You are here

Article Text

Article menu

Download PDFPDF

Original article
Self-efficacy and risk of persistent shoulder pain: results of a Classification and Regression Tree (CART) analysis
  1. Rachel Chester1,
  2. Mizanur Khondoker2,
  3. Lee Shepstone2,
  4. Jeremy S Lewis3,
  5. Christina Jerosch-Herold1
  1. 1 School of Health Sciences, Faculty of Medicine and Health Sciences, University of East Anglia, Norwich, UK
  2. 2 Norwich Medical School, Faculty of Medicine and Health Sciences, University of East Anglia, Norwich, UK
  3. 3 Department of Allied Health Professions, School of Health and Social Work, University of Hertfordshire, Hatfield, UK
  1. Correspondence to Dr Rachel Chester, School of Health Sciences, Faculty of Medicine and Health Sciences, University of East Anglia, Norwich NR4 7TJ, UK; r.chester{at}uea.ac.uk

Abstract

Objectives To (i) identify predictors of outcome for the physiotherapy management of shoulder pain and (ii) enable clinicians to subgroup people into risk groups for persistent shoulder pain and disability.

Methods 1030 people aged ≥18 years, referred to physiotherapy for the management of musculoskeletal shoulder pain were recruited. 810 provided data at 6 months for 4 outcomes: Shoulder Pain and Disability Index (SPADI) (total score, pain subscale, disability subscale) and Quick Disability of the Arm, Shoulder and Hand (QuickDASH). 34 potential prognostic factors were used in this analysis.

Results Four classification trees (prognostic pathways or decision trees) were created, one for each outcome. The most important predictor was baseline pain and/or disability: higher or lower baseline levels were associated with higher or lower levels at follow-up for all outcomes. One additional baseline factor split participants into four subgroups. For the SPADI trees, high pain self-efficacy reduced the likelihood of continued pain and disability. Notably, participants with low baseline pain but concomitant low pain self-efficacy had similar outcomes to patients with high baseline pain and high pain self-efficacy. Cut-off points for defining high and low pain self-efficacy differed according to baseline pain and disability. In the QuickDASH tree, the association between moderate baseline pain and disability with outcome was influenced by patient expectation: participants who expected to recover because of physiotherapy did better than those who expected no benefit.

Conclusions Patient expectation and pain self-efficacy are associated with clinical outcome. These clinical elements should be included at the first assessment and a low pain self-efficacy response considered as a target for treatment intervention.

  • shoulder
  • physiotherapy
  • longitudinal study
  • observational study
  • prospective

https://doi.org/10.1136/bjsports-2018-099450

Statistics from Altmetric.com

    Request Permissions

    If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

    What are the findings?

    • High levels of pain and disability at baseline are associated with high levels of pain and disability at 6-month follow-up. However, this ‘predicted’ poor outcome is modified to a predicted better outcome if the patient has high pain self-efficacy and a greater expectation of treatment.

    • Pain self-efficacy is the extent or strength of the patient’s belief in their ability to complete tasks and reach a desired outcome despite their shoulder pain.

    • Low levels of baseline pain and disability are associated with low levels of follow-up pain and disability. This predicted better outcome is modified to a predicted poor outcome if the patient has low pain self-efficacy.

    How might it impact on clinical practice in the future?

    • We recommend that pain self-efficacy and patient expectation of outcome as a result of physiotherapy treatment should be formally assessed and discussed at the first physiotherapy appointment.

    Introduction

    Persistent musculoskeletal shoulder pain is common and frequently associated with substantial disability. Over a period of 1 month, between 16% and 31% of the general population in the UK will have suffered from musculoskeletal shoulder pain lasting at least 24 hours.1–3 Experiencing shoulder pain concerns many and accounts for up to 3% of visits to general practitioners (GP) annually4 5 and up to 48% of people with shoulder pain visit their GP more than once over a 3-year period due to ongoing symptoms.5 6 The most effective treatment is not yet known; clinical trials comparing surgical and non-surgical management, including exercises prescribed by physiotherapists report equivocal effects.7–9 Between 8% and 11% of patients visiting their GP with shoulder pain are referred to see a physiotherapist at initial consultation,5 10 11 rising to 18% over a 3-year period.5

    Response to physiotherapy is variable. In a multicentre cohort study of 1030 patients with musculoskeletal shoulder pain attending physiotherapy, of mean duration 14 months (SD 28 months), 69% of patients reported complete recovery or being much improved by 6 months follow-up; 17% reported only slight improvement and 14% reported no change or a worsening of symptoms.12 A multivariable general linear model (GLM) was used to identify prognostic factors associated with patient-rated pain and disability.13 Several factors were consistently associated with a better outcome at 6 months. One limitation of the GLM approach is the difficulty of practical use, particularly in a clinical setting. All predictor variables within the model are used simultaneously requiring lengthy calculations (particularly when there are many predictor variables).

    Classification and Regression Tree (CART) analysis is an alternative method of providing prognostic guidance. CART analysis considers the predictive value of prognostic factors sequentially, that is, in a hierarchy of importance. CART typically results in a simple and readily interpretable decision ‘tree’ (or ‘what if’ flow diagram). This can be graphically represented easily and requires no numeric calculations.14 This can help guide clinicians to prioritise their initial prognostic assessment to those factors which are most influential. When modifiable, prognostic factors may become targets for interventions and inform shared decision making between clinicians and patients. The objective of these analyses was to provide clinicians with a guide to the most influential factors that predict outcome for people undergoing management for non-surgical musculoskeletal shoulder pain.

