Types of studies

We included only RCTs.

Types of participants

We included RCTs involving the following types of participants: male or female; aged 18 years or older; treated for LBP; in the acute, subacute or chronic phases, with or without sciatica. We excluded studies involving people with LBP due to specific causes (e.g. tumour, metastasis, fracture, inflammation, osteoporosis, rheumatoid arthritis).

Types of interventions

We included RCTs using any type of traction, such as mechanical traction, manual traction (unspecific or segmental traction), computerized traction, auto‐traction, underwater traction, bed rest traction, inverted traction, continuous traction and intermittent traction. Additional treatment was allowed, provided that traction was the main contrast between the intervention and control groups. We included studies with any type of control group (i.e. those that used placebo, sham, no treatment or other treatments).

Types of outcome measures

The four primary outcome measures that we considered to be the most important were pain intensity (e.g. measured by a visual analogue scale (VAS) or a numerical rating scale (NRS)), back‐pain‐specific functional status (e.g. measured by the Roland Morris Disability Questionnaire or Oswestry Disability Index (ODI)), a global measure of improvement (e.g. overall improvement, proportion of participants recovered, subjective improvement of symptoms) and return to work (e.g. measured by return to work status or days off work). We also considered reported adverse effects.

These outcomes could be measured immediately after the end of one traction session, immediately after a course of traction sessions, in the short‐term after the end of the traction sessions (up to three months), or in the long‐term (around one year).

Electronic searches

We used the results of the literature search listed in Appendix 1, updating the three previous versions of this review (Clarke 2006a; Clarke 2006b; Van der Heijden 1995a). This included a computer‐aided search the Cochrane Back Review Group Specialized Register (August 2012), the Cochrane Central Register of Controlled Trials (2012 Issue 8), MEDLINE (January 2006 to August 2012), EMBASE (January 2006 to August 2012) and CINAHL (January 2006 to August 2012).

Searching other resources

Furthermore, we screened reference lists of relevant reviews and identified RCTs, as well as references in personal files of the review authors.

Selection of studies

Two review authors independently selected the trials to be included in the systematic review using title, abstract and keywords. The same two review authors independently applied the selection criteria to the studies that were retrieved by our literature search. We used consensus to resolve disagreements concerning selection and inclusion of RCTs. There was the option to consult a third review author if disagreement had persisted, although this was not necessary. We only evaluated full papers and excluded papers written in languages other than English, Dutch, German, French and Swedish.

Data extraction and management

Two review authors (IW and ISW) independently extracted the data (using a standardized form) considering the study population (e.g. number of participants, age, gender, type and duration of back pain), the interventions (type, intensity, and frequency of index and reference interventions) and the primary outcomes (type and duration of follow‐up). We used consensus to resolve disagreements and we would have consulted a third review author (GH) if disagreement persisted, although this was not necessary. We summarized key findings in a narrative format. We did not blind data extraction.

Assessment of risk of bias in included studies

We used the Cochrane Back Review Group's 'Risk of bias' tool to assess the risk of bias of the included RCTs (Furlan 2009). The 12 criteria are listed in Appendix 2. Studies included in the previous version of the review had not been assessed using this tool. Therefore, we re‐assessed these studies according to the updated methods. We could not obtain two articles (Lind 1974; Reust 1988) and two articles were written in a language that the review authors did not master (Bihaug 1978; Walker 1982). We transformed the previous risk of bias assessments of these four trials to the new format without re‐assessing them. As a result, supporting statements for the risk of bias assessments are missing for these studies. Two review authors (IW and ISW) independently assessed the methodological quality. Review authors resolved their initial discrepancies during discussion; the presented results are based on their full consensus. We did not blind quality assessment with regard to the authors, institution and journal. We did not contact study authors for additional information, because half the trials were published in the late 1990s. If the article did not contain the required information for the scoring of a specific item, we scored the item as 'unclear'.

We scored the criteria as 'low risk', 'high risk' or 'unclear risk', and reported them in the 'Risk of bias' table. We defined a study with a low risk of bias as one fulfilling six or more of the criteria and having no fatal flaws. In the previous review, a sensitivity analysis was performed in which six was considered the cut‐off point for low risk of bias. A second sensitivity analysis was performed in which half of items that had been scored 'unclear' in each trial were included as 'positive'. The same cut‐off point of six for low risk of bias is supported by empirical evidence (Van Tulder 2009).

Blinding of participants and care providers to treatment allocation is nearly impossible in trials of traction therapy. Given that some of the primary outcomes assessed in this review are subjective measures (i.e. pain and functional status), any attempt to blind the outcome assessor regarding these outcomes can be considered irrelevant. However, most studies also assessed objective outcome measures. If the care provider assessing those outcomes was blinded, the item was scored as 'low risk'.

