The SQRP registers Patient-Reported Outcome Measures (PROMs) data from a majority of the Swedish specialist chronic pain departments (35). This study investigates adult patients (i.e., ≥18 years) with chronic (≥3 months) non-malignant pain registered in the SQRP between 2008 and 2016. Most of the patients were referred by primary care, but some were referred by rheumatology and orthopaedic specialist departments. Patients enrolled in SQRP have complex pain conditions that often include comorbidities (e.g., depression and anxiety), insufficient coping strategies, long sick leave, low social participation, and/or unresponsiveness to monomodal treatments (e.g., pharmacological, and physiotherapeutic interventions). As SQRP is a clinical registry of patients with chronic pain conditions, the SQRP does not have strict inclusion and exclusion criteria. General inclusion criteria include disabling chronic pain (i.e., chronic pain that results in sick leave or major interference in daily life), 18 years and older, medically fully investigated, and written consent to participate. General exclusion criteria include severe psychiatric morbidity, abuse of alcohol and/or drugs, diseases that do not allow physical exercise, and specific pain conditions with other treatment options available (i.e., conditions associated with “red flags”).

The study was conducted in accordance with the Helsinki Declaration. It was approved by the Ethical Review Board in Linköping (Dnr: 2015/108-31). All the participants received written information about the study and gave their written consent.

The PROMs capture a patient's background including demographic aspects, pain intensity, pain-related cognitions, and psychological distress symptoms as well as activity/participation aspects and health-related quality of life variables. PROMs are completed by patients on up to three occasions: before the first visit (baseline assessment), immediately after treatment for those who participated in interdisciplinary pain rehabilitation programs (IPRP; approximately 50%–55% of the patients assessed), and at a 12-month follow-up. In the present study, only baseline data are analysed. In the present baseline study, 19 variables were initially considered for inclusion in the PLS-SEM analyses – i.e., six background variables and 13 of 22 mandatory outcome variables (i.e., variables that all participating clinics/units collect from patients. In addition to this, there are other – non-mandatory – variables that the clinic/unit can include if desired) (Table 1).

We used three statistical packages: IBM SPSS Statistics (version 28.0; IBM Corporation, Route 100 Somers, New York, United States); SIMCA-P + (version 17.0; Sartorius Stedim Biotech, Umeå, Sweden); and Smart-PLS version 3 (Ringle, C. M., Wende, S., and Becker, J.-M. 2015. “SmartPLS 3” Boenningstedt: SmartPLS GmbH, http://www.smartpls.com). In the text and tables, the mean value ± one standard deviation (±1 SD) of continuous variables and percentages (%) for categorical variables are reported. To compare groups, Student's t-test for independent samples and Chi squared tests were applied. In large samples, small differences may be significant and effects size can be used to evaluate the clinical importance of the significant differences. Effect sizes (ES; Cohen's d) for within group analyses were computed and Hedges' g – a measure of effect size weighted by the relative size of each sample – was used for between group ESs. The absolute effect size was considered clinically insignificant for <0.20, small for 0.20–0.49, moderate for 0.50–0.79, and large for ≥0.80 (69).

The retrieved SQRP data included missing data (Table 1). Our analysis of missingness mechanisms indicated that data were not missing completely at random (MCAR) (i.e., Little's Test of Missing Completely at Random was not significant) and therefore data were missing either at random (MAR) or not missing at random (NMAR) (70). One choice may have been to use listwise deletion, but, multiple imputation gives less biased results than listwise deletion (70). The multiple imputation option in SPSS was used for the imputation of our variables (Random generator: Mersenne Twister with a fixed starting point for the active generator initialization; all variables were included in five imputations, which were pooled). The imputed data are shown in Table 3.

Previous studies have discussed the necessity of using advanced multivariate analyses (MVDA) when accounting for system-wide aspects and multicollinearity problems (53, 71). Advanced Principal Component Analysis (PCA), which is available in SIMCA-P+, can be used to determine multivariate outliers and multivariate correlation patterns. Outliers were identified using score plots in combination with Hotelling's T² and distance to model in X-space. PCA can conceptually be viewed as a kind of multivariate correlation analysis that extracts and displays systematic variation in the data matrix. To identify nontrivial components (p), we used a cross validation technique. Variables loading on the same component p are correlated; variables with high loadings but opposing signs are negatively correlated. Variables with high absolute loadings were considered significant. The obtained components are not correlated but are arranged in decreasing order with respect to explained variation. R² describes the goodness of fit – the fraction of sum of squares of all the variables explained by a principal component (72). Q² describes the goodness of prediction – the fraction of the total variation of the variables that can be predicted using principal component cross validation methods (72).

