a recommendation by Green [35] to have a minimum of 162 subjects when undertaking multiple regression analysis with 14 predictor variables.

The internal consistency of the measures was tested using Cronbach's alpha. If this was less than 0.6, then questionnaire items were removed from each measure to achieve the highest Cronbach's alpha possible. For constructs with only two questions, a correlation coefficient of 0.25 was used as a cut off.

For each of the three outcome variables, we examined the relationship between predictor and outcome variables within the structure of each of the theories individually. Spearman's correlation (for behaviour outcome) and Pearson Correlation Coefficients (for behavioural simulation and intention outcomes) between the individual constructs and the outcome measures were calculated. Given the distribution of the behavioural data, we used negative binomial regression (NBR) to model the predictive ability of individual theoretical constructs and complete theories. NBR is used to model count exhibiting over dispersion, as in the case of the behaviour outcome data in this study. We reported incidence rate ratios (IRR) from the NRB models. IRRs estimate the change in the rate of the dependent variable associated with changes in the independent variables. NBR does not generate a direct equivalent of an R² statistic to estimate the proportion of variance in the dependent variable explained by models. However, it is possible to compute a number of different R² statistics to explore the goodness of fit of the model [36]. The pseudo-R² we chose to use was McFaddens' adjusted R² because it penalizes models in the spirit of adjusted R² in linear regression for adding more variables to a model (see Additional File 2 for further discussion). Linear regression was used for intention and behavioural simulation. For the five 'perceived cause of illness' questions in the CS-SRM, responses were dichotomized into scores of five to seven (indicating agreement that the cause in question was responsible for low back pain) versus anything else (indicating disagreement). These dichotomous variables were then entered as independent variables into the regression models. The relationship between II and intention was not explored as it is a post-intentional theory. For the analysis of the PAP, respondents were dichotomized into two groups (decided to reduce or have already reduced x-rays versus other responses) and the relationship between predictive and outcome variables were examined using regression models. Finally, for predictors p < 0.25 irrespective of whether or not they came from the same theory, we conducted a cross-theoretical construct analyses that examined the relationship between predictive and outcome variables.

The study was approved by the UK South East Multi-Centre Research Ethics Committee (MREC/03/01/03).

The three outcome measures were significantly (though weakly) correlated with each other: for behaviour and behavioural simulation, the Spearman's rho statistic was 0.169 (p = 0.004); similarly for behaviour and behavioural intention it was 0.165 (p = 0.005); and for behavioural simulation and behavioural intention the Pearson's r was 0.313 (p < 0.001).

The mean number of lumbar spine x-rays was 5.0 per 1,000 patients registered per year. The results of analyses are shown in Table 3. Individual construct analyses suggested that constructs from TPB (attitudes, intention, and perceived behavioural control), SCT (risk perception, self efficacy), OLT (anticipated consequences) and CS-SRM (cause - aging) significantly predicted the lumbar spine referrals. To aid interpretation of the results, we provide the following example; intention had a mean score of 2.1 (SD 1.0), the IRR was 1.29 -- this suggests that for every point increase in intention (equivalent in this example to one SD), lumbar spine referrals would increase by 29.0%. Theory-level analyses (Table 3) suggested that TPB (perceived behavioural control), SCT (risk perception), OLT (anticipated consequences), CS-SRM (control - by patient, cause - poor prior medical care, cause - patients' own behaviours, cause - aging) predicted behaviour. II, PAP, and knowledge did not predict behaviour. However, the goodness to fit measures suggested that the theoretical models did not predict behaviour data in this dataset (McFadden's pseudo R² range from 0 to 0.004, see also Additional file 2 for addition goodness to fit measures). In the cross-theoretical construct analysis, constructs from TPB (attitudes) and CS-SRM (coherence, cause - poor prior medical care, control - by patient) were retained in the regression model; again the goodness of fit models performed poorly (Table 4).

In response to the five clinical scenarios, the respondents indicated that they would refer for lumbar spine x-ray in a mean (SD) of 1.5 (1.2) cases. The median number of referrals was 1 with a range of 0 to 3. From Table 5, the individual constructs that predicted behavioural simulation (i.e., what primary care physicians said they would do in response to the specific clinical scenarios) were: TPB (attitudes, social norms, perceived behavioural control, and intention), SCT (risk perception, outcome expectancies, and self efficacy); II; OLT (anticipated consequences, evidence of habitual behaviour); CS-SRM (control - by treatment, control - by patient, control - by doctor, cause - ageing, emotional response treatment). Neither knowledge nor PAP predicted behavioural simulation.

The results of the theory-level analyses are shown in Table 5. The TPB explained 11.6% of the variance in behavioural simulation, SCT explained 12.1%, II explained 1.5%, and OLT explained 8.1%. In the cross-theoretical construct analysis, constructs from TPB (perceived behavioural control), II and CS-SRM (cause - ageing) were retained in the regression model, together explaining 16.5% of the variance in the scenario score (Table 4).