    Methods

    Data were available on 1030 people with shoulder pain recruited from primary and secondary care to a multicentre longitudinal cohort study in the East of England, UK, between November 2011 and October 2013. People aged 18 years or over were eligible to participate if they were referred to physiotherapy for the management of musculoskeletal shoulder pain and complained of shoulder or arm pain reproduced on movement of the shoulder. Those presenting with shoulder fractures, traumatic shoulder dislocations, systemic source of shoulder symptoms, cervical radiculopathy or had undergone shoulder surgery were excluded. Referral and treatment pathways were unaffected by participation in the study. The study protocol has been published.15

    Outcome variables

    Two validated patient-rated outcome measures were collected at baseline and via postal questionnaire at 6-month follow-up: the Shoulder Pain and Disability Index (SPADI),16 17 and Quick Disability of the Arm, Shoulder and Hand (QuickDASH).18 The SPADI is a joint specific questionnaire designed to measure two domains: shoulder pain and disability. Thirteen items, five comprising a pain subscale and eight a disability subscale, are scored from zero to 10, where zero represents no pain or disability and 10 represents the worst pain imaginable or so difficult it requires help. For this analysis, each domain carried equal weighting in the overall score. The QuickDASH is an upper limb region specific questionnaire that includes items related to symptoms, daily activities, sleep, social and work function. Eleven items are scored from one to five where one represents no difficulty and five represents unable. Each item carries equal weighting in the final score. Scores are converted to a scale of zero to 100, where zero represents no pain or disability and 100 represents maximum pain and disability.

    Baseline predictor variables

    Data for potential prognostic factors were collected prior to and during the participant’s first physiotherapy appointment using bespoke questionnaires and clinical record forms. These variables included demographics, patient expectations and beliefs, lifestyle, general health, work, shoulder history and presentation and clinical examination findings.15 All factors statistically associated with outcome (p≤0.05) in at least one of the multivariable linear models from our previous analysis13 were included in the CART analysis. In addition, baseline factors measured, but not found to be statistically significant, were entered if reported as a significant prognostic factor for outcome in reviews of other musculoskeletal studies.19 20 A description of the variables included in the CART analysis are detailed in online supplementary file 1.

    Supplemental material

    Patient involvement

    Patient and public representatives were involved in the design of the study, in particular, details associated with the timing and procedures for recruiting and follow-up of participants, and the design and layout of questionnaires for data collection. A lay version of results, designed with patient and public representatives, were disseminated to all study participants who at their final data collection replied that they would like a copy. Patients were not involved in the actual recruitment or conduct of the study.

    Statistical analysis

    We used regression trees algorithms, a subclass of CART,21 for continuous outcome variables, in our analyses. CART uses a recursive partitioning of the study sample to produce subgroups as homogenous as possible with respect to the outcome of interest. This partitioning is based on a binary split of each predictor variable. It is a more flexible approach for uncovering complex variable relationships than traditional linear modelling as it does not rely on any functional relationship between the outcome and predictor variables, nor does it require any distributional assumption regarding the outcome variable. CART is also less sensitive to outlying data, well suited for a large number of predictor variables and therefore offers a suitable alternative for building prediction models where the relationships among variables are unspecified and existing parametric statistical methods are not suitable to guide the model building.22 The prediction accuracy of CART is comparable with parametric regression models and it can be more accurate when the relationship between the outcome and predictor variables is non-linear.21 Furthermore, the partitioning in CART can be represented graphically as an easily interpretable decision tree14 that may then be used to inform clinical practice.

    We constructed four regression trees for each of the four outcome variables, that is, SPADI overall score, SPADI pain subscore, SPADI disability subscore and QuckDASH score, respectively. We used the R (R Core Team, 2015) package rpart.23 For each of the 6-month outcome variables the respective baseline score was included within the list of predictor variables. Including the respective baseline scores, the number of predictor variables entered in different models ranged between 32 and 34. The pain and disability subscales were not included in the total SPADI tree analysis therefore using only 32 variables (see online supplementary file 1 for a complete list and definitions of variables).

    The procedure for building a regression tree in rpart is discussed below.24

    Building a tree

    First, the predictor variable is found which best splits the sample into two subgroups. The ‘best’ is defined as the split that maximises the between groups sum-of-squares (or, equivalently, minimises the within-group error sum-of-squares). This process is applied separately to each subgroup recursively until the subgroups either reach a minimum size (set to seven in our analysis, the default in rpart) or until no improvement can be made in the model fit.

    Pruning the tree

    The resultant model is typically too complex and likely to overfit the data. The second stage of the procedure consists of using cross-validation to trim back the full tree. We used 10-fold cross-validation to evaluate model fit at a series of model complexities and chose the optimal (pruned) tree by inspecting the plot of the cross-validated error against model complexity. Statistical analyses were completed by a statistician without prior knowledge of the clinical area or results of earlier analyses. Based on the cross-validated predicted residual error sum of squares statistic, these trees provided the best predicting models. Retention of more baseline variables within the trees did not improve their predictive capacity. See online supplementary file 2 for the plot for cross-validated errors versus model complexity for SPADI (total score) at 6 months.

