Explaining clinical behaviors using multiple theoretical models

Explanatory measures were chosen to reflect both volitional (processes that are aware, deliberative, reflective, and may therefore be considered subject to volition) and non-volitional (processes that are associative, automatic, impulsive, and may therefore be considered not subject to volition; here represented by Learning Theory, or LT) theoretical constructs that have been found to predict behavior. All measures were included in each study. Measures were derived for the constructs in the Theory of Planned Behavior (TPB)[15], Social Cognitive Theory (SCT)[16, 17], and for illness representations specified by the Common Sense Self Regulation Model (CSSRM)[18]. We constructed self-report measures of two constructs from LT (habitual behavior and anticipated consequences)[19, 20], a measure of implementation intentions (II)[21], and questions to assess stage in the process of change as specified by the Precaution Adoption Process (PAP)[22, 23].

The TPB[15] by Aizen and Fishbein proposes that the strength of an individual’s behavioral intention to engage in a behavior, and the degree of control they feel they have over that behavior (perceived behavioral control) are the proximal determinants of engaging in it. The perceived behavioral control construct is closely related to the concept of self-efficacy in SCT. The theory of planned behavior also proposes that behavioral intention strength is determined by three variables: attitudes towards the behavior (the extent to which the behavior will result in consequences that the person values); subjective norms for the behavior (the belief that other people whose opinion influences the person think the person should engage in the behavior); and perceived behavioral control over it. These variables in turn are based upon salient behavioral, normative, and control beliefs about the behavior. Each predictor variable may be measured directly (by asking respondents about their overall attitude), or indirectly (by asking respondents about specific behavioral beliefs and outcome evaluations). Aizen suggests that behavior change is brought about by interventions that change these beliefs.

Bandura’s SCT[16] proposes that behavior is determined by three kinds of expectancies: situation-outcome expectancies, outcome expectancies, and self-efficacy expectancies. Situation-outcome expectancies are beliefs about how events are connected (e.g., a diagnosis of cancer is likely to result in death). Outcome expectancies refer to beliefs about the consequences of performing a behavior (e.g., if I stop smoking, I will put on weight). Self-efficacy expectancies are beliefs about one’s ability to perform the behavior (e.g., I can stop smoking). Self-efficacy has been found to be the strongest determinant of behavior, and the theory proposes four methods of changing self-efficacy in order to change behavior: providing mastery experiences (i.e., successful behavioral performance); modelling (i.e., observation of successful behavioral performance); verbal persuasion; and addressing attributions for physiological and affective states

Leventhal’s CSSRM[18] proposes that individuals attempt to make sense of illness by making use of pre-existing knowledge or schemas to develop both cognitive representations of the illness (including its identity, consequences, timeline, cause, and cure/control) and emotional representations of the threat involved. These representations give rise to behavioral responses. For patients, behavioral responses may include going to see a doctor, taking prescribed or non-prescribed medicines, or participating in rehabilitation programmes. For clinicians, responses may include referring a patient for diagnostic tests, prescribing a drug, or restoring a carious tooth. The model is described as self-regulatory because this is seen as a dynamic process. The aim of the actions that people take in response to a health threat is to reduce the degree of threat or to restore their own (or their patient’s) emotional equilibrium. The individual’s own understanding of the health condition and their personal emotional reaction are central to the self-regulatory approach. It seems likely that health professionals also experience emotional reactions to some conditions (e.g., terminal conditions in children) or patient groups, for example ‘heart sink’ reactions, which may influence their practice. In this model, behavior change results from changes in the cognitive representations or in the responses available to manage either the threat or the emotions.

