The study involves a causal mediation analysis of a 2-group participant and assessor-blinded RCT [24,25]. The RESOLVE trial was prospectively registered (ACTRN12615000610538) and approved by the University of New South Wales Human Research Ethics Committee (HC15357).

Participants were recruited from primary care practices and the community in Sydney, Australia. The eligibility criteria are comprehensively described in the trial protocol [25]. Briefly, the RESOLVE trial included people reporting nonspecific LBP [26] (intensity rated at least 3/10), with or without leg pain, that had persisted for at least 12 consecutive weeks. Participants were aged between 18 and 70, fluent in English, able to access the internet, and had a trusted person to assist with the home portion of the intervention. The RESOLVE trial excluded people with LBP due to serious pathology, and people with contraindications to physical activity, transcranial direct current stimulation, cranial electrical simulation, low-intensity laser therapy, or short-wave diathermy. Finally, the RESOLVE trial excluded people who were pregnant or had given birth in the previous 6 months, had undergone spinal surgery in the previous 12 months, were scheduled for major surgery in the next 12 months, or had an uncontrolled mental health condition that would impede participation.

Eligible participants were randomly allocated in a 1:1 ratio to the RESOLVE treatment or the sham-control treatment. The allocation schedule was generated a priori by a scientist independent to the trial using blocked randomization. The allocations were printed and placed in 276 sealed, opaque, sequentially numbered envelopes. Participants and assessors were blind to both group allocation and study hypotheses throughout the trial and follow-up period.

The treatments are comprehensively described in the trial protocol [25] and briefly here. Each treatment group received twelve 30-60-minute one-on-one treatment sessions with a clinician, scheduled approximately weekly over 12-18 weeks. The treatment sessions were supplemented with a home treatment component entailing 30 minutes of training 5 times per week. Concomitant interventions were allowed and recorded on a weekly treatment diary.

The RESOLVE treatment comprised 4 treatment components delivered with a standard progression protocol. The components were pain education, graded sensory training, movement simulation training, and graded precision-focused feedback-enriched functional movement training. The intent was to help people in pain understand that it is safe and helpful to move, feel safe to move, and experience safety with movement as they progress toward reengagement with meaningful functional goals.

Pain education was delivered throughout the treatment period to improve the participants’ understanding of pain and their CLBP problem, address maladaptive beliefs, improve engagement with treatment, and emphasize the value of movement and physical activity. The pain education was based on the Explain Pain model [27,28], delivered according to a standard curriculum, and individualized to the patients lived experience and CLBP narrative. The educational material was delivered by the study clinicians and included the use of graphical media, video, metaphor, and narrative [27,29,30].

Sensory precision training comprised tactile localization training, discrimination of sharp/blunt sensations, and graphesthesia training. The movement simulation component was grounded in graded motor imagery, developed for pathological limb pain [31], and included left–right recognition training using the Recognise software [32] and implicit and explicit motor imagery training using a series of custom-designed videos on low back movements.

Graded precision-focused feedback-enriched functional movement training included individualized movement training related to the patients’ goals. Training progressed from part practice to whole task practice within a visual and proprioceptive feedback-enriched environment [33].

The sham treatment was composed of 3 treatment components to match for time and clinician interaction, individualization, and relevance [34]. These include passive discussion of the participant’s back pain experience, detuned low-intensity laser therapy, and detuned short-wave diathermy. Participants also received a home training program of sham cranial electrical stimulation to control for the home training requirements of the RESOLVE treatment group.

We will test the mechanisms of the RESOLVE treatment strategy for adults with CLBP by estimating the extent to which the 7 hypothesized mediators explain the effect of the treatment on the participants’ pain and disability scores. Using causal mediation analysis, we will partition the total effect (TE) of the treatment into an indirect effect which operates through the mechanism(s) of interest, and a direct effect which operates through all other possible mechanisms [56,57].

The RESOLVE treatment strategy was designed around the Maladaptive Perceptions Model [58], an explanatory framework for the development and persistence of LBP that is grounded in a broad scope of literature concerning behavioral (eg, movement avoidance), neurobiological (eg, altered cortical representations), and cognitive (unhelpful and inaccurate beliefs about the biology of pain and the structural integrity of the back) characteristics of CLBP [20]. The RESOLVE treatment integrates contemporary understandings of pain with known features of best practice care to address the biopsychosocial contributors to the CLBP experience, including maladaptive conceptualizations of the pain problem [59-61], altered sensory function [62,63], altered motor function [64-68], and altered self-perception of the back [43].

While the causal mechanisms that underpin improvement or recovery from CLBP are not well established [13], the Maladaptive Perceptions Model [58] proposes several intermediary variables through which effects might occur. These are cognitions about the back, pain, and movement [69]; back-specific body representations [43]; fidelity and weighting of sensory information from the back [70,71]; and spinal control, movement coordination, and functional tolerance for meaningful activities [33,72]. Components of the RESOLVE treatment were designed to target these factors alongside pain and function.

