Journal of Rehabilitation Medicine 51-1CompleteIssue | Page 11

8 J. Oosterhaven et al. disability with dropout compared with other studies. The patients who dropped out in this study had lower levels of pain intensity and reported less disability than the completers (28). However, with regard to the unusually high dropout rate of 51% compared with 10–23% in the other studies (24, 30–33), the results of this study may be related to unique elements in this study (poor selection for treatment or inexperienced therapists). Therefore, we may conclude that higher levels of pain intensity and higher severity of disability are predictors for dropout that are worthwhile for future research (24, 29, 33). Since all studies in this review investigated different combinations of potential predictors, it might not be surprising that conflicting results were demonstra- ted, for instance, for age, depression, pain intensity, disability and other predictors. Another explanation may be the heterogeneity in study populations due to differences in study design and differences in type of chronic pain treatment (28, 31–33)). The 3 types of interventions identified in this review differed in main programme goal (return to work vs pain management), duration of the programme (4–20 weeks) and selec- tion criteria for the study population. Depending on in which country the study was performed (USA or in European coun­tries), differences in the organiza- tion of the healthcare system and referral and funding patterns might have also caused differences in study populations. Study limitations The results of this systematic review have to be in- terpreted in the light of some limitations. First of all, the finding that the methodological quality was con- sidered to be low in all 8 studies (24, 27–33). Quality assessment was a difficult process caused by the lack of consistent and clear reporting of study- and dropout characteristics. Due to a serious risk of bias in most studies for: study attrition and statistical reporting and analysis, the internal validity of the evidence found in this systematic review can be considered to be low. This was probably caused by the retrospective design of most cohort studies (24, 27–33). Because the lack of conceptualization and operatio- nalization of dropout, the methodological quality of the domain of study attrition of the QUIPS was difficult to judge. Only 4 of the 8 studies provided a definition of dropout (24, 29, 32, 33). This may have influenced the internal validity of this review. The huge differences in dropout rates (10–51%) found in this systematic review can partly be explained by the variation in definitions that were used to describe dropout. A high risk of bias was found for the statistical analysis and reporting www.medicaljournals.se/jrm for 4 out of 8 studies in this review. Judgement of the statistical results in univariate analyses was most often based on the description in the methods sections of the original articles, as the results were not reported. The statistical analyses in these studies followed a more data-driven approach, which often tend to give a too optimistic estimate of performance of the prediction model and may cause overfitting of the data (36). Only 3 studies reported statistical analyses based on a conceptual framework (28, 29, 31). The results in this review may suggest that dropout is entirely a patient characteristic (see predictors in the sociodemographic, patient and disease domains). How­ ever, as demonstrated in the mental health literature, the experience level of therapists and the therapeutic alliance may be important moderators in association with dropout (10, 12, 13). In this literature it has been recommended that it may be more relevant to focus on the interaction between patient-related predictors and therapist/therapy-related predictors and not on a single key predictor alone (10, 12, 13). Judging the available evidence in this review, strong limitations concerning the external validity may be raised. A factor that will influence the generalizability of the multiple logistic regression models for dropout presented in this review is related to the differences in chronic pain treatments in the studies. The evidence found for predictors in the 3 different interventions first has to be validated outside the context in which it was gathered (19). For example, most predictors associated with dropout were found in a large prospective cohort study within a functional restoration programme in the USA. These findings may not be generalized to other contexts (24). Another factor that may limit the generalizability of the predictors found in the RCT (33) is the fact that RCTs are known to attract very highly motivated patients, which does not reflect patients with low treatment motivation, who are often seen in daily clinical practice. It is remarkable that none of the multiple logistic re­ gression models that were identified in this review were tested in independent samples. Currently, there is little information available on the performance of the mul- tiple logistic regression models. Three parameters can be described to gain more insight in the performance of the logistic regression models: calibration, discrimina- tion and clinical usefulness (37, 38). Only calibration was described with the goodness-of-fit for 2 multiple logistic regression models in 2 studies (28, 30). Due to the abovementioned limitations, the decision was made not to perform a best-evidence synthesis (25, 26), but to systematically generate a broad overview of all po- tential predictors found in the literature associated with dropout in chronic pain management programmes. The