    Supplemental material

    Estimating a model and evaluating prediction accuracy on the same dataset generally overestimates model performance. It is, therefore, recommended to evaluate the model performance on an independent dataset (ie, a dataset that was not used to estimate the model). In the absence of an independent test dataset for model evaluation, cross-validation approach is an alternative way to create independent datasets for model assessment by holding apart a small portion of the sample for model evaluation. More specifically, 10-fold cross-validation involves randomly splitting the data into 10 parts of similar size, holding aside one part (1/10th of the whole sample) for testing and using the rest (9/10th) for model estimation. The process is repeated 10 times, meaning that each of the 10 folds is used as independent test set for model evaluation. Overall model performance is typically assessed by calculating prediction errors on each of the 10 folds and averaging across all of these results.

    Cross-validation is a widely used and acceptable way of validating model accuracy/performance and we used this approach to select the optimal CART model, that is, to select the variables that are most predictive of the respective outcome variables (and also to remove those not contributing enough to the prediction model). The results of this validation process ensures that our selected CART models are optimal, despite not having a separate independent validation dataset.

    Results

    One thousand and fifty-five participants were assessed by physiotherapists and subsequently recruited and consented into the study. One thousand and thirty participants were found to be eligible for the study and provided adequate baseline data. There were no potential prognostic factors at baseline for which >2% of data were missing. Eight hundred and eleven participants (79%) provided outcome data at 6-month follow-up. One participant was excluded due to incomplete outcome data (see flow diagram in figure 1).

    Figure 1
    Figure 1

    Strengthening the Reporting of Observational Studies in Epidemiology flow diagram. Participant recruitment and follow-up.

    All participants providing complete outcome data at 6 months were included in the CART analysis for the SPADI and QuickDASH (n=810). There were differences between those participants who provided complete data at 6 months and those who did not. Completers were older by a mean of 10 years, had greater pain self-efficacy by a mean of almost 4 out of a possible 60 points, were almost twice as likely to exercise and had a twofold lower likelihood to report anxiety or depression. A summary description of baseline characteristics for all participants (n=1030) for each of the 34 variables entered into the CART analysis are provided in online supplementary file 3.

    Supplemental material

    Figures 2–5 represent the resulting pruned regression trees for the total SPADI, SPADI pain subscale, SPADI disability subscale and QuickDASH at 6-month follow-up. Three variables were identified as important predictors of 6-month outcomes: (1) baseline pain or disability levels, (2) pain self-efficacy and (3) patient expectation of ‘change as a result of physiotherapy treatment’. All three variables were collected prior to the participant’s first physiotherapy attendance. Pain self-efficacy is the extent or strength of the patient’s belief in their ability to complete tasks and reach a desired outcome despite their shoulder pain.25 Pain self-efficacy was measured using the pain self-efficacy questionnaire (PSEQ),26 which comprises 10 items rated 0 to 6, zero representing minimum pain self-efficacy and six representing maximum pain-self efficacy. The total score is out of 60, a higher score representing higher pain self-efficacy. Patient expectation of change was collected in response to the following question ‘How much do you expect your shoulder problem to change as a result of physiotherapy treatment’ and was measured on a 7-point Likert scale ranging from ‘completely recover’ to ‘worse than ever’.15

    Figure 2
    Figure 2

    Regression tree for total Shoulder Pain and Disability Index (SPADI) score. Cut-off points for the SPADI and pain self-efficacy questionnaire (PSEQ) have been rounded up or down to whole numbers. The four box plots at the bottom of the figure illustrate the distribution of total SPADI scores at 6-month follow-up. The median SPADI score at 6-month follow-up, (represented by the horizontal line dissecting the box), is lowest (better outcome) in the subgroup represented by the box furthest left and highest (poorer outcome) in the subgroup represented by the box furthest right.

    Figure 3
    Figure 3

    Regression tree for Shoulder Pain and Disability Index (SPADI) pain subscale score. Cut-off points for the SPADI pain subscale scores and pain self-efficacy questionnaire (PSEQ) have been rounded up or down to whole numbers. The four box plots at the bottom of the figure illustrate the distribution of total SPADI pain subscale scores at 6-month follow-up. The median SPADI pain subscale score at 6-month follow-up (represented by the horizontal line dissecting the box) is lowest (better outcome) in the subgroup represented by the box furthest left and highest (poorer outcome) in the subgroup represented by the box furthest right.

    Figure 4
    Figure 4

    Regression tree for Shoulder Pain and Disability Index (SPADI) disability subscale score. Cut-off points for the SPADI disability subscale scores and pain self-efficacy questionnaire (PSEQ) have been rounded up or down to whole numbers. The four box plots at the bottom of the figure illustrate the distribution of total SPADI disability subscale scores at 6-month follow-up. The median SPADI disability subscale score at 6-month follow-up (represented by the horizontal line dissecting the box) is lowest (better outcome) in the subgroup represented by the box furthest left and highest (poorer outcome) in the subgroup represented by the box furthest right.

    Figure 5
    Figure 5

    Regression tree for Quick Disability of the Arm, Shoulder and Hand (QuickDASH). Cut-off points for the QuickDASH scores have been rounded up or down to whole numbers. The four box plots at the bottom of the figure illustrate the distribution of QuickDASH scores at 6-month follow-up. The median QuickDASH score at 6-month follow-up (represented by the horizontal line dissecting the box) is lowest (better outcome) in the subgroup represented by the box furthest left and highest (poorer outcome) in the subgroup represented by the box furthest right.

