Network treatment nodes

Using the framework proposed by Caldwell et al,51 we will use a splitting approach to classify the psychological interventions. A splitting approach was chosen because psychological interventions are typically complex and heterogeneous in nature. For example, two separate trials involving cognitive behavioural therapy may focus on using different psychological principles or strategies and incorporate different additional co-interventions (eg, exercise, passive therapies). Failing to adequately account for the variability, as best as possible, may potentially result in inaccurate estimates of treatment effects. In attempts to account for heterogeneity, we will first scrutinise intervention descriptions to classify the psychological interventions into five treatment nodes based on five key approaches (behavioural, cognitive behavioural, mindfulness-based, counselling-based and pain education). We will also form a separate treatment node using a lumping method to account for combination approaches (eg, two or more psychological approaches delivered together). Then, we will further differentiate whether additional non-psychological co-interventions are involved, which will be subclassified as exercise, passive treatments or physiotherapy. If present, the combination of the psychological approach with a non-psychological co-intervention will form a separate treatment node (eg, cognitive behavioural therapy plus exercise).

The following treatment nodes will be formed for the psychological interventions:

• Behavioural therapy-based interventions (eg, relaxation-based interventions, biofeedback, operant conditioning), which we will consider as psychological approaches focussed on facilitating the removal of positive reinforcement of pain behaviours and promoting health behaviours, in the absence of cognitive strategies;17 18

• Cognitive behavioural therapy-based interventions, which we will consider as the combination of behavioural therapies with an additional focus of changing unhelpful cognitions (ie, thoughts, beliefs and attitudes), and/or promoting emotion regulation and problem-solving;17

• Mindfulness-based interventions, which we will consider as psychological approaches focussed on practicing techniques such as meditation, non-judgemental attention control and awareness (eg, mindfulness-based stress reduction, acceptance and commitment therapy);52 53

• Counselling-based interventions, which we will consider as psychological approaches focussed on using supportive communication and active listening techniques to facilitate healthy behaviour change (eg, health coaching, motivational interviewing);54 55

• Pain education-based interventions, which we will consider as psychological approaches focussed on improving understanding and knowledge about pain. These interventions may involve a biomechanical explanation of LBP, but are clearly focussed on the reconceptualisation of beliefs about the pain experience;56

• Combinations of psychological interventions (eg, pain education combined with behavioural therapy), which we will consider as the delivery of two or more psychological approaches together, in the absence of a non-psychological co-intervention.

Non-psychological co-interventions will be classified into the following treatment nodes:

• Exercise, which we will define as interventions that formally prescribe a structured exercise programme (eg, consisting of aerobic, strengthening, stretching, stabilisation, motor control exercises) and/or direct instructions to increase physical activity levels;

• Passive treatment, including but not limited to spinal manipulative therapy, massage and electrotherapies;

• Physiotherapy, which we will define as interventions delivered by a physiotherapist, which may involve a combination of exercise and passive treatments.

Comparison interventions will be classified into the following treatment nodes:

• Exercise, defined above;

• Passive treatment, defined above;

• Physiotherapy, defined above;

• General practitioner care, which we will define as interventions considered as standard care provided by general practitioners;

• Advice, which we will consider as interventions providing general advice that is not psychologically-informed;

• No intervention (eg, waitlist control, no intervention).

For comparison interventions described as ‘usual care’ by study authors, we will scrutinise the authors’ descriptions of the intervention to classify them into the above treatment nodes.

Figure 1 represents all possible combinations of treatment nodes. Consensus will be sought regarding accurate classification of interventions prior to conducting statistical analyses.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Network plot of all theoretically possible network comparisons.

Prior to data analysis, we will consult clinical experts from the review team to establish the appropriateness of further lumping treatment nodes together if there are inadequate number of studies are available for a given treatment node (eg, less than two studies available). Any post-hoc alternative network geometrics formed using this approach will be clearly identified and justified in the final review. A decision set and supplementary set will be formulated for the final review.

