The importance of emotional distress, cognitive behavioural factors and pain for life impact at baseline and for outcomes after rehabilitation – a SQRP study of more than 20,000 chronic pain patients

2.1 The Swedish Quality Registry for Pain Rehabilitation (SQRP)

The SQRP is a national registry based on Patient Reported Outcome Measures (PROM). All clinical departments within specialist care throughout Sweden provide data to the registry [32]. The SQRP (https://www.ucr.uu.se/nrs/) captures patients’ backgrounds, pain aspects, and psychological distress symptoms (e.g. depression and anxiety together with activity/participation aspects and health-related quality of life variables (Hr-QoL). Patients complete the SQRP questionnaires up to three times: at the first visit to the clinical department (baseline) and for those who participate in multimodal rehabilitation programs^[1] (MMRPs); immediately after the MMRP; and at a 12-month follow-up.

2.2 Subjects

Sweden is divided in 21 regions, which are responsible for health care in their geographical area. Unimodal rehabilitation interventions such as physiotherapy, occupational therapy and psychological treatment for patients with non-malignant chronic pain are provided within primary health care. In 2008, the Swedish government introduced a rehabilitation warranty, which included MMRPs delivered by primary health care. Hence, MMRP for patients with relatively uncomplicated clinical presentations were intended to be delivered by the primary care while patients with more complex pain conditions should be referred to specialist clinics for MMRP. However, the implementation of MMRP in primary care varied across the regions and when the rehabilitation warranty ended, access to MMRP within primary care even decreased in some regions. Most patients at the specialist clinics are referred from primary health care. A minority of patients are referred from other specialist clinics e.g. orthopedic and rheumatology clinics. The proportions of patients within primary health care with chronic pain conditions are not exactly known but 10–20% are estimates [40], [41]. Furthermore, the proportion of chronic pain patients within primary health care that are referred to specialist clinics is not known.

This study included SQRP data from women and men ≥18 years old who had chronic (≥3 months) non-malignant pain. Patients were selected to participate in SQRP if they had complex chronic pain conditions, were in need of a bio-psycho-social assessment and were referred to a specialist clinic corresponding to SQRP categories between 2008 and 2016. These patients were characterized as complex because they had psychological comorbidities with difficult relationships with pain aspects, their condition severely affected their working life and participation in social activities, and/or they did not respond to routine pharmacological/physiotherapeutic treatments delivered in a monodisciplinary fashion. Exclusion criteria were substance abuse and ongoing major somatic or psychiatric disease. The total number of patients referred to specialist clinics is not known, but the steering committee for SQRP has estimated the response rate of >90% for complex chronic pain conditions. In the present study, an additional requirement was that patients had completed one or two cognitive behavioural instruments reflecting emotion regulation aspects (see below).

2.3 MMRP

The SQRP does not provide detailed information about the MMRPs at the individual centres. Generally, the programs are delivered by a team of professionals (most often including a physician, a psychologist, an occupational therapist, a physiotherapist, and a social worker) and are based on a bio-psycho-social model of chronic pain [42]. In Sweden, MMRPs are mainly out-patient group-based programs. In addition, MMRPs provide opportunities for individual interventions based on the clinical picture and the aims of the patient. The group intervention includes cognitive behavioural treatment, physiotherapy (including physical exercise), interventions targeting improved ergonomics, and occupational therapy. In addition, basic pain physiology and pain management lectures are offered to patients, relatives, friends, and colleagues. The programs generally last several weeks (4–8 weeks) and include group-based activities (20–30 h/week). In addition, patients might be asked to complete tasks at home such as physical exercise.

2.4 Variables

Pain, life impact, background, emotional distress, and cognitive behavioural variables were selected from SQRP and used in the analyses. Data were retrieved from SQRP on three occasions (see above).

2.5 Statistics

All statistics were performed using the statistical package IBM SPSS Statistics (version 24.0) and SIMCA-P+ (version 13.0; Umetrics Inc., Umeå, Sweden). As this study had a large number of subjects, a probability of <0.001 (two-tailed) was considered significant. The mean value, median, and range of continuous variables are reported as ±one standard deviation (±1 SD). Categorical variables are reported as percentages (%). SQRP uses predetermined rules when handling single missing items of a scale or a subscale; details about this have been reported elsewhere [54]. For group comparisons, student’s t-test (un-paired observations), analysis of variance (ANOVA), and Chi-square test were used. Pearson’s test was used for bivariate correlation analysis. Between group effect sizes (ES; Cohen’s d) were computed when appropriate using a calculator (https://www.socscistatistics.com/effectsize/default3.aspx). When calculating within group ES (Hedges’ g) we used the following calculator: https://webpower.psychstat.org/models/means01/effectsize.php. The absolute ES was considered very large for ≥1.3, large for 0.80–1.29, moderate for 0.50–0.79, small for 0.20–0.49, and insignificant for <0.20 [55].