    The first ‘node’ (at the top of the trees) represents the sample (ie, all 810 participants). This then divides into two, based on cut-off values for baseline pain or disability (SPADI, SPADI subscale score or QuickDASH). The baseline score was therefore considered the most important variable in predicting the respective 6-month outcome. In addition to baseline pain or disability, each pruned regression tree retained only one other variable of the 34 variables considered: baseline pain self-efficacy or patient expectation. Either pain self-efficacy or patient expectation led to classification of participants into four subgroups. The number of participants in these subgroups ranged from 48 to 487.

    The cut-off point for baseline SPADI scores (total, pain and disability subscores) at the first node of each tree ranged from 62 to 75. When subdividing patients with lower baseline SPADI or baseline SPADI pain subscores into two groups using baseline pain self-efficacy scores, the cut-off for the PSEQ was 40 and 41, respectively. When subdividing patients with higher baseline SPADI pain or disability scores into two groups using pain self-efficacy scores, the cut-off point for the PSEQ was consistently 48.

    Table 1 and figures 2–4 show that the size of any subgroup with low pain self-efficacy ranged from 16% (n=127) to 20% (n=161) of participants in the cohort. The SPADI pain tree includes two subgroups with low pain self-efficacy and this constitutes as 36% (n=288) of the cohort. The discrimination between median outcome scores associated with different pain self-efficacy scores differs between trees and baseline pain and/or disability. For example, the median difference in subgroups is most marked for participants with high baseline SPADI pain subscores (≥75) and least for participants with lower baseline total SPADI scores (<68). Twenty-three per cent (n=239) of the cohort had a baseline of 41–59 on the QuickDASH, their outcomes were differentiated by their expectation of ‘change as a result of physiotherapy treatment’: the median difference between subgroups at outcome being 23/100 on the QuickDASH.

    View this table:
    Table 1

    Median (IQR) outcome (SPADI total, SPADI pain subscale, SPADI disability subscale and QuickDASH) for each subgroup for each tree

    Validation

    External validation of the results was not possible as we were unable to identify an external dataset containing the same or similar variables. We have, however, conducted an informal internal validation of the results by partitioning the QuickDASH outcome data based on the classifications of the SPADI regression trees and comparing the distribution of QuickDASH outcome within each subgroup with that of the SPADI outcomes. Comparison of QuickDASH distributions corresponding to the total SPADI, SPADI pain and SPADI disability trees are displayed in online supplementary files 4, 5 and 6, respectively. The similarity of the pattern distributions of the SPADI outcome on the left and QuickDASH outcome on the right demonstrate the replicability of the SPADI tree.

    Discussion

    The objective of these analyses was to identify important predictors of outcome for patients presenting with non-surgically managed musculoskeletal shoulder pain. We identified that only 3 of 34 baseline variables considered in the classification trees were predictive of outcome. These were i) baseline pain or disability measured by the SPADI or QuickDASH, ii) pain self-efficacy measured by the PSEQ26 and iii) patient’s expectation of ‘change as a result of physiotherapy treatment’, measured on a 7-point Likert scale.15 As expected, there was a positive association between pain and disability at baseline and at 6-month follow-up, that is, those with higher scores at baseline tended to have higher scores at follow-up. However, in all three SPADI classification trees higher pain self-efficacy influenced this relationship: for patients with high baseline pain or disability (cut-off points 75 and 62, respectively), higher pain self-efficacy (PSEQ≥48) reduced the likelihood of continued high levels of pain and disability at 6-month follow-up. Between 16% and 19% of participants were at risk of continued high levels of pain and disability (measured by SPADI pain and disability subscores) at 6 months due to (i) high baseline pain and disability and (ii) low pain self-efficacy. For patients with moderate levels of baseline pain and disability measured with the QuickDASH (41–59), the association was influenced by patient expectation: participants who expected to completely recover or much improve as a result of physiotherapy did better than patients who expected to only slightly improve, stay the same or worsen. Participants at risk of continued high levels of pain and disability at 6-month follow-up due to a lower expectation of recovery constituted 7% of our cohort at 6-month follow-up.

    Pain self-efficacy also influenced outcome for patients with low levels of baseline pain and disability: for patients with low baseline SPADI and SPADI pain scores (<68 and <75, respectively), low pain self-efficacy (PSEQ <40 and <41, respectively) increased the likelihood of persistent pain. Perhaps surprisingly, patients reporting low baseline pain but low pain self-efficacy (n=161, 20% of cohort) had a similar or worse outcome on the SPADI pain subscale to patients with high baseline pain but high pain self-efficacy (n=48, 6% of cohort).

    Our regression tree analyses provide a useful and simple clinical guide, highlighting the influence of patient beliefs and expectations of treatment on outcome, irrespective of baseline pain and disability. While these findings are consistent with those from the GLM analysis,13 the CART analysis selects variables based on prediction power rather than statistical significance or p values. Variables are included in order of importance; the most predictive variable is included first, the analysis then searches for the second most important variable among the rest and so on. The prediction error curve estimated using cross-validation gives a clear indication at what point in the selection process the additional predictors are not contributing enough to the prediction model. The prediction-based variable selection combined with cross-validation for assessing model performance ensures that only the relevant and most predictive variables are included in the optimal model.