Risk of bias in the included studies

Two reviewers will independently assess risk of bias in the included studies using the Revised Cochrane risk-of-bias tool for randomised trials (RoB 2).57 58 We will use the licensed Microsoft Excel tool to implement the RoB 2. We will pilot-test the risk of bias assessment procedure on a small number of articles. Authors will be contacted for additional information where necessary. The RoB 2 assesses five domains: (1) bias arising from the randomisation process; (2) bias due to deviations from intended interventions; (3) bias due to missing outcome data; (4) bias in measurement of the outcome; and (5) bias in selection of the reported result. Each domain will be graded as low risk of bias, some concerns or high risk of bias, and the results will be summarised in a table. For cluster RCTs, we will use the Cochrane cluster RCT variant of the RoB 2 tool, which assesses an additional domain: bias arising from identification or recruitment of individual participants within clusters.59 An overall risk of bias judgement (low risk of bias, some concerns, or high risk of bias) will be made based on the five (or six) domain-level judgements, as described in Sterne et al.57 Generally, the overall risk of bias judgement corresponds to the worst risk of bias in any of the five (or six) domains, however studies with multiple domains graded as ‘some concerns’ may be judged as high risk of overall bias.57 Disagreements will be resolved through consensus or a third reviewer.

Data analysis

Characteristics of the publications, study designs, study populations, interventions and comparators and outcome measures will be summarised descriptively and presented in a table. Pairwise meta-analysis and NMA will be performed in Stata60 using the metan command (with Knapp-Hartung adjustment applied), and the network package61–63 and network graphs package,64 65 respectively.

Measures of treatment effect

For continuous outcomes that use the same rating scale across all studies, we will use mean differences (MD) and 95% CIs. If different rating scales are used for comparable outcomes, all continuous data for the given outcome will be converted to a common standardised 0 to 100 scale. If data is reported as dichotomous, we will use ORs and 95% CI.

Dealing with missing outcome data and missing statistics

For continuous outcomes, we will impute missing data by converting standard errors, p values or CI into SD.44 If a study only reports the median or IQR, SD will be calculated by dividing the IQR by 1.35, and we will consider the median to be equivalent to the mean. If relevant information is provided in figures, we will extract data from the graphs. If data cannot be obtained, we will attempt to contact authors.

Geometry of the network

The network diagram will be used to graphically depict the available evidence. Nodes will be used to represent the different interventions and comparators, and the weight of the edges will be used to visually represent the proportional number of studies comparing two connected nodes within the network.

Pairwise meta-analysis

We will perform traditional pairwise meta-analyses of all direct comparisons for which there are at least two studies available. We will apply the khartung command to adjust for the Hartung-Knapp-Sidik-Jonkman random-effects method, which has less error rates compared with the DerSimonian and Laird approach in particular across comparisons with greater heterogeneity and when the number of studies is small.66 We will assume the heterogeneity variance for each pairwise comparison is different. We will use the Q statistic to test for statistical heterogeneity in pairwise comparisons. We will use alpha <0.10 as we anticipate a few studies per comparison. We will calculate Higgins I² statistic to indicate the proportion of variability in effect estimates due to heterogeneity and interpret I² >50% as suggesting substantial heterogeneity.44 Forest plots will be created to graphically depict individual and pooled effect sizes. Narrative analysis will be performed if we are unable to impute missing data or cannot contact authors for data, inadequate number of studies are available for a given comparison (eg, <2 studies), or there is substantial heterogeneity.

Assessment of transitivity assumption

Transitivity implies the assumption that distribution of clinical and methodological variables that could potentially act as effect modifiers across available treatment comparisons is balanced within a network.37–39 67 Given the lack of conclusive evidence on treatment effect modifiers for LBP68 or psychological interventions,12 69 70 we will consider the following factors to be potential effect modifiers: age,68 gender,71 sample size,72 baseline physical function, baseline pain intensity, baseline fear avoidance,73 sciatica (leg pain with nerve root compromise). We anticipate that we will have difficulty assessing the distribution of effect modifiers, due to insufficient reporting the potential effect modifiers within individual studies and few studies available per pairwise comparison to make reasonable judgements.74 To assess transitivity, we will use Stata to adjust the weight of the edges within the network plot, proportional to the baseline distribution of the pre-specified effect modifier and visually inspect comparability within the network.67 If minor intransitivity is suspected (ie, minor or negligible dissimilarities in the distribution of a given effect modifier across comparisons based on clinical judgement), we will proceed with the NMA and perform network meta-regressions or subgroup analyses (or both) to explore the influence of suspected factors on the results. If the distribution of a given effect modifier is clearly dissimilar across comparisons, we will exclude network nodes. If intransitivity persists, we will consider not proceeding with NMA.