Measures of treatment effect

We analyzed dichotomous outcomes by calculating the risk difference. We analyzed continuous outcomes by calculating the mean difference (MD) when the same instrument was used to measure outcomes, or the standardized mean difference (SMD) when different instruments were used to measure the outcomes. We converted VAS or NRS scales to a 100‐point scale. We expressed uncertainty using with 95% confidence intervals (CI).

We grouped outcomes by timing when they were measured: immediately after, short term and long term.

Unit of analysis issues

In several studies, we compared more than two intervention groups. We included these studies by making pair‐wise comparisons between all possible pairs of intervention groups with traction being one of the intervention groups. The same group of participants was included more than once in these examples (e.g. underwater traction versus underwater massage and underwater traction versus balneotherapy in the study performed by Konrad 1992). These participants were not counted twice in the meta‐analysis.

Dealing with missing data

In cases where data were reported as a median with an interquartile range (IQR), we assumed that the median was equivalent to the mean and the width of the IQR equivalent to 1.35 times the standard deviation in accordance with Cochrane Handbook for Systematic Reviews of Interventions, section 7.7.3.5 (Higgins 2011). If standard deviations were not given, we calculated them from the 95% CIs, P values based on a two‐sided t‐test or standard errors. We did not include data reported in graphs in this review.

Assessment of heterogeneity

We tested heterogeneity using the Chi² test and I² statistic; however, the decision regarding heterogeneity was dependent upon the I² statistic (Higgins 2011). We defined substantial heterogeneity as an I² greater than 50%, and where necessary, the effect of the interventions were synthesised narratively when the I² statistic was greater than 50%.

Assessment of reporting biases

We searched ClinicalTrials.org and ISRCTN.org for the protocols of included studies. When protocols were available, we checked studies for selective outcome reporting.

Data synthesis

A quantitative analysis had been planned, but most of the studies did not provide sufficient data to enable statistical pooling (e.g. some trials reported the mean score but not the standard deviation, other trials reported median and IQR; some trials reported only post‐intervention means and other trials reported mean change scores; some trials did not report any numerical data. Therefore, we used a descriptive analysis to summarize the data. In this analysis, we used a rating system of levels of evidence to summarize the results of the studies in terms of the strength of the scientific evidence. To accomplish this, we used the GRADE approach, as recommended in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011), and adapted in the updated Cochrane Back Review Group method guidelines (Furlan 2009). The system consists of five levels of evidence, based on performance against five principal domains or factors:

• high‐quality evidence ‐ consistent findings among at least 75% of RCTs with low risk of bias, consistent, direct and precise data and no known or suspected publication biases. Further research is unlikely to change either the estimate or our confidence of the results;

• moderate‐quality evidence ‐ one of the domains is not met. Further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate;

• low‐quality evidence ‐ two of the domains are not met. Further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate;

• very‐low‐quality evidence ‐ three of the domains are not met. We are very uncertain about the results;

• no evidence ‐ no RCTs were identified that addressed this outcome.

Factors that may decrease the quality of the evidence are: study design and risk of bias (downgraded when > 25% of the participants were from studies with a high risk of bias), inconsistency of results, indirectness (downgraded when > 50% of the participants were outside the target group), imprecision (downgraded when the total number of participants was less than 400 for continuous outcomes and 300 for dichotomous outcomes) and other factors (e.g. reporting bias).

Because the majority of studies contained a mix of participants with acute, subacute and chronic LBP, we did not separate out these groups in our analyses, other than in several trials involving only people with chronic LBP. We categorized studies as including people 'with sciatica' if more than66% of the participants were described as having sciatica (this may or may not have included those with nerve root symptoms) or if there was a separate analysis of outcomes in those with sciatica.

Subgroup analysis and investigation of heterogeneity

Predefined subgroup analyses included:

• different types of comparison (traction versus placebo, sham or no treatment; physiotherapy with traction versus physiotherapy without traction; different types of traction and traction versus other treatments);

• different symptom patterns in subjects (mixed population of people with LBP with and without sciatica; people with LBP with sciatica and people with LBP without sciatica).

However, we were not able to conduct these analyses, because of reasons stated above. Instead, the results were synthesized narratively. 'Summary of findings' tables were generated for all analyses of different types of comparison. Primary outcome measures at a follow‐up duration of 12 to 16 weeks were included in the 'Summary of findings' tables.

Sensitivity analysis

In the previous review, sensitivity analyses were carried out to determine the cut‐off for high‐quality studies. The cut‐off point was set at six criteria for risk of bias, which is supported by empirical evidence (Van Tulder 2009). We considered that studies that met six or more of the criteria for risk of bias carried low risk of bias, whereas studies that met fewer than six of the criteria carried high risk of bias. We did not plan or carry out any new sensitivity analyses.