    CART analysis has advantages over traditional regression modelling in that it does not require a specified distribution of outcome data or a large sample size.27 In terms of predictive power CART analysis is comparable to traditional modelling.21 However, CART does have limitations. Defining subgroups based on data-driven cut-off points for continuous measures (ie, the PSEQ) is subject to sampling variability, but the CART methodology does not provide a measure of uncertainty (eg, SEs or CIs) associated with the cut-off points. A different cut-off point may be selected in a different sample, but it was not our intention to provide a ready to deploy clinical tool with definitive cut-off points at this stage. We rather aimed to demonstrate that an easily interpretable prediction tool with potential for clinical applications can be developed which can be further examined in bigger and external cohorts to derive more generalisable cut-offs. However, use of cross-validation approach for model selection should make the derived models sufficiently robust at least for the population represented by the study cohort. Also, being a multicentre study with broad eligibility criteria increases the generalisability of the results to the wide range of patients and presentations of shoulder pain commonly seen by physiotherapists within primary and secondary care. This is further supported by similar patterns of the distributions of the QuickDASH outcome based on classification of participants using the SPADI trees in our informal internal validation.

    With regard to non-surgically managed shoulder pain, this study is one of only two known using a CART analysis to investigate the hierarchy of predictive factors associated with outcome. Vergouw28 compared the results of CART and logistic linear regression for 587 patients with musculoskeletal shoulder pain attending general practice in the Netherlands; however, they did not include patient expectation of change as a result of physiotherapy and pain self-efficacy. A positive association between patient expectation and outcome has been consistently reported for a range of health conditions,29–31 although ours is the first to investigate patient expectation of outcome in non-surgically managed shoulder pain. The association between pain self-efficacy and chronic non-cancer pain has also been consistently reported for a range of health conditions.32 Ours is one of only two studies to investigate self-efficacy in non-surgical management for shoulder pain. A randomised controlled trial of 102 participants33 did not find an association between baseline pain self-efficacy and the outcome of supervised exercise or radial extracorporeal shockwave therapy.

    Based on our findings that pain self-efficacy and patient expectation are important predictors of outcome, we recommend that they be formally assessed in all patients with musculoskeletal pain. There is currently no standardised method of measuring patient expectation and we therefore recommend using a patient-rated Likert scale that includes a worsening as well as improvement of shoulder pain.29 There are several validated measurement tools for pain self-efficacy and for the busy clinician we recommend using shortened patient-rated versions such as the PSEQ-234 comprising two items. Standardised questionnaires like the PSEQ-2 and a single question on expectation of outcome provide an opportunity to openly discuss patient beliefs and expectations, which healthcare practitioners may find challenging otherwise.35 36 Such patient-clinician dialogues around the potential impact of expectations and beliefs further supports shared decision-making. Our results suggest that cut-off points will vary according to baseline pain and disability and therefore the use of specific cut-off points for stratification is not justified. Further research is also needed to validate our point estimates in an external cohort.

    It is plausible that patient expectation and pain self-efficacy are mediating factors.29 Adherence to non-surgical management is reportedly low.37 The therapeutic effect of a home exercise and/or self-management programme cannot be realised if not enacted by the patient.38 One of the suggested mechanisms by which higher patient expectation is associated with outcome is through an increased motivation to engage and adhere to an intervention that participants believe will have a beneficial outcome.39

    Although not previously reported for those experiencing shoulder pain, high self-efficacy has been shown to be significantly associated with greater exercise adherence40–42 as well as other health behaviours such as physical activity43–45 and taking medications as prescribed.46 A consistent and statistically significant association between all three factors, changes in self-efficacy, adherence and outcome, has yet to be demonstrated. Further studies are needed to explore if moderating self-efficacy affects outcome.

    Further development and testing of educational interventions targeting healthcare practitioners with strategies to increase patients’ pain self-efficacy and expectations of treatment is needed. A number of promising interventions exist for increasing patients’ self-efficacy and include positive feedback on performance, observation of mastery in others, graded activity, identifying realistic goals for which the patient is likely to succeed and selecting tasks and activities relevant to the patient.32 43 Variability in reported effectiveness suggests that the purpose, content and delivery may need to be tailored to each patient, requiring a person-centred approach.

    Conclusion

    This is the first known study to subgroup people with shoulder pain of musculoskeletal origin attending physiotherapy into risk groups for persistent pain and disability based on a range of baseline personal, clinical, activity and participatory variables. This multicentre study provides evidence that for a given baseline measure of shoulder pain and disability, pain self-efficacy and patient expectation of change as a result of physiotherapy, are the most influential predictors of patient-rated outcome at 6-month follow-up. Additionally, this is the first study to demonstrate that for people with shoulder pain higher pain self-efficacy reduced the likelihood of continued high levels of pain and disability at 6-month follow-up, for those with high baseline pain or disability. The likelihood of persistent pain increased in the subgroup that were categorised as having low levels of baseline pain and disability and concomitant low pain self-efficacy. Of importance those identified as having low baseline pain and low self-efficacy had similar or worse outcome on the SPADI pain subscale to those with high baseline pain and high pain self-efficacy.

    Although our findings are applicable to people referred to physiotherapy for the management of shoulder pain of any duration and in primary and secondary care, they are likely to be applicable beyond this group.

    Based on our findings, we suggest that pain self-efficacy and patient expectation should be formally assessed and discussed at the first physiotherapy appointment. Further research should investigate whether these factors can be targeted and modified by therapeutic interventions and improve patient outcomes.