Network meta-analysis

A NMA will be performed using a frequentist approach to simultaneously compare direct and indirect evidence. We will assume the heterogeneity variance across different comparisons within the NMA model will be the same.75 We will use heterogeneity variances from the NMA model as an index of global network heterogeneity. Mean rank and relative treatment rankings will be estimated for each intervention node according to the surface under the cumulative ranking curve (SUCRA) values.

Assessment of inconsistency

Valid NMA results rely on the assumption of consistency, which describes statistical agreement between direct and indirect evidence for each comparison within a network.39 40 Global inconsistency of the entire network will be assessed using the design-by-treatment interaction model,76 which is a goodness-of-fit test. The presence of inconsistency will be inferred based on p<0.10. Local inconsistencies within closed loops will be assessed with the loop specific approach (Bucher method),77 and by fitting side-splitting models.61 The loop specific approach (Bucher method) will be implemented in Stata using the ifplot command. We will infer the presence of local inconsistencies using a threshold of p<0.10 for either approach. If inconsistencies are identified, we will first check for errors in data extraction. Then, we will examine the potential influence of the pre-specified effect modifiers within inconsistent loops using network meta-regression models or subgroup analyses, and conduct sensitivity analyses excluding studies that may be the source of inconsistency (eg, high risk of bias, studies measuring physical function using the SF-12 or SF-36). If substantial inconsistency remains and the origin remains unexplained, we will consider not proceeding with NMA.

Sensitivity and subgroup analysis

To examine robustness of results, we will conduct a sensitivity analysis by excluding studies with high risk of bias, provided that the original network structure remains the same. We will also perform a sensitivity analysis by excluding studies measuring physical function using the SF-12 or SF-36, which may be a potential source of heterogeneity, provided that sufficient data for physical function is available and the original network structure remains the same. We will also perform network meta-regressions or subgroup analyses on the following covariates, if sufficient data is available: age, gender, sample size, baseline physical function levels, baseline pain levels, baseline fear avoidance, sciatica (leg pain with nerve root compromise). We will assume that for each network meta-regression model, the regression co-efficient for each covariate will be the same across all comparisons in the network. We specify the following assumptions about the direction of effect for each covariate:

• Age (continuous): Increasing magnitudes of the covariate reduces the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

• Gender (continuous): Gender will be summarised as the proportion (percentage) of males. Increasing magnitudes of the covariate reduces the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

• Sample size (continuous): Increasing magnitudes of the covariate reduces the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

• Baseline physical function (continuous): Increasing magnitudes of the covariate increases the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

• Baseline pain intensity (continuous): Increasing magnitudes of the covariate reduces the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

• Baseline fear avoidance (continuous): Increasing magnitudes of the covariate reduces the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

• Sciatica (leg pain with nerve root compromise)(continuous): Presence of sciatica will be summarised as the proportion (percentage) of participants reporting sciatica at baseline. Increasing magnitudes of the covariate reduces the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

Further, subject to the availability of data, we will attempt to perform meta-regressions to explore the effects of intervention parameters relating to dosage and/or frequency (eg, total length (in weeks) of the intervention, total intended hours of the intervention during the intervention period). We make the following assumption about the direction of effect for intervention dosage and/or frequency (continuous): Increasing magnitudes of the covariate increases the differences in effect sizes between the intervention and comparator (compared with trials in which the covariate is less).

We will also perform the following subgroup analyses, provided that sufficient data is available and the original network structure remains the same:

1. Delivery format of psychological intervention (eg, face-to-face, telephone-administered, web-based, self-help booklets), the hypothesis is that face-to-face delivery format will result in greater improvements in disability and pain intensity.

2. Individual versus group-based intervention delivery, the hypothesis is that group-based interventions will result in greater improvements in disability and pain intensity.

Publication bias

Publication bias in the NMA will be evaluated by visual inspection of comparison-adjusted funnel plots for asymmetry. As described above, meta-regression using sample size and effect estimates will be performed to detected small study effect.78

Confidence in cumulative evidence

Judgements of the confidence in cumulative evidence will be evaluated using the Confidence in Network Meta-Analysis (CINeMA) framework,79–81 a web application of the Grading of Recommendations Assessment, Development and Evaluation ratings approach. The framework assesses six domains: within-study bias, across-studies bias, indirectness, imprecision, heterogeneity and incoherence.

Patient and public involvement

Patients will not be involved.