Common methods such as logistic regression (LR) and multiple linear regression (MLR) can quantify the level of relations of individual factors but disregard interrelationships among different factors and thereby ignore system-wide aspects [56]. Moreover, such methods assume variable independence when interpreting results [57] and there are several risks when considering one variable at a time [58]. Because of the context of our aims, the problems handling missing data (see below), and the obvious risks for multicollinearity, we used advanced multivariate analyses (MVDA). Hence, advanced principal component analysis (PCA) for the multivariate correlation analyses of all investigated variables and Orthogonal Partial Least Square Regressions (OPLS) for the multivariate regressions were applied using SIMCA-P+. These techniques do not require normal distributions of the included variables [59]; furthermore, in contrast to LR and MLR, these techniques take advantage of the fact that the regressors may be highly correlated so multicollinearity is not an issue.

PCA was used to extract and display systematic variation in the investigated data matrix. If necessary, all variables were log transformed before the statistical analyses. A cross validation technique was used to identify nontrivial components (p). Variables loading on the same component p were correlated, and variables with high loadings but with opposing signs were negatively correlated. Variables with high absolute loadings were considered significant. The obtained components are per definition not correlated and are arranged in decreasing order with respect to explained variation. R² describes the goodness of fit – i.e. the fraction of the sum of squares of all the variables explained by a principal component [60]. Q² describes the goodness of prediction – i.e. the fraction of the total variation of the variables that can be predicted by a principal component using cross validation methods [60]. Outliers were identified using two methods: score plots in combination with Hotelling’s T² and distance to model in X-space. No extreme outliers were detected.

OPLS was used for the multivariate regression analyses of MPI-Pain-interfer and MPI-LifeCon [60]. Variable Influence on Projection (VIP) indicates the relevance of each X-variable pooled over all dimensions and Y-variables – the group of variables that best explain Y. VIP≥1.0 was considered significant if VIP had 95% jack-knife uncertainty confidence interval non-equal to zero. P(corr) was used to note the direction of the relationship (positive or negative). This is the loading of each variable scaled as a correlation coefficient and thus standardizing the range from −1 to +1 [59]. P(corr) is stable during iterative variable selection and comparable between models. An absolute P(corr)>0.5 is considered significant [59]. A regressor was considered significant when VIP>1.0 and absolute P(corr)>0.50. R², Q², and the p-value of a cross-validated analysis of variance (CV-ANOVA) are reported for each regression. SIMCA-P+ uses the NIPALS algorithm when handling missing data (variables/scales: max 50% missing data and subjects: max 50% missing data).

To identify subgroups at baseline, we performed PCA based on relevant variables, followed by hierarchical clustering analysis (HCA) and, based on the identified clusters (subgroups) defined by HCA, partial least squares – discriminant analysis (PLS-DA). We also applied a bottom-up HCA to the principal component score vectors using the default Ward linkage criterion to identify relevant subgroups of patients. HCA complements PCA in the sense that while PCA identifies distinct clusters in multivariate space, HCA can find subtle clusters. In the resulting dendrogram, clusters were identified; based on these groups, PLS-DA was performed using group belonging as Y-variables and psychometric data as predictors (X-variables). The PLS-DA model was computed to identify associations between the X-variables and the subgroups. These associations were visualized on a corresponding loading plot. Based on the groups defined by HCA, traditional inferential statistics were computed using SPSS.

3.1 Investigated variables

This study includes 22,406 chronic pain patients. The investigated variables are displayed in Table 1. Most patients were women. Significant gender differences were found for place of birth and education: more men were born outside Europe (men: 15.2% vs. women: 10.6%) and more women had a university education (men: 17.8% vs. women: 25.7%) (Table 1). Significant gender differences were found except for anxiety, pain interference, and life control (Table 1). According to the ES, the significant gender differences were moderate for pain spreading (PRI) and generally non-significant or small for the other variables (Table 1).