We have chosen the proposed mediators, pain self-efficacy, back beliefs, pain catastrophizing, kinesiophobia, back perception, tactile acuity, and movement coordination, based on theoretical predictions from the Maladaptive Perceptions Model [58] and the results of pilot studies [33,69-71,73,74].

We will estimate the effect and corresponding uncertainty for the treatment–mediator relationship. This is the average unstandardized effect of the RESOLVE treatment on each independent mediator, compared with sham control. We will also estimate the effect and corresponding uncertainty for the mediator–outcome relationship. This is the average unstandardized effect of the mediator on the outcome (Figure 1).

If inference is considered feasible given the causal assumptions, we will also estimate natural (in)direct effects of the RESOLVE treatment on the outcomes considering (1) the mediators independently, while assuming independence of the mediators, and (2) the mediators simultaneously as a joint mediator.

The identification of natural (in)direct effects relies on several strong and untestable causal assumptions including (1) no treatment–outcome confounding, (2) no mediator–outcome confounding, (3) no treatment–mediator confounding, and (4) no mediator–outcome confounder that is itself affected by the treatment [75]. Adjustment for a sufficient set of observed confounders and correct specification of the statistical models may provide sufficient conditions to identify mediation effects and causal interpretation [76]. Assumption (4) may not hold because there are possible causal relationships between mediators. We will also assess the mediators simultaneously as a joint mediator which relies on weaker assumptions for identification [77,78]. The causal model is presented in Figure 1.

Analyses will be performed in R (version 3.6.1; R Foundation for Statistical Computing) [79]. We will use the “mediation” package [80] to estimate independent mediated effects and the “medflex” package [78] to estimate joint mediated effects.

We will estimate effects for the treatment–mediator and mediator–outcome relationship with 2 regression models: the mediator model and the outcome model. We will specify the mediator model as a linear regression of the mediator (dependent variable) on treatment allocation and baseline values of the mediator. The outcome model will be specified as a linear regression of the outcome at 18 weeks (dependent variable) on the mediator at baseline and follow up, treatment allocation, and possible confounders of the mediator-outcome relationship, and a treatment allocation x mediator interaction term. We will also include an interaction term (treatment allocation × mediator) in the outcome models to increase model flexibility [57].

A model-based inference approach developed by Imai et al [81] will be used to estimate independent mediated effects for each mediator. We will use the “mediate” function [80] to compute the average treatment effect, the average causal mediation effect (ACME), and the average direct effect (ADE). We will use 1000 bootstrapped simulations to estimate 95% CIs. We will interpret conditional estimates of the ACME and ADE separately if there is evidence for a significant (P<.05) intervention–mediator interaction. If there is no evidence for an interaction, we will interpret the average of the conditional effects for the ACME and ADE.

An imputation-based approach using a class of natural effect models introduced by Lange et al [82] and Vansteelandt et al [83] will be used to estimate the joint mediated effect of all mediators simultaneously. We will use the “neModel” function [78] to compute the natural indirect effect (NIE), natural direct effect (NDE), and the TE. We will use 1000 bootstrapped simulations to estimate 95% CIs.

We will assess the proportion and patterns of missing mediator and outcome data. We will conduct all analyses on complete cases if the proportion of missing data is less than 15% for all variables in a given model. If missing data exceed 15%, we will impute missing data through multiple imputation by chain equations using the “mice” package [84] in R.

We will conduct sensitivity analyses to determine the robustness of the estimated ACME to bias introduced by unmeasured pretreatment confounding in the independent mediated effect models [85]. We will use the “medsens” function [80] to estimate the magnitude of residual confounding that would cause the point estimate of the ACME to reach 0. The level of residual confounding is represented by the correlation between the residuals (error terms) in the mediator and outcome models, denoted ρ. By estimating the ACME, including point estimates and 95% CIs, at all possible levels of ρ (between the extremes of –1 and +1), we can determine how strong the effects of residual confounding would need to be to reduce the ACME to 0 (ie, no mediating effect).

We will conduct a sensitivity analysis to determine the robustness of the estimated NIE and NDE to possible unmeasured confounding in the joint mediated effect model [85]. The mediational E-value [86] will be used to assess the minimum strength of the relationship between an unmeasured confounder and the mediator, conditional on measured confounders, that would reduce the NIE and NDE to 0. A comparatively large E-value in relation to known confounders implies that considerable unmeasured confounding would be required to reduce the NIE and NDE to 0. A comparatively small E-value implies that little unmeasured confounding would be required to reduce the NIE and NDE to 0. We will use the “EValue” package [87] in R to estimate the mediational E-value.

If appropriate, we will conduct a sensitivity analysis to assess possible violations to the temporal ordering of the mediator–outcome relationship, excluding participants for whom the mediators and outcomes were assessed concurrently at 18 weeks.

If there is evidence of a mediator–outcome effect, we will investigate the magnitude of change required in the mediator(s) to produce a minimally clinically important difference (MCID) in pain intensity (of 1 point out of 10) [88] and disability (of 2 points out of 24) [89].

The University of New South Wales Human Research Ethics Committee granted ethical approval (HC15357), and all participants provided written informed consent to participate.