    Supplemental material

    Supplemental material

    Supplemental material

    Acknowledgments

    The authors would like to thank Christine Christopher and members of the study steering group, including representatives from the Public & Patient Involvement in Research (PPIRes) project, for their comments and suggestions during the study. The authors would also like to thank the Principal Investigators and physiotherapists who contributed and collected data for the study and to the patients who generously gave their time and participated in the study.

    References

    1. ↵
      2. Parsons S ,
      3. Ingram M ,
      4. Clarke-Conwell A , et al
      . A heavy burden. The occurrence and impact of musculoskeletal conditions in the United Kingdom today: The University of Manchester, 2011. Accessible online. http://www.medicine.manchester.ac.uk/musculoskeletal/aboutus/publications/heavyburden.pdf.
    2. ↵
      2. Pope DP ,
      3. Croft PR ,
      4. Pritchard CM , et al
      . Prevalence of shoulder pain in the community: the influence of case definition. Ann Rheum Dis 1997;56:308–12.doi:10.1136/ard.56.5.308
      OpenUrlAbstract/FREE Full Text
    3. ↵
      2. Urwin M ,
      3. Symmons D ,
      4. Allison T , et al
      . Estimating the burden of musculoskeletal disorders in the community: the comparative prevalence of symptoms at different anatomical sites, and the relation to social deprivation. Ann Rheum Dis 1998;57:649–55.doi:10.1136/ard.57.11.649
      OpenUrlAbstract/FREE Full Text
    4. ↵
      2. Bot SD ,
      3. van der Waal JM ,
      4. Terwee CB , et al
      . Incidence and prevalence of complaints of the neck and upper extremity in general practice. Ann Rheum Dis 2005;64:118–23.doi:10.1136/ard.2003.019349
      OpenUrlAbstract/FREE Full Text
    5. ↵
      2. Linsell L ,
      3. Dawson J ,
      4. Zondervan K , et al
      . Prevalence and incidence of adults consulting for shoulder conditions in UK primary care; patterns of diagnosis and referral. Rheumatology 2006;45:215–21.doi:10.1093/rheumatology/kei139
      OpenUrlCrossRefPubMedWeb of Science
    6. ↵
      2. van der Windt DA ,
      3. Koes BW ,
      4. de Jong BA , et al
      . Shoulder disorders in general practice: incidence, patient characteristics, and management. Ann Rheum Dis 1995;54:959–64.doi:10.1136/ard.54.12.959
      OpenUrlAbstract/FREE Full Text
    7. ↵
      2. Dorrestijn O ,
      3. Stevens M ,
      4. Winters JC , et al
      . Conservative or surgical treatment for subacromial impingement syndrome? A systematic review. J Shoulder Elbow Surg 2009;18:652–60.doi:10.1016/j.jse.2009.01.010
      OpenUrlCrossRefPubMedWeb of Science
    8. ↵
      2. Ketola S ,
      3. Lehtinen JT ,
      4. Arnala I
      . Arthroscopic decompression not recommended in the treatment of rotator cuff tendinopathy: a final review of a randomised controlled trial at a minimum follow-up of ten years. Bone Joint J 2017;99-B:799–805.doi:10.1302/0301-620X.99B6.BJJ-2016-0569.R1
      OpenUrl
    9. ↵
      2. Saltychev M ,
      3. Äärimaa V ,
      4. Virolainen P , et al
      . Conservative treatment or surgery for shoulder impingement: systematic review and meta-analysis. Disabil Rehabil 2015;37:1–8.doi:10.3109/09638288.2014.907364
      OpenUrlCrossRefPubMed
    10. ↵
      2. Kooijman M ,
      3. Swinkels I ,
      4. Dijk Cvan , et al
      . Patients with shoulder syndromes in general and physiotherapy practice: an observational study. BMC Musculoskelet Disord 2013;14:1–13.doi:10.1186/1471-2474-14-128
      OpenUrlCrossRefPubMed
    11. ↵
      2. Reilingh ML ,
      3. Kuijpers T ,
      4. Tanja-Harfterkamp AM , et al
      . Course and prognosis of shoulder symptoms in general practice. Rheumatology 2008;47:724–30.doi:10.1093/rheumatology/ken044
      OpenUrlCrossRefPubMed
    12. ↵
      2. Chester R ,
      3. Jerosch-Herold C ,
      4. Lewis J , et al
      . The SPADI and QuickDASH are similarly responsive in patients undergoing physical therapy for shoulder pain. J Orthop Sports Phys Ther 2017;47:538–47.doi:10.2519/jospt.2017.7195
      OpenUrl
    13. ↵
      2. Chester R ,
      3. Jerosch-Herold C ,
      4. Lewis J , et al
      . Psychological factors are associated with the outcome of physiotherapy for people with shoulder pain: a multicentre longitudinal cohort study. Br J Sports Med 2018;52:1–8.doi:10.1136/bjsports-2016-096084
      OpenUrl
    14. ↵
      2. Loh W-Y
      . Classification and regression trees. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2011;1:14–23.doi:10.1002/widm.8
      OpenUrl
    15. ↵
      2. Chester R ,
      3. Shepstone L ,
      4. Lewis JS , et al
      . Predicting response to physiotherapy treatment for musculoskeletal shoulder pain: protocol for a longitudinal cohort study. BMC Musculoskelet Disord 2013;14:1–8.doi:10.1186/1471-2474-14-192
      OpenUrlCrossRefPubMed
    16. ↵
      2. MacDermid JC ,
      3. Solomon P ,
      4. Prkachin K
      . The shoulder pain and disability index demonstrates factor, construct and longitudinal validity. BMC Musculoskelet Disord 2006;7:1–11.doi:10.1186/1471-2474-7-12
      OpenUrlCrossRefPubMed
    17. ↵
      2. Roach KE ,
      3. Budiman-Mak E ,
      4. Songsiridej N , et al
      . Development of a shoulder pain and disability index. Arthritis Care Res 1991;4:143–9.doi:10.1002/art.1790040403
      OpenUrlCrossRefPubMed
    18. ↵
      Institute for Work and Health. The QuickDASH outcome measure. Information for Users. 2006. Available at http://www.dash.iwh.on.ca/sites/dash/files/downloads (Accessed Nov 2016).
    19. ↵
      2. Artus M ,
      3. van der Windt DA ,
      4. Afolabi EK , et al
      . Management of shoulder pain by UK general practitioners (GPs): a national survey. BMJ Open 2017;7:e015711.doi:10.1136/bmjopen-2016-015711
    20. ↵
      2. Chester R ,
      3. Shepstone L ,
      4. Daniell H , et al
      . Predicting response to physiotherapy treatment for musculoskeletal shoulder pain: a systematic review. BMC Musculoskelet Disord 2013;14:203.doi:10.1186/1471-2474-14-203
      OpenUrlCrossRefPubMed
    21. ↵
      2. Breiman L ,
      3. Friedman J ,
      4. Stone CJ , et al
      . Classification and regression trees: CRC press, 1984.
    22. ↵
      2. Taxman FS ,
      3. Kitsantas P
      . Availability and capacity of substance abuse programs in correctional settings: a classification and regression tree analysis. Drug Alcohol Depend 2009;103:S43–53.doi:10.1016/j.drugalcdep.2009.01.008