3.2 Multivariate correlation patterns

A mix of variables showed large variations across patients and had significant importance for the clinical presentation. In addition, these variables were significantly intercorrelated. Hence, a PCA (R²=0.32, Q²=0.22) of the variables under investigation revealed one significant component (p(1)) (Fig. 1). According to p(1), the most important variables (i.e. absolute loadings above 0.25) with positive loadings were the two acceptance scales and life control, whereas the most important variables with negative loadings were pain severity, pain intensity, pain interference, the two emotional variables, and fear of movement/(re)injury. Age, pain spreading, and the categorical variables gender, University and outside-Europe had little importance in the multivariate context (all with absolute loadings <0.15) and hence only showed low correlations with the other variables (Fig. 1).

Fig. 1:

Loading plot (i.e. relations between variables) from the PCA of background, pain, psychological distress, emotion regulator, and life impact variables (n=22,370). The horizontal axis is the first component (p(1)) and the vertical axis the second component (p(2)); note that only one significant component p(1) was obtained (horizontal axis).

PCA=principal component analysis; Outside_Europe=born outside Europe; University=university education; NRS_7days=pain intensity previous 7 days according to a numerical rating scale; MPI=Multidimensional Pain Inventory; MPI_Painseveri=the pain severity scale of MPI; PRI=spreading of pain on the body; HAD=the Hospital Anxiety and Depression Scale; HAD_A=the anxiety scale of HAD; HAD_D=the depression scale of HAD; CPAQ=Chronic Pain Acceptance Questionnaire; CPAQ_AE=the activity engagement scale of CPAQ; CPAQ_PW=the pain willingness scale of CPAQ; TSK=Tampa Scale for Kinesiophobia, measures fear of movement/(re)injury; MPI_Paininterfer=the pain interference scale of MPI; MPI_LifeCon=the life control scale of MPI.

3.3 Regressions of pain interference and life control

The associations between Pain interference and life control and the other variables were investigated in two regressions (Table 2). For both, regressions were highly significant. Pain interference correlated with a mix of factors (i.e. cognitive behavioural variables, pain intensity aspects, and emotional distress factors) (Table 2, left part). Hence, CPAQ-AE showed the strongest association followed by pain severity and depression. The significant pain intensity variables (MPI-Pain-severity and NRS-7days) and depression had positive correlations with pain interference, and the two acceptance variables correlated negatively with the dependent variable pain interference.

The analysis of life control showed that the two emotional distress variables (i.e. the HAD scales) had the strongest correlations (Table 2, right part). However, activity engagement according to CPAQ and the two pain intensity variables (MPI-pain severity and NRS-7days) also correlated significantly with life control. Activity engagement showed a positive correlation with life control, but the other significant regressors correlated negatively.

In both analyses, the fear of movement/(re)injury variable TSK, spreading on the body nor the sociodemographic variables were significant.

3.4 Identifying clusters at baseline

In order to identify clusters/subgroups of patients a HCA based on a PCA of the most important variables was performed (Fig. 1). Hence, the variables used as input variables were the two pain intensity variables (MPI-Pain-severity and NRS-7days), the two emotional distress variables (HAD-A and HAD-D), the two acceptance (CPAQ) variables, pain interference and life control. Three subgroups/clusters were identified.

Based on the HCA, a PLS-DA model was obtained with group belonging as Y-variable. The model had two latent variables (R²=0.41, Q²=0.41, CV-ANOVA p<0.001) and the three groups are visualized in Fig. 2. The corresponding loading plot of the PLS-DA model (i.e. the variables in relation to the three identified subgroups) is shown in Fig. 3.

Fig. 2:

Score plot (i.e. relations between patients) of the PLS-DA model showing the three clusters of patients; each dot represents a patient. The two axes – i.e. scores t[1] on the horizontal axis and t[2] on the vertical axis – represent the two latent variables of the model. The latent variables are mathematical constructs that “summarise” the variables registered in the study. PLS-DA: partial least squares – discriminant analysis. Green=group 1, Red=group 2, and Blue=group 3.