With the range of possible scores for intention of 1 to 7, the mean (SD) intention score was 2.1 (1.0); the median intention score was 1.6 with a range of 1 to 5.5. The constructs that predicted behavioural intention were: TPB (attitudes, subjective norms, perceived behavioural control); SCT (risk perception, outcome expectancy, self efficacy); OLT (anticipated consequences, evidence of habitual behaviour); CS-SRM (control - treatment, control - patient, control - doctor, cause - stress, emotional response, and coherence); knowledge; and PAP (Table 5).

The results of the theory level analyses are shown in Table 5. The TPB explained 25% of the variance in behavioural intention, SCT 21.5%, OLT 26.3%, CS-SRM 11.3%, knowledge 2.3%, and PAP explained 2.6%. In the cross-theoretical construct analysis, constructs from TPB (perceived behavioural control), OLT (evidence of habitual behaviour, outcome expectancy), CS-SRM (control - treatment) were retained in the regression model, together explaining 33.5% of the variance in intention (Table 4).

Low back pain is a frequent presenting problem in primary care settings. However, the use of x-rays in clinical management of low back pain is relatively infrequent. In the theory level analysis predicting clinical behaviour, constructs relating to beliefs about consequences (SCT (risk perception) and CS-SRM (cause - poor prior medical treatment, cause - patient's own behaviour and cause-ageing, control - patient) and beliefs about capabilities (TPB (perceived behavioural control)) all significantly predicted behaviour. Looking across our two other outcome measures, there are also suggestions that beliefs about consequences (attitudes, outcome expectancies, risk perception, anticipated consequences) and beliefs about capabilities (PBC, self efficacy) may be important. In addition, II predicted behavioural simulation and OLT (evidence of habitual behaviour) predicted intention. The theories individually explained a significant proportion of the variance in behavioural simulation and intention, but overall were poorly predictive of behaviour. Together, these findings suggest both beliefs about consequences and beliefs about capabilities are likely determinants of lumbar spine x-ray requests.

This is a correlational study, so the causative aspects of the theories and their constructs 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 predictive elements should have the greatest likelihood of success in influencing the implementation of this evidence-based practice.

The PRIME study has evaluated the predictive value of a range of theories across different behaviours (prescribing antibiotics for upper respiratory tract infections, or URTIs, taking dental radiographs, placing preventive fissure sealants), target professional groups (primary care doctors, dentists), and contexts [17, 19, 20, 37]; we have demonstrated that different constructs predicted different proportions of the variance in the intention and behaviour. This raises the question of how best to identify relevant theories specific to different behaviours and clinical groups. One option would be to undertake preliminary work to identify the key construct domains that are likely to influence the target behaviours, and use them to specify potentially relevant theories [38, 39].

Operationalising our behaviour of interest in the surveys that reflected the available behavioural data was challenging. Our behaviour of interest was managing patients with low back pain without referral for lumbar spine x-ray. However, we could only get behavioural data on the number of lumbar spine x-ray referrals ordered by primary care physicians. In general, we tried to word the survey questions to correspond to the available behavioural data (e.g., 'when a patient presents with back pain, I have in mind to refer them for X-ray'). However, we found it difficult to frame some questions that corresponded to the behavioural data and clinically sensible. As a result the final questionnaire, included some questions worded in terms of doing the behaviour (e.g., in general, referring patients with back pain for an X-ray would...) and some worded in terms of not doing the behaviour (e.g., without an x-ray, how confident are you in your ability to...). This raises the issue of whether doing and not doing a behaviour are two sides of the same behaviour, or whether they represent linked but alternate behaviours. If the latter, the predictive ability of our survey instrument would be likely to be reduced.

Operationalising the constructs with theoretical fidelity was also challenging. A number of the models (OLT, II, CS-SRM) had not been operationalised in this way prior to the PRIME studies. OLT and II are usually used as intervention methods to change behaviour. However, both predicted behavioural simulation, and OLT predicted intention and behaviour. Since we undertook this study, some of the models have been adapted or enhanced, and different approaches to measurement have been developed -- for example, the post intentional action-coping planning enhancements of the TPB [40, 41] and Verplanken's Self Reported Habit Index [42].

The CS-SRM pattern of results mirrored the overall picture of beliefs about consequences and capabilities being important. However, they did not predict behaviour, behavioural simulation, and intention particularly well. 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 the utility of this theory to predict clinician behaviour.