Orthogonal Partial Least Square Regressions (OPLS) was used for the multivariate regression analyses of pain intensity and psychological distress variables. The 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 explains Y (72). A variable was considered significant when VIP > 1.0 (or VIPpred if > one component) and had a 95% jack-knife uncertainty confidence interval non-equal to zero. P(corr) was used to note the direction of the relationship (positive or negative) – i.e., the loading of each variable was scaled as a correlation coefficient and therefore standardised the range from −1 to +1 (71). Thus, a variable/regressor was considered statistically significant when VIP or VIPpred > 1.0. A regression model was obtained, including one or several components (the first is always the predictive component), if certain predefined criteria were fulfilled. The validity of the model was estimated using cross validation. Hence, for each regression, we report R², Q², and the P-value of a cross-validated analysis of variance (CV-ANOVA).

SMART-PLS version 3 (Ringle, C. M., Wende, S., and Becker, J.-M. 2015. “SmartPLS 3.” Boenningstedt: SmartPLS GmbH, http://www.smartpls.com) was used for the Partial Least Squares Structural Equation Modelling (PLS-SEM). This non-parametric method enables researchers to simultaneously model and estimate complex relationships among multiple dependent and independent variables. The method has a causal-predictive approach to SEM, which intends to explain the variance of the dependent variables; for details of this method, see (46). The recommendations by Hair Jr et al. were followed when performing the PLS-SEM analyses (46). This method is based on constructs (latent variables) that are generally associated with several indicators. The indicators of a certain component must have the same direction – i.e., they cannot be negatively intercorrelated. The measurement model (outer model) displays the relationships between the constructs and its indicator variables. In the present study, a reflective relationship was assumed for all indicator-construct relationships (not relevant for single indicator constructs). Exogenous latent variables are constructs that explain other constructs – i.e., endogenous latent variables (46). The structural model displays the relationships (paths) between the constructs. Indicator reliability was determined using the outer loadings (possible range: −1–1), and it is generally required that the absolute loadings are >0.708. It is recommended that indicators with absolute loadings ≤0.40 be omitted, and indicators with absolute loadings between 0.40 and 0.708 be omitted only if this result increased internal consistency or convergent validity. Internal consistency reliability was determined using exact (consistent) reliability coefficient (ρ_A; possible range: 0–1), and >0.50 was required. Convergent validity was determined using Average Variance Extracted (AVE; possible range: 0–1), and >0.50 was required. Heterotrait-monotrait ratio (HTMT) was used as an indicator of discriminant validity, and we required values <0.90, preferably <0.85.

To evaluate the structural/inner model, presence of collinearity was checked using variance inflation factor (VIF), and values < 5 were acceptable, preferably <3 (46). Path coefficients (i.e., standardized regression coefficients displaying direct effects) were determined, including mean ± SD, t-values, p-values, and 95% confidence intervals (95% CI). To determine path coefficients and total effects, indirect effects and specific indirect effects, we performed bootstrapping (options: 10,000 samples, complete bootstrapping, percentile boot strap, two-tailed, p = 0.05). Explanatory power was determined from coefficient of determination (R², possible range: 0–1) as well as f² effect size. The latter expresses the change in R² when a specific predecessor construct is omitted from the model (<0.02 = no measurable effect; 0.02–0.14 = small effect; 0.15–0.34 = medium effect; and ≥0.35 large effect). Bootstrapping was performed to obtain t-values, p-values, and 95% CI of these coefficients. In this explorative study, predictive power was not in focus. However, using the PLS_predictive option (10 folds and 10 repetitions), we determined Q²_predictive values. Values > 0 indicated that the model had predictive relevance. The greater value of Q², the greater predictability of the model: < 0.02 = no predictive relevance, 0.02–0.14 = small relevance, 0.15–0.34 = relevance, and ≥0.35 = large predictive relevance (73).

The imputed data are shown in Table 3. Most patients were women, a large majority were born in Europe, and one-fourth had a university education. Patients had on average a relatively high pain intensity and reported high levels of psychological distress and negative effects upon interference, physical and social function, and life control. The prevalence of definite cases for anxiety and depression according to HAD were 38.0% and 33.9%. Hence, this clinical pattern is consistent with non-imputed data for this cohort (12, 53).

Using a PCA in SIMCA-P+, we found two strong multivariate outliers (MCEID: 12543 and 16266), so we excluded them from the subsequent analysis. Thereafter, a PCA of the included variables (R²_cummulative= 0.38, Q²_cummulative= 0.25) was conducted (some MPI variables and all sf36 variables were reversed so that high values indicated a troublesome/negative situation) (Figure 2), which identified two significant components. Important variables for the first component (R²= 0.27, Q²= 0.19) were psychological variables (sf36-mental health, HAD-tot, MPI-Distress, and sf36-role emotional) as well as pain intensity variables (sf36-bodily pain, NRS-7d, and MPI-Pain severity). Also, interference (MPI-Pain interference), physical function aspects (sf36-physical function and sf36-role physical), social aspects (sf36-social function), and control (MPI-Control) had importance for this component. The second component (R²= 0.12, Q²= 0.05) mainly reflected social support (MPI-Social support) but also mental health (sf36-mental health) and control aspects (MPI-Control). Several of the sociodemographic variables showed low importance (loaded near zero) – i.e., pain duration, age, gender, university, outside-Europe, and PRI.