Fig. 3:

Loading plot of the PLS-DA model. The loading plot is complementary to the score plot in Fig. 2 and summarises how the X-variables relate to each other as well as to group belonging (Y-variable symbolized by a group dot in blue: SM7.DA1=group1, SM7.DA2=group 2, and SM7.DA3=group 3). X-variables located near a group dot are positively associated with that group. PLS-DA: partial least squares – discriminant analysis. NRS_7days=pain intensity previous 7 days according to a numerical rating scale; MPI=Multidimensional Pain Inventory; MPI_Pain_severity=the pain severity scale of MPI; PRI=spreading of pain on the body; HAD=the Hospital Anxiety and Depression Scale; HAD_A=the anxiety scale of HAD; HAD_D=the depression scale of HAD; CPAQ=Chronic Pain Acceptance Questionnaire; CPAQ_AE=the activity engagement scale of CPAQ; CPAQ_PW=the pain willingness scale of CPAQ; MPI_Pain_interfer=the pain interference scale of MPI; MPI_LifeCon=the life control scale of MPI.

The identified clusters were compared for the input variables and those variables not included in the HCA and PLS-DA by omnibus statistical testing including post-hoc tests (Table 3). According to the input variables, cluster 1 was associated with the worst situation, cluster 2 with the best situation (Table 3), and cluster 3 with the intermediary situation. The pair-wise ES of Cohen’s |d| for the identified clusters are shown in Table 4. A very large ES for the input variables was found between cluster 1 and 2. Between cluster 1 and cluster 3, the ES was large to very large for the input variables. Between cluster 2 and cluster 3, a more heterogeneous picture emerged: i.e. very large ESs for the two pain intensity variables and pain interference, large for CPAQ-AE and life control, and small to moderate for the rest of the input variables.

Table 4:

Pair-wise effect sizes by Cohen’s d (absolute figures) for variables included and not included (variable name in bold type) in the identification of the three clusters.

1. Interpretation of Cohen’s |d|:

2. Black cell: ≥1.3, i.e. very large effect size.

3. Dark grey cell: 0.80–1.29, i.e. large effect size.

4. Grey cell: 0.50–0.79, i.e. moderate effect size.

5. Light grey cell: 0.20–0.49, i.e. small effect size.

6. White cell: <0.20, i.e. insignificant effect size.

7. Outside-Europe=born outside Europe; University=university education; NRS-7days=pain intensity previous 7 days according to a numerical rating scale; MPI=Multidimensional Pain Inventory; MPI-Pain severity=the pain severity scale of MPI; PRI=spreading of pain on the body; HAD=the Hospital Anxiety and Depression Scale; HAD-A=the anxiety scale of HAD; HAD-D=the depression scale of HAD; CPAQ=Chronic Pain Acceptance Questionnaire; CPAQ-AE=the activity engagement scale of CPAQ; CPAQ-PW=the pain willingness scale of CPAQ; TSK=Tampa Scale for Kinesiophobia, measures fear of movement/(re)injury; MPI-Pain-interfer=the pain interference scale of MPI; MPI-LifeCon=the life control scale of MPI.

For the variables not used as input variables when identifying the clusters, cluster 1 to a greater extent included more subjects born outside Europe, fewer subjects with a university education, more subjects with spreading of pain (PRI), and more subjects with higher fear of movement/(re)injury according to TSK. Subjects belonging to cluster 2 reported the least spreading of pain (PRI), the lowest fear of movement/(re)injury, the highest proportion with university education, and the lowest proportion born outside Europe. Cluster 3 were intermediary to cluster 1 and 2 with respect to these variables. When scrutinizing ES for the variables not used as input variables when identifying the clusters in all comparisons, age had insignificant ES, spreading of pain had small to medium ES, and fear of movement/(re)injury had moderate to very large ES (Table 4).

3.5 Comparing the three clusters after MMRP and at the 12-month follow-up

There were significant differences in proportions participating in MMRP (cluster 1: 32.5%, cluster 2: 31.2%, and cluster 3: 37.5%; p<0.001; cluster 3 differed significantly from clusters 1 and 2). The pattern of the input variables at baseline remained for those in the three clusters participating in MMRP when evaluated both immediately after MMRP and at the 12-month follow-up (Table 5; left part). In fact, the three clusters were significantly different for all input variables at these two time-points.