Learning theory proposes that behaviors that have positive consequences for the individual (such as remuneration) are likely to be repeated, whereas those that have unpleasant consequences will become less frequent[20]. Learning happens when consequences occur if and only if the behavior has occurred and can take a variety of forms, from material rewards (e.g., financial incentives), through social rewards (e.g., receiving praise) to personally salient rewards (e.g., achieving a desired goal). The principle that positive consequences promote repetition of behavior is well established and has been widely and successfully used to understand behavior and behavior change[24]. As behaviors are repeatedly rewarded, they may become ‘habitual’ and, if reward is scheduled appropriately, the behavior may be maintained with a reduced frequency of reward; if a behavior is increased by being rewarded and then removed without planned scheduling, the learned behavior will be ‘extinguished’ unless some other form of reward is available,

‘Implementation Intentions’ are explicit ‘if-then’ action plans about when and where a goal intention will be achieved[21]. Gollwitzer[21] has made the distinction between goal intentions and II. A goal intention is an intention to perform a behavior or achieve a goal (e.g., I intend to reduce the number of referrals I make for lumbar spine x-rays). This is conceptually close to the behavioral intention construct in the TPB. Gollwitzer argues that by creating an II, people effectively transfer control of the behavior to the environment—establishing cues to action. For example, by saying that ‘When a patient tells me about their low back pain, I will explain the pros and cons of an x-ray to them.’

Stage theories propose that behavior change occurs in a stepwise process, rather than a linear fashion as implied by motivational or action theories. From a stage theory perspective, interventions to facilitate change will be most effective if they are tailored to the stage an individual has reached within this process. While there are differences between the stage models in the number and nature of stages proposed, stage theories typically distinguish motivation and action steps[17, 22, 25]. In the PAP model, Weinstein has proposed an additional early stage when individuals may be unaware of the need for behavior change, a stage which may have particular relevance to the early stages of the implementation of evidence[22].

A common implicit model is the Knowledge-Attitudes-Behavior model. This assumes that a change in knowledge will produce a change in attitude, and this will, in turn, produce a change in behavior. While this model has some value, it should be tested rather than assumed before being used as a basis for behavior change. Accordingly, we examined the ability of knowledge to predict behavior.

The questions were developed based on the standard methods for each theory where possible. Additionally, for each behavior five knowledge questions, for which there was good evidence, were developed by the study team. Table1 provides a summary of the explanatory measures used in each of the studies (with examples from the Managing Low Back Pain Without Ordering Lumbar Spine X-rays study). Questions were rated on a seven-point Likert scale from ‘strongly disagree’ to ‘strongly agree.’ Scores were adjusted so that a higher score equated with a greater propensity towards evidence-based behavior. Exemplar questionnaires are available through previous publications[11, 13, 14] and in Additional file2.

The three dental and two medical behaviors are shown in Table2. In each case, the description of the clinical behavior contained in the questionnaires was specified according to the TACT principle—in terms of its Target patient, the Action to be performed, the Context or conditions for the action and when the action would be performed (Time). For example, for managing low back pain, the target is patients presenting with back pain, the action is managing the patient without referring for lumbo-sacral spine (lumbar) X-rays, the context is general practice, and the time is (implicitly) now.

For each of the five behaviors, we collected two measures of behavior, an objective measure (Table2) collected from routine data systems and gathered for a 12-month period to control for seasonal variations, and a measure of simulated behavior. For simulated behavior, key elements that may influence clinicians’ decisions to perform the behaviors were identified from the literature, opinion of the clinical members of the research team, and the interviews with clinicians. From this, we constructed five clinical scenarios that described patients presenting in relation to the index behaviors. Respondents were asked to decide what they would do, and responses were summed to create a total score out of a possible maximum of five (Table1).

We also elicited behavioral intention using three questions assessing intention to perform the behavior of interest. With a behavior specific stem the questions were worded ‘I have in mind to,’ ‘I intend to,’ ‘I aim to’ (rated on a seven-point scale from ‘strongly disagree’ to ‘strongly agree’).

Clinicians were sent an invitation pack comprising a letter of invitation, a questionnaire consisting of theory-based cognition measures (explanatory measures), two interim outcome measures (behavioral intention and behavioral simulation), and demographic measures, plus a consent form to allow access to their data (objective behavior measure) from the routine systems, and a reply-paid envelope. Three postal reminders were sent to non-responders at two, four, and six weeks from the first mailing.

Each study included a large number of predictor variables, and a power calculation suggested that, for each of the five studies, a minimum sample of 200 clinicians was required to detect an effect size of 0.40, with alpha of 0.05, 95% power.