In the next step, the four psychological variables and the three pain intensity variables were regressed (Supplementary Tables S1, S2). As expected, the psychological distress variables and the three pain intensity variables were generally intercorrelated. Pain intensity and psychological distress exhibited different strengths in their relationships to the impacts. Hence, psychological distress aspects were associated with control (MPI-Control), and social function (sf36-social function) in the four regressions and with interference in three of the regressions (MPI-Pain interference). Generally, MPI-Control was a stronger regressor than MPI-Pain interference. In the regressions of the pain intensity variables, pain interference was an important regressor in three regressions, and physical function (sf36-physical function) and social support (MPI-Social support) were important regressors (X-variables) in two regressions.

Modern clinical pain management relies on the biopsychosocial model of pain. The model is usually presented graphically in a relatively simple way (74–76). However, associated explanations of the model suggest significant complexity between the constituents of the three subcomponents. But these relationships are only partially known. PLS-SEM can contribute to the understanding of these relationships. Hence, this real-life study of more than 40,000 patients investigated to what extent and how Pain intensity, Psychological distress, and Social support interact with Interference and Lack of life control aspects.

Although high prevalence of psychological distress was found in the cohort, relatively low correlation and explanatory power existed for the Pain intensity - Psychological distress relationship. Pain intensity had stronger effect on Interference than on Lack of life control, whereas Psychological distress had stronger effect on Lack of life control than on Interference. Depending on the underlying assumption concerning the Pain intensity - Psychological distress relationship, Pain intensity and Psychological distress could complementarily mediate each other in relation to the two impact constructs (Interference and Lack of life control). Social support showed very similar absolute correlations with Interference and Lack of life control. Interference and Lack of life control showed relatively weak associations. Psychological distress level was a moderating factor for several of the significant and important associations/paths.

Chronic pain conditions are associated with high prevalence of anxiety and depressive symptoms (77–80). This does not necessarily mean that high correlations exist between pain intensity and psychological distress levels. The present and other studies have shown that the intercorrelations between pain intensity and psychological distress levels are relatively low (24, 81, 82) – 16% of the variance explained (R²) in this study. As noted above, pain intensity and psychological distress are intercorrelated both in cross sectional and longitudinal studies. Also, day-to-day pain variability studies generally indicate reciprocal associations with negative affect factors (83, 84). Closely related to day-to-day fluctuations are the concepts of trait and state, which are relevant both for pain intensity and psychological distress (85–87). Several models including pain and psychological distress relationships (e.g., the misdirected problem-solving model and the fear avoidance model) suggest mediating factors between these two constructs (7, 88, 89). Cohort heterogeneity for the relationships between pain and mood aspects further complicates the understanding of pain intensity and psychological distress interactions both in short- and long-term perspectives (91, 94). Future studies of mediating (transdiagnostic) factors (e.g., attention, cognition, coping strategies, and emotion regulation) for pain intensity and psychological distress may shed further light on their intricate relationship even though understanding causal directions may be problematic.

As expected in the literature, we observed that both Pain intensity and Psychological distress were positively and significantly associated with the Interference and Lack of life control constructs (24, 81) (Figure 3). Relatively similar explanatory power (R²) was found for Interference and Lack of life control (0.45 vs. 0.48) in the total cohort (Figure 3). Thus, also other factors than Pain intensity, Psychological distress, and Social support contribute to the explanatory power of these constructs.

For both models of the total cohort, Pain intensity had stronger total effects on Interference than on Lack of life control, whereas the reverse was found for Psychological distress vs. Interference and Lack of life control (Table 4A). These results are consistent with results from another study of this cohort using other methods. The study found that pain intensity variables were more important for interference than for depressive symptoms, whereas life control was more strongly associated with symptoms of anxiety and depression (38). In our network study, anxiety and depression levels had stronger relationships than pain intensity with life control aspects (24).

The relative importance of Pain intensity and Psychological distress for the two impact constructs depend on the underlying assumption concerning their interrelationship. Hence, model differences were also obvious. For instance, the total effects of Pain intensity vs. both Interference and Lack of life control were markedly larger in model 2 than in model 1 (Table 4A). Psychological distress had relatively weaker total effects on Interference in model 2 than in model 1 (Table 4A). Differences also existed across the two models when comparing the relative contributions of Pain intensity and Psychological distress for the explanation of Interference and Lack of life control. For example, Pain intensity and Psychological distress in model 1 had very similar total effects on Interference, whereas Pain intensity in model 2 showed a significantly stronger total effect than psychological distress on Interference (0.56 vs. 0.26). In addition, the indirect (mediating) effects were significant and model dependent (Tables 4A,B). Psychological stress and Pain intensity can complementarily mediate each other in relation to Interference and Lack of life control (Tables 4A,B).