The changes in the input variables after MMRP were larger for cluster 1 than cluster 3, and cluster 2 had the smallest changes. The same pattern was basically found at the 12-month follow-up (Table 5, right part). In agreement with these observations we found that the within-group ES were consequently higher i.e. generally moderate (d>0.50) in cluster 1 both immediately after MMRP (Table 6 and Fig. 4) and at the 12-m follow-up (Table 6). The lowest within-group ES were found in cluster 2 and with cluster 3 as intermediary both after MMRP (Fig. 4) and at the 12-month follow-up. The ES in cluster 3 were generally small while the ES in cluster 2 generally were insignificant (ES<0.20) except for the two CPAQ variables.

Fig. 4:

Within-group absolute effect sizes (ES; Cohen’s d) for the three clusters.

When scrutinizing the pairwise ES comparisons between clusters were found that cluster 1 vs. cluster 2 showed the highest ES (Table 6): moderate ESs for four out of eight variables (NRS-7days, MPI-Pain-severity, HAD-D and MPI-LifeCon) after MMRP while the corresponding figure was three out of eight variables (NRS-7days, HAD-D and MPI-LifeCon) at the 12-month follow-up. The other variables showed small ESs for this comparison.

4.1 Variations in clinical presentation

The three cognitive behavioural variables (TSK, CPAQ-AE, and CPAQ-PW) had relatively high absolute loadings on the significant component of the PCA (Fig. 1) and hence contributed to the variations in clinical presentation. TSK and CPAQ had different signs for their loadings and were thus negatively correlated, as previously reported [36], [61]. The two pain intensity variables, the anxiety and depression subscales of HADS and the two life impact variables were also of importance for the clinical presentation and correlated with the cognitive behavioural variables since they all had high absolute loadings on the significant component (Fig. 1). Hence, the results from this very large sample of patients confirm that the clinical picture is multifaceted.

In contrast, pain spreading, sociodemographic variables, and gender contributed relatively little to the variation in clinical presentation. Relatively small differences were found for gender (i.e. non-significant or small ES except for pain spreading) (Table 1). Female patients had greater spread of pain (Table 1), which agrees with other studies [62], [63].

The results of the PCA support the bio-psychological parts of the bio-psycho-social model of chronic pain. However, several relevant aspects of the social context may not have been included as variables but may nevertheless be of importance. Indeed, as discussed below, the identified clusters showed prominent differences in the social variables investigated.

4.2 Variables associated with pain interference and life control

The relative importance of the regressors (x-variables) differed somewhat between the two analyses (Table 2), but an important conclusion is that a blend of aspects (pain intensity/severity, emotion, and acceptance) is important for the two aspects of life impact. Our results agree with other studies that show that a mix of variables (pain intensity, emotion variables, fear, etc.) influence disability aspects such as life impact variables [32], [35], [64]. Thus, life impact aspects are not mainly determined by one variable. The two acceptance variables of CPAQ were significantly associated with life impact; activity engagement was in fact the strongest regressor of pain interference. When scrutinizing the individual items of CPAQ and the two life impact variables, there are several similarities and thus the important role of the activity engagement subscale of CPAQ could be because this variable is a bit “contaminated” with pain interference and life impact. Another observation from this mix of intercorrelated variables important for the variations in the two life impact variables is their different signs. Hence, while high pain intensity and emotional distress were positively associated with pain interference and low life control, the acceptance variables had, as previously noted, an opposite negative effect that possibly moderated the effects of pain intensity and emotional distress [4], [65].

Although the regression analyses in this large sample of patients were highly significant, only 43–53% of the variations in the two life impact variables were explained (Table 2). Hence, we have an incomplete understanding of what factors (except for pain intensity, emotion, and cognitive behaviour) determine the aspects of life impact. A broader selection of independent variables may increase the explained variation, including work environment factors, sick-leave use and duration, catastrophizing aspects, and family relationships.

In the multivariate context, the fear of movement/(re)injury variable in TSK was borderline significant in the regressions of life impacts (Table 2). On a superficial level, it may seem to contradict a meta-analysis of pain-related fear and disability, which concluded a moderate to large positive association [66], but our results for this variable must be seen in a relative context of other cognitive behavioural variables tapping, at least partly, into the same underlying phenomenon. Moreover, TSK showed significant correlations with pain variables and emotional distress variables according to the PCA (Fig. 1). Hence, although the TSK was not retained in the multivariate regression analyses, it cannot be concluded that the fear of movement/(re)injury variable generally is unimportant.