      OpenUrlCrossRefPubMedWeb of Science
    23. ↵
      R Core Team R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. 2015. Accessible online: https://www.R-project.org
    24. ↵
      2. Therneau TM ,
      3. Atkinson EJ
      . An introduction to recursive partitioning using the RPART routines. Mayo Foundation: Technical report 1997;61:452.
      OpenUrl
    25. ↵
      2. Bandura A
      . Self-efficacy: toward a unifying theory of behavioral change. Psychol Rev 1977;84:191–215.doi:10.1037/0033-295X.84.2.191
      OpenUrlCrossRefPubMedWeb of Science
    26. ↵
      2. Nicholas MK
      . The pain self-efficacy questionnaire: Taking pain into account. Eur J Pain 2007;11:153–63.doi:10.1016/j.ejpain.2005.12.008
      OpenUrlCrossRefPubMedWeb of Science
    27. ↵
      2. Phelps MC ,
      3. Merkle EC
      . Classification and regression trees as alternatives to regression. Proceedings of the 4th Annual GRASP Symposium: Wichita State University, 2008.
    28. ↵
      2. Vergouw D
      . Methodological issues of clinical prediction models for shoulder pain in general practice: EMGO Institute for Health and Care Research, VU University, 2013.
    29. ↵
      2. Bialosky JE ,
      3. Bishop MD ,
      4. Cleland JA
      . Individual expectation: an overlooked, but pertinent, factor in the treatment of individuals experiencing musculoskeletal pain. Phys Ther 2010;90:1345–55.doi:10.2522/ptj.20090306
      OpenUrlAbstract/FREE Full Text
    30. ↵
      2. Bishop MD ,
      3. Mintken PE ,
      4. Bialosky JE , et al
      . Patient expectations of benefit from interventions for neck pain and resulting influence on outcomes. J Orthop Sports Phys Ther 2013;43:457–65.doi:10.2519/jospt.2013.4492
      OpenUrlCrossRefPubMed
    31. ↵
      2. Mondloch MV ,
      3. Cole DC ,
      4. Frank JW
      . Does how you do depend on how you think you’ll do? A systematic review of the evidence for a relation between patients' recovery expectations and health outcomes. CMAJ 2001;165:174–9.
      OpenUrlAbstract/FREE Full Text
    32. ↵
      2. Jackson T ,
      3. Wang Y ,
      4. Wang Y , et al
      . Self-efficacy and chronic pain outcomes: a meta-analytic review. J Pain 2014;15:800–14.doi:10.1016/j.jpain.2014.05.002
      OpenUrlCrossRefPubMed
    33. ↵
      2. Engebretsen K ,
      3. Grotle M ,
      4. Bautz-Holter E , et al
      . Predictors of shoulder pain and disability index (SPADI) and work status after 1 year in patients with subacromial shoulder pain. BMC Musculoskelet Disord 2010;11:1–9.doi:10.1186/1471-2474-11-218
      OpenUrlCrossRefPubMed
    34. ↵
      2. Nicholas MK ,
      3. McGuire BE ,
      4. Asghari A
      . A 2-item short form of the Pain Self-efficacy Questionnaire: development and psychometric evaluation of PSEQ-2. J Pain 2015;16:153–63.doi:10.1016/j.jpain.2014.11.002
      OpenUrl
    35. ↵
      2. Alexanders J ,
      3. Anderson A ,
      4. Henderson S
      . Musculoskeletal physiotherapists' use of psychological interventions: a systematic review of therapists' perceptions and practice. Physiotherapy 2015;101:95–102.doi:10.1016/j.physio.2014.03.008
      OpenUrlCrossRefPubMed
    36. ↵
      2. Singla M ,
      3. Jones M ,
      4. Edwards I , et al
      . Physiotherapists' assessment of patients' psychosocial status: are we standing on thin ice? A qualitative descriptive study. Man Ther 2015;20:328–34.doi:10.1016/j.math.2014.10.004
      OpenUrl
    37. ↵
      2. Peek K ,
      3. Sanson-Fisher R ,
      4. Mackenzie L , et al
      . Interventions to aid patient adherence to physiotherapist prescribed self-management strategies: a systematic review. Physiotherapy 2016;102:127–35.doi:10.1016/j.physio.2015.10.003
      OpenUrl
    38. ↵
      2. Hayden JA ,
      3. van Tulder MW ,
      4. Tomlinson G
      . Systematic review: strategies for using exercise therapy to improve outcomes in chronic low back pain. Ann Intern Med 2005;142:776–85.doi:10.7326/0003-4819-142-9-200505030-00014
      OpenUrlCrossRefPubMedWeb of Science
    39. ↵
      2. Flood AB ,
      3. Lorence DP ,
      4. Ding J , et al
      . The role of expectations in patients' reports of post-operative outcomes and improvement following therapy. Med Care 1993;31:1043–56.doi:10.1097/00005650-199311000-00006
      OpenUrlCrossRefPubMedWeb of Science
    40. ↵
      2. Medina-Mirapeix F ,
      3. Escolar-Reina P ,
      4. Gascón-Cánovas JJ , et al
      . Predictive factors of adherence to frequency and duration components in home exercise programs for neck and low back pain: an observational study. BMC Musculoskelet Disord 2009;10:155.doi:10.1186/1471-2474-10-155
      OpenUrlCrossRefPubMed
    41. ↵
      2. Dumoulin C ,
      3. Alewijnse D ,
      4. Bo K , et al
      . Pelvic-Floor-Muscle Training Adherence: Tools, Measurements and Strategies-2011 ICS State-of-the-Science Seminar Research Paper II of IV. Neurourol Urodyn 2015;34:615–21.doi:10.1002/nau.22794
      OpenUrl
    42. ↵
      2. Hay-Smith EJ ,
      3. McClurg D ,
      4. Frawley H , et al
      . Exercise adherence: integrating theory, evidence and behaviour change techniques. Physiotherapy 2016;102:7–9.doi:10.1016/j.physio.2015.08.006
      OpenUrl
    43. ↵
      2. Ashford S ,
      3. Edmunds J ,
      4. French DP
      . What is the best way to change self-efficacy to promote lifestyle and recreational physical activity? A systematic review with meta-analysis. Br J Health Psychol 2010;15:265–88.doi:10.1348/135910709X461752
      OpenUrlCrossRefPubMedWeb of Science
    44. ↵
      2. Bachmann C ,
      3. Oesch P ,
      4. Bachmann S
      . Recommendations for Improving Adherence to Home-Based Exercise: A Systematic Review. Physikalische Medizin, Rehabilitationsmedizin, Kurortmedizin 2018;28:20–31.doi:10.1055/s-0043-120527
      OpenUrl
    45. ↵
      2. Tierney S ,
      3. Mamas M ,
      4. Woods S , et al
      . What strategies are effective for exercise adherence in heart failure? A systematic review of controlled studies. Heart Fail Rev 2012;17:107–15.doi:10.1007/s10741-011-9252-4
      OpenUrlCrossRefPubMedWeb of Science
    46. ↵
      2. Náfrádi L ,
      3. Nakamoto K ,
      4. Schulz PJ
      . Is patient empowerment the key to promote adherence? A systematic review of the relationship between self-efficacy, health locus of control and medication adherence. PLoS One 2017;12:e0186458.doi:10.1371/journal.pone.0186458