The reports of heterogeneity for the causal directions between pain intensity and psychological distress must reasonably have clinical implications (91–94). Thus, the clinical management must take this heterogeneity into account when designing individual treatment interventions.

Interference and Lack of life control showed relatively weak positive correlations, which suggests that they are different constructs of the umbrella construct of life impact. Clinically, they should be assessed separately. Although several large studies from the SQRP, including this one, have focused on factors that intercorrelate with these two separate constructs of life impact, more studies from large real-life cohorts that can identify variables with significant associations with interference and lack of life control are desirable (12, 24, 38). One possible limitation is that we chose indicators from different instruments – e.g., MPI and sf36 – to cover the two constructs. However, the instruments and their subscales are well established in pain research.

Chronic pain increases the risk for social isolation, whereas social support can improve coping, management of treatments, and faster recovery of the pain condition (40, 95). Ferreira-Valente et al. reported negative associations between social support aspects and pain interference (48). A cross-sectional analysis of social isolation, which can be viewed as a negative aspect of social support, showed significant correlations both with pain interference (positive association) and physical function (negative association) (49). As in these studies, we noted that Social support was associated with decreased levels of Lack of life control (i.e., a negative association). Paradoxically, Social support correlated positively with Interference. We have no definite explanation for this counterintuitive finding. It could be due to lack of a mediating factor in our theoretical model. Hence, a recent study investigated social isolation in relation to physical function and identified depression as a mediating factor (49).

The presence or non-presence of severe psychological distress according to HAD was a moderating factor. Thus, several intercorrelations between the constructs/latent variables depend on whether definite signs of anxiety and/or depression are present. Specifically, Pain intensity, Psychological distress, and Social support affect Interference and Lack of life control with different strengths in the two subgroups. For example, Pain intensity and Psychological distress were more strongly intercorrelated in the High distress subgroup than in the Low distress subgroup (Figures 4, 5). Moreover, Psychological distress showed significantly stronger direct effects on Lack of life control but weaker direct effects on Interference in the High distress subgroup compared to the Low distress subgroup. In addition, Interference was associated with lower coefficients of determination (R²) in the High distress subgroup.

Such heterogeneity was also noted for the effects of Social support on Interference and Lack of life control (Figures 4, 5). Hence, the buffering effect of Social support is different and, in some aspects, diminished in the most severe chronic pain conditions, which might be interpreted as social support being of greater value for the ability to be physically active and lead meaningful social lives in chronic patients with a lesser degree of psychological co-morbidity. We also found that the three most important complementary mediating effects were stronger in the High distress subgroup.

Our results taken together indicate cohort heterogeneity and emphasise the need to subgroup patients with complex chronic pain conditions. In the literature there are several studies identifying subgroups [for references see (12)]. There is no agreement concerning the important input variables for such subgroupings; most studies have been hypothesis driven. This study as well as other studies indicate that psychological variables are important (96–101). There is a need for data driven methods in large cohorts of chronic pain patients for identifying the relevant input variables (12, 38). Studies using such methods also identify pain intensity aspects as important for the heterogeneity. Furthermore, studies including from SQRP have found that subgroups show different outcomes of IPRP (12, 51, 53). The different treatment outcomes could partially be due to the presence of moderating effects from definite anxiety and/or depression as well as from other currently unknown variables. Future research needs to focus more on identifying moderating factors. Covering such a knowledge gap may be important for improving outcomes of rehabilitation interventions.

A strength is the large number of chronic pain patients with nation-wide representation. Our results are relevant for patients referred to specialist care. Women are overrepresented in specialist clinics in Sweden compared to the community prevalence (12, 96). In an earlier large SQRP study we could not confirm reports that women assessed at specialist clinics have a more severe clinical situation than men or are judged to be more prone to behavioral change (50, 102). The reasons for this skewed assortment/selection are unclear and need to be addressed. All constructs except Social support had several indicators, which were scales of well-established instruments. This is an advantage from a measurement error point of view compared to path analyses, which only use single items or scales representing one aspect/construct. A limitation is that our theoretical model and subsequent analyses is based on cross-sectional data. For example, it was not possible to determine the most common causal relationship between pain intensity and psychological distress, which is a disadvantage of our models. Our construct Interference covers pain interference and physical function. These aspects are often used interchangeably, but recent reports suggest that it may be necessary to differentiate these measures in, for example, longitudinal perspectives (47, 103).