1. Introduction

Pain is common in childhood and adolescence ^[1][2], presenting after an injury or procedure, as a consequence of disease, or without any identifiable cause ^[3]. Effective interventions for acute and chronic pain in young people are critically needed. Many interventions encompass a variety of single and multi-modal treatments including pharmacological, psychological, and physical interventions. The efficacy of such treatments for children and adolescents with acute pain and chronic has been reviewed (e.g., ^[4][5][6][7]). Some seem effective, but effect sizes are small. One potential explanation is that small effect sizes may represent wide variability in patient response, which could be addressed by better understanding the underlying mechanisms of treatment effects. For example, acceptance and commitment therapy (ACT) is thought to operate via the proposed mechanism of changing psychological flexibility ^[8], and cognitive behavioural therapy (CBT) is thought to operate by changing maladaptive cognitions (e.g., pain catastrophizing) ^[9]. Understanding treatment mechanisms would allow targets for paediatric pain interventions to be informed by empirical evidence, rather than presumptive theories. Clinicians would therefore be able to refine and optimize the effectiveness of interventions by selectively targeting mechanisms known to improve outcomes ^[10].

Mediation analysis is the most frequently used quantitative method for evaluating the mechanisms of interventions ^[11]. Mediation analyses answer questions about how or why an intervention works, or does not work, by estimating the extent to which interventions exert their effects on outcomes via mediating variables (i.e., ‘mediators’). For example, in a recent mediation analysis, Kendall, et al. ^[12] showed that cognitive-behavioural therapy (termed the ‘exposure’) reduced anxiety symptoms (termed the ‘outcome’) in youth through improvements in coping efficacy but not through reductions in anxious self-talk. Mediation analyses targeting how or why an intervention works require longitudinal data because the timing of effects between the treatment, mediator and outcome, needs to be established. These analyses also require comparison groups for the intervention because a causal contrast against the treatment cannot be evaluated without one ^[13]. For the purpose of making causal inferences, mediation analyses are best conducted on randomized designs that eliminate potential confounding of the intervention-mediator effects, but non-randomized designs are acceptable if potential confounders are controlled for.

Mediation analyses that examine potential mediators of interventions for paediatric pain exist, but a systematic evaluation of the field is lacking. Systematic reviews of mediation analyses have examined potential mediators of interventions for adults with musculoskeletal pain ^[14] and back pain ^[15], but whether we can generalize these findings to children and adolescents with pain, who are developmentally unique and the interventions are modified accordingly, remains unknown. Some mechanisms are common to paediatric acute and chronic pain conditions; others are not ^[16]. Little is known if the same principle applies to treatment.

2. Mechanisms of treatments for paediatric pain

The evidence in this review provides more guidance about mechanisms that do not explain why interventions for youth with pain work, rather than mechanisms that do. That is, 41 of the 53 models did not reveal mediators of treatment effect. While these findings would need to be replicated in a larger number of methodologically-sound studies to increase validity, they provide preliminarily evidence that treatments for youth with pain may not operate via current theoretical explanations. For example, Kashikar-Zuck, et al. ^{[17][18][19][20][21][22][23][24]}[33] found that neither pain catastrophizing, coping efficacy or pain coping mediated improvement in functioning or depressive symptoms following CBT. This indicates that CBT works through other undiscovered mechanisms and identifying such mechanisms would increase the ability to adapt CBT to improve its effectiveness in this population. However, identifying variables that do not mediate treatment effects does not necessarily mean that theoretical explanations for how treatments work are unsupported. Two studies in this review ^[18][20] assessed mediation models including variables that did not align with hypothesised theories of how interventions work, and other variables that did. For example, Wicksell, et al. ^[18] assessed six potential mediators for the effect of ACT on depression and pain interference. By demonstrated mediation effects in the variables that aligned with the hypothesised theory of ACT (i.e., pain impairment belief, pain reactivity), but not in the variables that did not (i.e., self-efficacy, kinesiophobia, pain catastrophizing, pain intensity), finding from this study support theoretical explanations for how ACT works.

This review revealed mediators that may be promising targets of interventions for youth with pain. Two studies in this review assessed the outcome of GI-symptoms for children ^[17] or adolescents ^[20] with GI-related pain, revealing that GI- specific avoidance behaviour may be an important target of exposure-based internet CBT interventions. However, in other included studies, mediators were tested across different populations, types of pain, for different treatments and outcomes, making comparison difficult, and highlighting the limited state of the evidence. While mediation analyses in paediatric pain are common, research with designs able to identify mediators of interventions are rare. Longitudinal data are often used to conduct mediation analysis to investigate mechanisms of the development of symptoms (e.g. ^[22][23]), rather than mechanisms of interventions. When mediation analyses are used to investigate interventions, some studies use designs that limit the ability to conduct meaningful mediation analyses, such as longitudinal designs without comparison groups (e.g. ^[24]). Ideally, mediation analyses would be embedded into clinical trials, as a valuable tool to allow us to move beyond efficacy and investigate how intervention for youth with pain do or do not work. More studies of this kind, for each type of intervention with consistent mediators and outcomes are needed to draw firm conclusions about specific mechanisms of interventions for youth with pain.