The measures of the theoretical constructs were tested for acceptable internal consistency. If the criterion value (0.60) was not reached, items were dropped from the variable measures until the maximum possible Cronbach’s alpha was achieved or, for two items, a correlation greater than 0.25.

We examined the relationship between predictive and dependent variables using correlation, except for the PAP where ANOVA was used. For the two general medical behaviors for the PAP analysis, respondents were dichotomized into two groups (decided to perform desired behavior or have already performed desired behavior versus all other responses), and the relationship between predictive and outcome variables were examined using regression models.

Multiple regression analyses were used to examine the predictive value of each of the theoretical models. For the behavior ‘managing low back pain without referring for a lumbar spine x-ray,’ given the distribution of the behavioral data, we used negative binomial regression (NBR) to model the predictive ability of individual theoretical constructs and complete theories.

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

Questionnaire response rates were: 48% (214/450) for taking dental radiographs; 29% (130/450) for performing dental restorations; 29% (120/407) for placing fissure sealants; 21% (230/1100) for managing upper respiratory tract infections (URTIs) without prescribing antibiotics; and 27% (299/1100) for managing low back pain without ordering lumbar spine x-rays. We achieved the target sample size for three of the five studies.

The rates of performing the objective behaviors and the scores for simulated behavior and intention are shown in Table2.

Summaries of the mean scores for the explanatory variables are shown in Table3 and Additional file3. In general, the mean scores were above the midpoint on the scale, with median values across the five behaviors generally being above four out of seven and the range being from 1.53 to 6.01.

The proportion of variance in the three dependent variables of intention, behavioral simulation, and behavior explained by the theories for each of the five behaviors are shown in Table4 and Additional file4.

Across all of the theories, the highest proportion of the variance explained was always for intention and the lowest was for objective behavior. The Knowledge-Attitudes-Behavior Model performed poorly across all five behaviors and three dependent variables. CSSRM performed poorly for predicting behavior, though it predicted intention for two of the behaviors. For TPB, SCT, II, and LT across the five behaviors, the theories predicted median R² of 25% to 42.6% for intention, 6.2% to 16% for behavioral simulation, and 2.4% to 6.3% for (three of the five) objective measures of behavior.

Comparing the performance of the models, TPB was predictive of three behaviors (taking dental radiographs, managing URTIs without prescribing antibiotics, and managing low back pain), SCT and LT were predictive of (the same) two behaviors (taking dental radiographs and managing URTIs without prescribing antibiotics), action planning predicted one behavior (taking dental radiographs) while other models were not predictive of behavior.

The TPB, SCT, and LT performed well in predicting intention, confirming their effectiveness in predicting key elements of the models (behavioral intention in TPB; goals in SCT). However, LT predicted more variance in intention than either TPB or SCT for four of the five behaviors, despite having fewer predictive variables and not usually being used as a model of cognitions. The PAP (Additional file 4) did not predict any dependent variable for ‘taking dental radiographs,’ predicted behavioral simulation and intention for ‘performing dental restorations,’ predicted intention for ‘placing fissure sealants,’ predicted all three dependent variables for ‘managing URTIs without prescribing antibiotics’ and predicted intention for ‘managing low back pain without ordering lumbar spine x-rays.’

A major strength of conducting five studies and having multiple theories/models operationalized in the same way across a wide range of behaviors that spanned disciplines was that this allowed the identification of variation in the performance of the models, both between models and across behaviors. Such variation would not be apparent from single condition, single theory studies. It allows an additional degree of interpretation that is not otherwise possible.

The fact that we used several theories/models and tried to adhere to the best practices for measuring constructs related to each theory (e.g., having three items measuring each construct to achieve a reliable measure of the construct) meant that the questionnaire was both long and appeared repetitive (e.g., When I see patients with URTIs, I automatically consider managing them without an antibiotic; it is my usual practice not to prescribe antibiotics for patients with URTIs; I am not to prescribe antibiotics for patients with URTI). In addition, the wording specified by each of the theories also meant that, for many questions the format of the questions may have appeared constrained and non-intuitive (e.g., the symptoms of an URTI are puzzling to me). This almost certainly affected our response rates.