There is a stepwise decrease in the proportion of variance across our dependent variables from intention to behavioural simulation (to behaviour) (Tables 2 and 4) as found in previous PRIME studies. Godin's review [16] of the predictive value of social cognitive models on professional behaviour showed a similar pattern, with social cognitive models explaining means of 13% of the variance in objectively measured behaviour (from 11 studies), 44% of self-reported behaviour (from four studies), and 59% of intention (from 72 studies). Our results are each lower than Godin's average figures, but all are within the range reported by other studies. However, our explanation of behaviour is at the very lowest limit of the reported range. In the previous PRIME studies, we have been able to explain 16% of the variance in general dental practitioners' use of dental radiographs [37] and 6% of primary care physicians prescribing of antibiotics for patients who present with an URTI [20]. This suggests that our operationalisation of the models was likely to have been good, and raises the question of why the models did not work as well for ordering lumbar spine x-rays by primary care physicians.

We can identify three potential explanations. Firstly, there was poor correspondence between the behaviour specified in the survey and the measured behaviour as mentioned above. This highlights the importance of clear and consistent framing of the questions and concordance with the measured behaviour. In the previous PRIME papers, the behaviours specified in the surveys and the measured behaviours were: dental radiographs (survey - use of intra-oral radiographs in patient management, data - the number of intra oral radiographs taken per course of treatment (good concordance)) [18], and antibiotics (survey - prescribing an antibiotic for patients presenting with URTIs, managing patients without an antibiotic, data - number of likely URTI relevant antibiotic prescriptions per 100 patients registered (weak concordance)) [20].

Second, there was potentially excess observational error (noise) in our behaviour measure. X-ray-ordering data was chosen because it was available from routine data sources, and was therefore inexpensive to collect. Low back pain was chosen because it was more likely that a request for an investigation would be attributed to the primary care doctor who issued it. Despite this, anecdotally we believe that there may be errors in the attribution of x-rays to doctors, with radiology departments reporting that requests could be reported to the correct practice but attributed to the wrong primary care doctor. In addition, we attempted to standardise our behaviour by the number of patients registered with the primary care doctor to reflect differences in workloads of the participating primary care doctors. We only had data on the total number of patients and number of primary care doctors in each practice, and so calculated an average list size per primary care doctor within each practice. This is a relatively crude standardisation approach that does not take account of likely variations of workload within practices (not all primary care doctors in the same practice will have the same workload) and variations in presentation of the target condition (not all primary care doctors will have same rate of presentation of low back pain). In the previous PRIME studies, these issues were likely to have been more problematic in the antibiotic study rather than the fissure sealant study (where data were abstracted from a claims database). These issues reflect some of the challenges of using routine data to measure behaviour relating to the level of clinical detail available (we could not estimate the number of patients each primary care doctor saw presenting with back pain) and problems of attribution of clinical actions to specific primary care doctors. In future studies of this kind, it will be important to invest more in the measurement of the behavioural data. These issues are likely to be less problematic in population-based large administrative database facilities where there may be detailed understanding of the content of the available data and their limitations. Alternatively it could be possible to collect behavioural data directly.

Thirdly, we used a different analytical approach to analyse the behavioural data. Previous PRIME studies have used multiple regression analyses and used the adjusted R² statistic from ordinary least squares (OLS) regression to quantify the proportion of variance explained by the models. In the current study, when we conducted multiple regression analyses of behavioural simulation and intention, we observed similar magnitude R² statistics for behavioural simulation and intention models. However given the distribution of the lumbar spine x-ray data, we had to use negative binomial regression for the behavioural analysis. A direct equivalent to the adjusted R² statistic does not exist for negative binomial regression. There are several pseudo R² statistics that mimic OLS R² in the sense that they can range over the scale 0 to 1 with higher values indicating a better fit of models to data. We present the results for various goodness to fit models that suggest that, in general, the resulting models overall were poorly predictive of the behavioural data. However these pseudo R² values cannot be used to compare the performance of competing theoretical models across different data sets, making comparisons of proportion variation explained with previous PRIME study surveys qualitative only. To explore the likely comparability of these results with previous PRIME studies, we undertook an OLS regression of square root transformed behavioural data and observed an R² statistic of 0.05, which is at the lower end of the observed R² statistics from previous PRIME studies. Together, we believe these data suggest that the models may be performing similarly to those in previous PRIME studies and the analytical approach required due to the negative binomial distribution is obscuring this.

Our final response rate was not high compared to what would be expected for a postal questionnaire survey to healthcare professionals. Following the report by Cummings et al. that up to 1995, response rates of surveys of healthcare professionals remained constant at approximately 60% [43], Cook et al. demonstrated that by 2005 response rates in surveys of healthcare professionals had slightly declined to an average of 57.5% [44]. Given this, we cannot exclude the possibility of selection bias in respondents and should be cautious about generalising from our respondents to the population of UK primary care physicians. However, this may be less of an issue at this exploratory stage of using these methods, as the purpose of the study was theory testing and an exploration of the predictive ability of theories to explain variations in behaviour. 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.