    Footnotes

    • Contributors RC conceived the study. RC, LS, JSL and CJ-H contributed to the design of the study. RC directed the study and managed data acquisition. RC had full access to all of the data in the study and takes responsibility for the integrity of the data. MK undertook and LS provided expert advice on the statistical analyses. RC and MK drafted the manuscript. LS, JSL and CJ-H provided additional important intellectual and substantial scientific input throughout the study and on all redrafts of the report.

    • Funding RC was funded by the National Institute for Health Research (NIHR CAT CDRF 10–008). CJ-H was funded by the NIHR (NIHR- SRF-2012-05-119). The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report or decision to submit the article for publication. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. The views and opinions expressed therein are those of the authors and do not necessarily reflect those of the NIHR, NHS or the Department of Health.

    • Disclaimer The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report or decision to submit the article for publication. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. The views and opinions expressed therein are those of the authors and do not necessarily reflect those of the NIHR, NHS or the Department of Health.

    • Competing interests RC had support from a National Institute of Health Clinical Doctoral Research Fellowship for the submitted work.

    • Ethics approval This study was obtained from the National Research Ethics Service, East of England, Norfolk, UK in July 2011 (reference 11/EE/0212).

    • Provenance and peer review Not commissioned; externally peer reviewed.

    • Data sharing statement No additional data are available.

    • Patient consent for publication Not required.