We did not achieve our target number of responses for two of the surveys. This means that we may be underpowered for a number of our analyses and may underestimate the performance of the models. Higher response rates would give greater confidence in the representativeness of the data, though this aspiration is set against the evidence of modest and slowly declining response rates to postal surveys among healthcare professionals—61% over the 10 years up to 1995 and 57% in the 10 years after[26]. Increased response rates might be achieved by fully compensating the recipients for their time. While financial incentives have shown little impact in increasing survey response rates, the amounts involved have usually been small and not equivalent to full recompense[27]; it would be reasonable to explore the effectiveness of full recompense and to use it as a strategy in the future[28].

Although we used standard methods and wordings whenever possible, this was not possible for II and LT where constructs have more often been manipulated in experiments than measured in predictive studies such as those we have conducted. In addition, a number of the theories contained constructs that were similar and difficult to differentiate (e.g., TPB attitudes (indirect) and SCT outcome expectancies), and we ended up using the same questionnaire items as a measure of both.

For any study of the behavior of healthcare professionals, there is the key decision about how to measure their behavior. We used measures of behavior that were available from routine sources. We were trading off any lack of specificity of these measures of behavior compared to the behaviors we aimed to predict against the costs of collecting the most specific data. This would have involved data collection within the practice of each recipient and would have been very costly.

We were aware that the concordance of the TACT specification of the behaviors in the questionnaires and the behavioral measures available within the various routine data sets varied. Of the five sets of behavior data, the three general dental measures were all claims-based measures from a comprehensive central claims database. We were confident that the dental radiographs data provided a good measure of the behavior we specified in the questionnaires, but there were problems with the other two dental behavior measures. The fee claim database was too imprecise to be used for the placing fissure sealants behavior (due to our interest in a specific age defined subgroup) and was not analysed, although a randomized controlled trial running during the time period of this study successfully collected practice based data on this behavior[29] and, using the same theories, successfully predicted behavior[30]. We were specifically interested in restorations in permanent teeth in children aged <17 years, but the dental restorations behavior data did not sufficiently specify this age group and tooth restored and so was less specific than we anticipated and would have wished.

The two general medical behaviors we specified within the questionnaires were in the context of not performing a behavior—doctors managing URTI without prescribing antibiotics, and managing patients with low back pain without taking x-rays. Our objective measures of these behaviors were rates of prescribing antibiotics and performing x-rays. This meant that we could never know the number of patients who presented with URTI or back pain, or the proportion of those presenting who were not treated nor referred. We had to assume that this proportion varied across practices in the same way as the data we collected. It is also the case for these two behaviors that ‘not performing the target behavior’ is unlikely to reflect one or more common alternative behaviors across all clinician’s practice. Clinicians are likely to have a number of alternative behaviors that they perform in order to avoid prescribing or ordering x-rays rather than doing nothing. We have previously shown that generating alternative clinical behaviors can increase intention to ‘not perform’ behaviors that are counter to evidence-based practice[26]. Identifying and specifying these may have improved the precision of our prediction in these clinical areas—assuming we could also have identified a measure of them. Given the low proportions of variance in behavior explained (with the possible exception of dental radiographs), it would seem prudent to exercise caution about the use of routinely collected data unless it is a good match to the specified clinical behavior in the survey.

The theories all predicted intention better than they predicted simulated behavior and, in turn, simulated behavior better than objective measures of behavior as is usually found[31] and is due at least in part to common method (self-report) variance. It seems logical that, as the length of the causal chain (attitudes to intention to self-report behavior to observable behavior) increases, there is more opportunity for other factors to come into play and prediction to be more imprecise.

The effectiveness of LT (and its consideration of habit) in predicting behavior raises the importance of considering non-volitional, non-deliberative aspects of performing clinical behaviors and the idea of dual processing models (i.e., models which propose both deliberative and associative or automatic processes) as explanatory models of professional behavior and as ways of considering analysing such data[26, 27, 32].

