The most important finding of this longitudinal study is that being employed and the level of disability before treatment are predictive factors for relevant improvement in CLBP patients’ functional status at 1-year follow-up. In contrast to our expectation, the pre-treatment degree of experienced pain intensity and belief in one’s ability to manage and to cope with CLBP complaints appeared not to be predictive of outcome. Moreover, the results revealed that 1 year after the programme, highly distressed patients who were referred to the programme were not at risk of being a failure.
Previously, this CPP programme has been evaluated for patients who met the inclusion criteria [
20,
21]. The present analysis was conducted to determine whether it would be possible to enhance the efficacy of the programme by further patient selection by identifying a subgroup of patients who could benefit of the programme. As the main goal of the intervention is to improve disability, success at 1-year follow-up was defined as having reached 22 points or lower on the ODI. We reasoned that less change is not clinically relevant.
A minimal clinical important difference (MCID) of ten points on the ODI has been recommended as a measure for clinical relevancy in CLBP [
33]. Although consensus has been reached for this MCID value, the value is still arbitrary because some of the studies upon which the consensus is based contain heterogeneous CLBP population samples and were derived from primary care [
33]. It is difficult to measure what is clinically relevant to patients [
34]. Patients who are highly disabled at pre-treatment assessment and who did reach the MCID value after treatment could be classified as improved success whilst in fact they are still disabled. Therefore, we decided to use ODI values seen in ‘normal’ healthy populations as a measure of success. The current study results show that at 1-year follow-up 217 patients (41.4 %) reached this ODI value. With the exception of one study, which was performed in primary care [
35] and included CLBP patients who were still at work and who were less disabled (ODI 20 [range 2–52]), we are not aware of any studies performed in secondary or tertiary care investigating factors predicting a functional outcome related to ‘normal’ and healthy populations.
Prediction model: pre-treatment ‘employed’ and pre-treatment ‘disability’
Being employed appeared to be the most important predictive factor (OR 3.61 [95 % CI 1.80–7.26]; dichotomised). To a lesser extent, the level of pre-treatment disability predicts the outcome (OR 0.94 [95 % CI 0.92–0.97]; decrease per point on ODI). These findings are consistent with the results of the systematic review by van der Hulst et al. [
6]. We recommend screening CLBP patients for these factors. It is known that CLBP patients who are significantly disabled and who are absent from work pre-treatment have a poor outcome [
36,
37]. The ODI might have screening potential as it has been shown to be of predictive value for chronicity [
37]. Patients who are moderately disabled and who are at least partially employed before treatment could be given a higher priority for entry into a CPP programme. From an organisational and economic perspective, patients who are at work and who are mildly disabled might benefit from a shortened programme. To substantiate these ideas, more research is needed.
In the current study, the prediction model (Table
3) has wider confidence intervals for the validation model (Table
4), and a lower explained variance (
R
2 22 % versus 40 % [Hosmer and Lemeshow]), resulting in a greater number of cases correctly classified (67 versus 75 %) for the validation model. Because of these discrepancies and to estimate the stability of the prediction model, we performed a post hoc multivariate logistic regression analysis on the random sample (
n = 252) using a bootstrap procedure that is 500 repeated samples with replacement. All potential prediction variables were then entered in one block. This result is comparable to the final prediction model (Model
χ
2 [5] = 68,157
p < 0.001;
R
2 24 % [Hosmer and Lemeshow]; 23 % [Cox and Snell]; 31 % [Nagelkerke]; 70 % correct classified). Based on these results, we conclude that the final prediction model, as initially developed, is robust. This model explains 22 % (Hosmer and Lemeshow) of the total variance. Moreover, 67 % of the patients were correctly classified. Although inconsistent evidence does exist for predictive factors that were identified for outcome of interventions with a physical and cognitive behavioural approach, a comparable and typical low amount of explained variance has been found [
38‐
41]; as well as the percentage correctly classified patients [
42]. Because physical and psychosocial factors only marginally contribute to treatment success, other non-specific or moderating factors such as clear treatment rationale, a highly structured programme, providing a pressure-cooker model programme, the dose of treatment, skilful staff, and the patient’s readiness to change pain-related behaviour have been proposed as being predictive for a successful outcome [
11,
43,
44]. There are two increasingly suggested specific contributing factors to functional treatment outcome in chronic musculoskeletal pain: expectancy of treatment outcome [
45] and central sensitisation [
46‐
48]. Central sensitisation includes features of referred pain, hypersensitivity to peripheral stimuli and neuropathic pain which are felt to represent peripheral manifestations of augmented central pain sensations. However, further research is required to determine which specific factors contribute to a successful outcome for CLBP patients in a CPP programme.
Some inconsistent qualitative evidence has been reported which is related to other potential and a priori predictive factors that might be expected for this study: experienced pain intensity [
6,
49], gender [
7,
23], or self-efficacy [
35,
50,
51]. However, no support for these predictive factors could be found in the present study. It has also been suggested that improvement of dysfunctional cognitive behavioural factors such as catastrophizing cognitions and fear of movement behaviour might contribute to a successful outcome [
11,
52]. This suggestion is endorsed by the fear avoidance model which postulates a causal relationship between pain catastrophizing, fear of movement, disability and experienced pain severity [
4]. Some studies have concluded that the impact of these dysfunctional cognitive behavioural factors on outcome measures as pain as well as functional status is diminished [
15,
53] or is even absent [
6], which is consistent with the results of the present study.
Studies investigating the predictive value of psychological distress have only yielded inconclusive and tentative evidence [
6,
15]. Self-rated depressive mood has been reported to be of prognostic value [
8,
18,
22,
39,
54]; furthermore, it has been suggested that patients with reported symptoms would benefit less from a multidisciplinary programme compared to patients with no or only mildly depressive symptoms [
7,
18,
23]. In the current study, despite a small association between the level of distress and being successful at 1-year follow-up (Pearson’s
r −0.23,
p < 0.001), no predictive value of psychological distress could be found in the final prediction model. This means that CLBP patients who are distressed at pre-treatment assessment might benefit from a CPP programme.
Strengths and limitations
The strengths of this study are the large sample size (
n = 524) and the wide range of available pre-treatment data. This means that there was enough statistical power to study the contribution of the different potential predictive factors towards successful treatment outcome over time. Although data were missing on at least one assessment for 67 (13 %) patients, no pre-treatment differences between non-responders and responders were seen. Our main results are based on the MI technique. MI is a technique that depends on model-based imputation of multiple values for each missing observation instead of only one estimate as in single imputation techniques. The major advantage of this method, over single imputation techniques or ‘complete cases only’, is that it does not underestimate variability. Single imputation methods could result in the estimated standard errors being too small, whereas multiple imputation results in the correct magnitude for estimated standard errors and confidence intervals [
55,
56], i.e. these imputed values reflect the uncertainty in estimation caused by the missing values [
56]. Thus, the information contained within the missing data seems similar in nature to the information actually documented. This implies that the conclusions based on the results obtained with MI are robust. Moreover, the large study sample gave us the opportunity to develop a prediction model in a 50 % random sample of the original set and to validate and check this final model with the remaining data.
Limitations in this study include possible selection bias. Therefore, generalisation to common clinical practice is limited as our findings are theoretically relevant only to specialised back care. There are no data for those patients not selected (28 %), it is possible that other factors could be predictive for a successful treatment outcome. It is possible that these patients were not ready or motivated to change pain-related behaviour. Although a selection criterion for treatment, we neither assessed this factor in a valid and reproducible way at pre-treatment nor assessed it systematically over time. Further research is needed to assess this factor and to evaluate its contribution to the outcome.