2.1. Methodological Quality of Included Studies

A key methodological strength of all studies in this review was the establishment of temporality, in that all mediators were measured prior to outcomes. Additionally, three studies conducted sensitivity analyses to test the assumption of temporal precedence. Two studies ^[17][20] included models that assessed mediators at repeated time points—an approach that allows for investigating a treatment with a gradual change in the mediator and outcome ^[25]. However, methodological limitations were common. No studies estimated a priori sample sizes for indirect or direct effects, which may reflect the complexities and lack of tools available for these calculations ^[26][27]. Despite their importance, less than half (21/53) of the mediation models in this review controlled for mediator-outcome confounders. Moreover, controlling for some confounders does not exclude the possibility of bias due to other, unmeasured confounders. Sensitivity analyses are recommended to assess the influence of unmeasured confounders ^[28], but they were not implemented in any study in this review.

2.2. Future Recommendations

In light of the methodological issues raised in this review, and the limited number of included studies, recommendations for future research are provided.

2.2.1. Study Effective and Ineffective Interventions

All included studies used mediation analysis to investigate mechanisms of effective interventions. However, there is also value in studying failed mechanisms of ineffective interventions. Mediation analysis of unsuccessful interventions can help identify where the hypothesized causal mechanisms break down. The results of such studies can then be used to reproduce or refine interventions. That is, if we identify mediators that have a causal effect on key outcomes (i.e., path b), but the intervention did not affect the mediator (i.e., path a), we may modify the intervention to specifically target the mediator. For example, an RCT ^[29] revealed that a pain science education program was no more effective at reducing functional disability than usual care for adults with chronic back pain, and a secondary mediation analysis was conducted to answer why the intervention did not work ^[30]. The mediation analysis found that illness perceptions (intermediate variable) were significantly associated with functional disability (outcome), but the intervention did not significantly influence illness perceptions. The authors conclude that illness perception may be an important target in future treatments. That no studies in this review investigated mechanisms of failed interventions is likely influenced by historical perspectives on mediation whereby a statistically significant treatment effect was required to proceed with mediation analysis. However, recent research has shown this requirement to be unnecessary, as indirect effects can be present in the absence of direct effects ^[31]. Therefore, mediation analysis of ineffective treatments is encouraged ^[32][33] and could benefit the paediatric pain field.

2.2.2. Assess Shared Mechanisms Across Interventions

Studying the same mediators across different interventions can help to identify if different treatments operate through shared mechanisms. Also, if shared mechanisms that affect multiple outcomes can be identified, then we can tailor interventions to specifically target those mechanisms ^[34]. For example, mechanistic research into treatments for adults with back pain suggest that different psychological interventions produce similar effects on outcomes through a common set of mediators (e.g., increased self-efficacy, reduced pain catastrophizing)—the same may be true of interventions for youth with pain. In this review, it is difficult to draw conclusions on shared mechanisms as only two mediators (pain catastrophizing and GI-specific avoidance behaviour) were assessed across multiple studies. Of these, one study; while both studies identified GI-specific avoidance behaviour as a mediator of treatment effect. More mediation analyses in the field will provide better insight into the role of shared mechanisms.

2.3. Strengths and Limitations of This Review

There are several strengths to this review, including preregistration (now recommended practice in the pain field ^[35]), a comprehensive search strategy based on established search terms, inclusion of a wide population of youth (aged 3 to 18) and use of two reviewers to independently screen and evaluate studies and extract data. Finally, a key strength of this review is the focus on formal mediation analyses with appropriate study designs (i.e., longitudinal designs with a comparison group), a choice that is reflected in our search terms.

This review also has limitations. Study heterogeneity meant we were unable to compare the magnitude of mediation effects across studies. Publication bias is also a possibility, but the limited number of included studies precluded formal evaluation of it. All but one study ^[21] in this review reported evidence in favour of one or more of the mediating pathways that were investigated. It is possible that post hoc mediation analyses were only conducted in studies that showed a significant treatment effect, and that primarily those with statistically significant results were published. It is also possible that our search strategy biased towards selecting studies that reported mediation analysis in the title or abstract, which may be more frequent when mediation analyses are significant. Finally, while we reported which studies conducted sample size calculations (none), we did not assess and exclude studies that were inadequately powered for mediation analysis.