4.3 Identification of subgroups at baseline

Patients within a certain diagnosis such as chronic low back pain can differ considerably in clinical presentation. In contrast to most other subgrouping studies, this large study used objective methods (i.e. PCA) to determine the input variables from a larger set of PROM variables and used advanced multivariate methods to identify the actual subgroups. Interestingly, this approach identified pain intensity/severity variables as clinically important for the clinical presentation; therefore, these were included as input variables, which partially contrasts with studies that only consider psychological variables [34], [35]. Patients appear to consider pain intensity as a very important aspect of their clinical situation [67]. The present study, as with another SQRP study, confirms that pain intensity/severity in fact carries important information [32].

At baseline, three subgroups with prominent differences in the input variables between cluster 1 and 2 were found; cluster 3 was an intermediary group. The ES between cluster 1 and 2 were generally very large (Table 4), while the ES was somewhat lower in the other pair-wise comparisons.

Cluster 1 had the worst situation with respect to the input variables and showed significantly more pain spreading (i.e. higher PRI) and higher fear of movement/(re)injury than cluster 2. In addition, other studies have identified a subgroup associated with high pain intensity, prominent emotional distress, cognitive behavioural factors, and disability [35], [36], [38], [68], [69], [70].

Previous studies using data from chronic pain patients at one or two clinical departments within SQRP have also found a subgroup (in the present study subgroup 2) with a relatively good situation – i.e. lower pain intensity, less psychological distress, and better coping skills [36], [38]. Furthermore, these studies and the present study identified intermediary groups. In a recent nationwide study (>35,000 patients) using the SQRP, two clusters were identified that differed significantly on emotional distress, but that study did not include cognitive behavioural variables [71]. In relation to this study, a reasonable conclusion is that cognitive behavioural factors such as acceptance further differentiate the clinical presentation beyond emotional distress and should be taken into account.

The clinical pictures of the three clusters were associated with relatively prominent differences in sociodemographic characteristics. This finding agrees with our earlier study that used emotional distress and fear of movement/(re)injury for subgrouping [36]. Hence, nearly 20% of the patients in the subgroup with the worst situation (i.e. subgroup 1) came from outside Europe compared to 3.6% in cluster 2. The highest proportion of university education was found in cluster 2 (36%) and the lowest in cluster 1 (18%). These circumstances may be important to assess and may influence the choice and planning of interventions.

4.4 MMRP outcomes in the three subgroups

The present study shows that the clinical situation at baseline gives clues to what outcomes will follow after MMRP. Hence, the most marked improvements and associated with the largest ES both immediately after MMRP and at the 12-month follow-up were found in cluster 1. That is, the subgroup with the worst situation at baseline and the smallest were found in cluster 2 (associated with the relatively best situation) and with cluster 3 as intermediary. While it could be that the relatively greater improvement in cluster 1 is due to regression to the mean, this cluster still improves at least as well from MMRP as those with markedly less emotional distress (cluster 2). SRs report that MMRP is an effective intervention with small to moderate effects for patients with chronic pain conditions [72], [73], [74], [75]. These results can be seen in the perspective of the absence of effects, small effects and/or lack of long-term follow-up according to SRs for common pain medications e.g. paracetamol, non-steroidal anti-inflammatory drugs and opioids [76], [77], [78]. Patients participating in randomized controlled trials (RCTs) are not necessarily representative of patients in clinically representative settings [79]. Hence, an intervention must show efficacy both in RCTs under highly controlled conditions and their effectiveness under clinically representative conditions [79], [80]. In the present study moderate within-group ES were found in cluster 1 and small to moderate ES in cluster 3. Hence, our results for cluster 1 and 3 confirm the results of SRs. In cluster 2 insignificant and small ES were generally found except for the two scales of CPAQ. The small to moderate ES for pain intensity aspects in cluster 1 and 3 agree with other studies in clinical routine care (n: 65–395). Hence, for long term follow up (6–12 months) such studies report small (d: 0.20–0.33 [81], [82]) to moderate (d: 0.59–0.70 [80], [83], [84]) ES for pain intensity. For psychological distress variables these studies generally have found small ES for long term follow up i.e. d: 0–0.38 for depressive symptoms [80], [81], [82] and d: 0.22–0.34 for anxiety [80], [82]. In the present study we observed moderate ES in cluster 1 and small ES in cluster 3. Moreover, an important observation from the present study is the broad effects of MMRP found in cluster 1 at 12-month follow-up i.e. all variables including disability aspects (i.e. Pain interference and Life control) show moderate ES. A similar pattern but with lower ES (small to moderate are observed in cluster 3). Cluster 2 shows a more heterogenous picture and with several insignificant ES.