Prediction of behavior by knowledge was included because it is one of the most common implicit models as a basis for changing behavior used in medicine[29]; based on its poor performance, it would be easy to justify omitting it from further study and sensible to suggest that this is not a viable model for promoting the uptake of evidence-based practice. However, given that we only included five questions and may not have sampled the critical knowledge, there may be different measures of knowledge that are predictive. The CSSRM performed less well than other models. The CSSRM is model that was developed specifically in the context of an individual’s illness and ‘health’ behaviors and thus has a different standpoint to the other theories. We operationalized it to assess a clinician's view about a condition they themselves were not experiencing, but one that they had to treat. The poor results of this model over our studies suggest that clinicians’ perceptions about a condition, in and of itself, are unlikely to be taken into account when choosing the behavior to treat it.

Several of the theories that we used performed well and should be further used in implementation research (TPB, SCT, II, LT). Given the length of any questionnaire using multiple models (and our experience of low response rates), it is important to consider ways of shortening the instruments. A number of simple steps seem worth considering. Researchers should consider removing models that do not consistently perform well (CSSRM and KAB) from these studies. In future, it might be possible to select theories that explain three key elements of behavior change: how motivation to change behavior is developed; how motivated individuals turn this motivation into action; and how behavior may be changed without the development of a deliberative intention to change.

It would also be important to include theories or models that suggest not only how to explain behavior change that has occurred, but also suggest how to change behavior; both LT and SCT give clear guidance on how to change behavior, while this is less clear for TPB. Further, II have been shown to be methods of changing behavior[32].

A further option as a prior step to developing an intervention could be to develop a screening instrument that uses single (or small numbers of) item measures of key constructs. For each behavior, an investigator could ask a single intention question, a single action question, and a single question tapping non-volitional processes in order to gain a profile of the influences on the behavior. The advantage of such a method is that the initial approach involves a very short instrument. If intention scores were high, there would be little additional benefit in measuring any pre-intentional measures because strengthening any of them could not produce a large increase in intention; under such circumstances, exploring further the post-intentional or contextual/organisational factors would be a better use of resources. A similar, though slightly longer approach would be to use the Theoretical Domains Framework[26, 29, 30, 32, 33] as an instrument to help identify potentially relevant domains and to target theoretical surveys addressing key domains.

Our studies represented a non-experimental empirical approach to understanding behavior and the processes involved in enacting them. They allowed the exploration of a range of theories and the potential for subsequent exclusion of theories that did not predict behaviors. They also allowed an understanding of both volitional and non-volitional processes. However, they do not have the ability to quantify behavior change or explain causal mechanisms, and to do that would require the use of experimental designs, either using interim outcomes[27, 28, 34] or actual clinical behaviors[35]. We have taken the ideas from these studies and one of these behaviors on to identifying and modelling the performance of an intervention[28, 36]. Further, we have tested a clinical intervention based on LT, contrasting it with an intervention based on a knowledge model, in increasing dentists’ use of fissure sealants for children; the LT intervention, providing contingent consequences for each behavior, was effective in changing behavior, while the knowledge based intervention, providing additional education, did not influence behavior[35].

Within these five studies, and many other theory-based studies reported in the literature, the behaviors are those enacted by individuals in isolation from other healthcare professionals. However, much of the work of primary care medical practice involves the behaviors of teams of individuals (as in the management of chronic diseases) where patients are managed by the (more or less) integrated actions of teams of healthcare professionals, and this is where much future interest in the utility of theory-based methods will lie. With the unit of analysis at the level of the team (which in primary care is often the same as the practice), there is legitimate interest in the cognitions of all members of a team and the need to explore the impact of these on practice level behavior[37, 38].

We operationalized multiple theories measuring across five behaviors. Some of the theories that we used successfully predicted behavior offering a basis for choosing between them. Continuing challenges that emerge from our work are: the need for better specification of behaviors; to better operationalize theories; to consider how best to appropriately extend the range of theories; to further assess value of theories in different settings and groups; exploring the implications of these methods for the management of chronic diseases; and moving to experimental designs to allow an understanding of behavior change.