Partially, our results challenge the conclusion that previous treatments based on subgroupings result have given limited success [85]. The relevance of offering extensive MMRP to patients belonging to cluster 2 must be questioned both due to the small and partly clinically insignificant effects on outcomes and for economic reasons [70]. For this cluster, relevant long-term changes seem to centre around pain acceptance. Possibly less extensive and more targeted intervention would suffice and patients in this cluster would have been good candidates for primary care based multimodal interventions. Hence, general recommendations for MMRP for patients with chronic pain are not justified according to our results [86]. This study gives some support to the hypothesis that a more appropriate matching between clinical presentation and the choice of an intervention improves outcomes [87]. A consequence of the present study is that patients belonging to cluster 1 need to be prioritized if marked improvements are required on a group level for MMRP. The present subgrouping may be an important pathway into more personalized practice of pain medicine within the field of MMRP [38] and reduce the trial-and-error basis for choice of treatments that depend on the experience of healthcare providers [86]. Severe emotional distress has been associated with worse treatment outcomes [8], [14], [88], [89], although the literature is not in total agreement [90], [91]. The fact that cluster 1, which had the most severe anxiety and depression symptoms, had larger improvements for all variables may seem unexpected, but it is important to recognize that MMRPs contain Cognitive Behavioural Therapy (CBT) interventions. As expected, the pattern of improvements in outcomes is not evident for MMRP management since relatively more patients in cluster 3 (the intermediary cluster) than in cluster 1 participated in MMRP. An alternative or complementary explanation is that patients, while highly distressed, also have a high somatic focus (fear of movement/(re)injury) and low acceptance of pain and therefore may refuse an offer of self-management interventions with a psychological focus and favour a more biomedically-oriented intervention instead. In our earlier study, we found that patients belonging to the subgroup with the worst situation participated in MMRP less often [36].

Despite these improvements, the principle pattern of the input variables at baseline remained after MMRP and at the 12-month follow-up in the three clusters; that is, cluster 1 still had the worst situation. Recently, two studies found a similar pattern, although these studies rely on smaller sample sizes [36], [92]. Nonetheless, this pattern may reflect the high baseline levels of psychological distress and pain intensities in cluster 1. The question arises whether further optimisation of the psychological interventions or other components such as exercise and education in MMRP will increase the outcomes for cluster 1 patients. When it comes to the psychological components, treatment that focuses on transdiagnostic factors such as emotion-focused exposure and acceptance-related strategies have produced promising results [93], [94]. Relatively large ESs were seen for acceptance within all the different clusters in this study indicating this may be a key process contributing to treatment outcomes throughout the clusters. Increased targeting of this process may help to improve treatment outcomes. Taken together, it seems clear that interventions need to be individualized to target the needs of the individuals within each cluster and these interventions may need to be different for each cluster. From several points of views, it seems that the most expensive resources should be offered to those patients with the most complicated clinical picture. For patients with a relatively less complicated clinical situation, Internet-based interventions may be an option [95]. Hence, clinical departments should offer different levels of MMRP in their range of treatments based on detailed assessments. Although improvements occurred in the three clusters, it can be questioned whether MMRP really affects neurobiological components of pain such as comprehensive alterations in peripheral tissues (e.g. muscles and blood), sensitisation mechanisms, altered descending control, and functional and morphological central nervous system alterations.

4.5 Strengths and limitations

A strength of this study is its large number of chronic pain patients with nation-wide representation. However, the patients represent a selection of the most difficult cases, so our results cannot be generalized. Another strength is the use of advanced multivariate statistics to handle obvious risks for multicollinearity. Although we used validated and well-known PROM instruments, a limitation might be that self-reports can be influenced by perceptions of social desirability, and changes in the social context over time may have occurred. No control group or treatment-as-usual group was available, which ethically is complicated to arrange for a registry of real-world practice patients, although this might have influenced our interpretation of improvements after MMRP. Another limitation is that individual differences in adherence to MMRP exist. Moreover, differences in the exact composition of MMRP exist among centres even though the general content of the MMRP is the same irrespective of where it is used.