Using Mediation Analysis to Understand How Treatments for Paediatric Pain Work: A Systematic Review and Recommendations for Future Research

Clinicians have an increasing number of evidence-based interventions to treat pain in youth. Mediation analysis offers a way of investigating how interventions work, by examining the extent to which an intermediate variable, or mediator, explains the effect of an intervention. This systematic review examined studies that used mediation analysis to investigate mechanisms of interventions on pain-relevant outcomes for youth (3–18 years) with acute or chronic pain, and provides recommendations for future mediation research in this field. We searched five electronic databases for clinical trials or observational longitudinal studies that included a comparison group and conducted mediation analyses of interventions on youth and assessed pain outcomes. We found six studies (N = 635), which included a total of 53 mediation models examining how interventions affect pain-relevant outcomes for youth. Five studies were secondary analyses of randomized controlled trials of psychological interventions for chronic pain; one was a longitudinal observational study of morphine for acute pain. The pain conditions studied were irritable bowel syndrome, functional abdominal pain, juvenile fibromyalgia, mixed chronic pain, and post-operative pain. Fourteen putative mediators were tested, of which three partially mediated treatment effect; seven did not significantly mediate treatment effect and four had mixed results. Methodological and reporting limitations were common. There are substantial gaps in the field with respect to investigating, and therefore understanding, how paediatric interventions work.


Supplementary
. Christidis et al. (2015) Participants not paediatric (3-18 years) Clementi et al. (2020) Does not compare two or more groups Conti et al. (2020) Participants not paediatric (3-18 years) Corinaldesi et al. (2009) Participants not paediatric (3-18 years) Cunningham et al. (2020) No statistical test for mediation performed Dekker et al. (2016) Study design was irrelevant to this review DiVasta et al. (2015) No statistical test for mediation performed Du et al. (2018) No statistical test for mediation performed Unsure There was a difference in outcome between groups, but it is not clear this was known prior to the analysis (primary data), and there is no mention that the analysis was dependent on a treatment effect being present.

Unsure
It is stated that a treatment effect in the primary RCT was found. It is not stated that mediation analysis was dependent on treatment effect.

Unsure
It is stated that a treatment effect in the primary RCT was found. It is not stated that mediation analysis was dependent on treatment effect. Mediation analysis was pre-planned in the trial protocol with no stipulation of the necessity of a treatment effect to proceed with mediation analysis.

Unsure
It is not explicitly stated that mediation analysis was dependent on treatment effect. However, it is stated that in the primary RCT that a significant treatment effect is present, and an analysis is conducted demonstrating that the putative mediators all had a moderate-large effect size change from baseline to post-treatment compared between groups.

Unsure
It is not explicitly stated that conduct of mediation was dependent on treatment effect, however outcomes (depression, disability) were chosen for mediation analysis because these were significant in the primary RCT, and others (e.g., pain intensity) were not.

Unsure
It is stated that a treatment effect was found. It is not stated that the conduct of the mediation was dependent on a treatment effect being present.

CONDUCT
2.1 Was multiple imputation (or other valid approaches) used to handle missing data? If a complete-case analysis was used, did they adjust for baseline covariates that were differentially distributed between responders and non-responders? Was a sensitivity analysis conducted to assess the impact of different approaches on the findings?
Missing data handled using listwise deletion Yes, a complete-case analysis was conducted. Yes, there was adjustment for baseline covariates using propensity scores.
No sensitivity analysis.
Missing data was assumed missing at random and handled using full information maximum likelihood.
No, a complete-case analysis was not used.
No sensitivity analysis.
Missing data was handled using full information maximum likelihood.
No, a complete-case analysis was not used.
No sensitivity analysis.
Missing data was handled using full information maximum likelihood.
No, a complete-case analysis was not used.
No sensitivity analysis.
No missing data.
Yes, a complete-case analysis was conducted. No adjustment for baseline covariates as no significant differences between groups for measured variables.
n/a -sensitivity analysis.
No imputation was used.
Unclear if a complete-case analysis was conducted. No reporting of adjusting for baseline covariates. Yes, product of coefficient approach justified assuming both models for the outcome and mediator are linear with no interaction.
Yes, product of coefficient approach justified assuming both models for the outcome and mediator are linear with no interaction.
Yes, product of coefficient approach justified assuming both models for the outcome and mediator are linear with no interaction.
Yes, structural equation modelling approach justified assuming both models for the outcome and mediator are linear with no interaction.
Yes, product of coefficient approach justified assuming both models for the outcome and mediator are linear with no interaction.
Yes, product of coefficient approach justified assuming both models for the outcome and mediator are linear with no interaction.
2.4 Does the study assess potential interaction(s) between treatment and confounding factors, treatment and mediator, mediator and mediator in the mediator and outcome models? Does the study evaluate the goodness-of-fit of each model?

No assessment of potential interaction(s)
No goodness-of-fit model.
No assessment of potential interaction(s).
No goodness-of-fit model.
Yes, the study assesses potential interactions between the treatment and mediator.
No goodness-of-fit model.

No assessment of potential interaction(s)
No goodness-of-fit model.

No assessment of potential interaction(s)
No goodness-of-fit model.

No assessment of potential interaction(s)
No goodness-of-fit model.

a Does the study adjust for mediator-mediator and mediatoroutcome confounders?
Mediator-mediator confounders -n/a Yes, adjusts for both exposure-mediator and mediatoroutcome confounders. The study adjusts for baseline covariates by using propensity scores in all analyses.

No adjustment for mediator-mediator confounders
Yes, adjusted for mediator-outcome confounders by adjusting for reciprocal mediators in the multiple mediator model.
No adjustment for mediator-mediator or mediator-outcome confounders.

No adjustment for mediator-mediator confounders
Yes, adjusted for mediator-outcome confounders by adjusting for reciprocal mediators in the multiple mediator model.
No adjustment for mediator-mediator or mediator-outcome confounders.
No adjustment for mediator-mediator or mediator-outcome confounders.
2.6 Does the study perform sensitivity analysis to assess sensitivity of the results to (1) the assumption of no measured mediator-mediator or mediatoroutcome confounders, (2) potential measurement errors of the mediators?

No sensitivity analyses
No sensitivity analyses

No sensitivity analyses
No sensitivity analyses

No sensitivity analyses
No sensitivity analyses 2.7 Does the study use apt strategies when some of the mediator-mediator or mediatoroutcome confounders are potentially affected by the treatment (e.g., by considering confounders as mediators themselves)?

REPORTING
3.1 Does the study report the approaches used for mediation and provide a causal diagram that underlies the analysis?
Yes, the approach for mediation is described, and a causal diagram is provided.
Yes, the approach for mediation is described, and a causal diagram is provided.
Yes, the approach for mediation is described, and a causal diagram is provided.
Yes, the approach for mediation is described, and a causal diagram is provided.
Yes, the approach for mediation is described.
No, a causal diagram is not provided.
Yes, the approach for mediation is described.
No, a causal diagram is not provided.
3.2 Does the study report the sample size calculation, the actual sample size of the mediation analysis and how the missing data is handled?
No sample size calculation provided.
Yes, actual sample size is described.
Yes, description of how missing data were handled.
No sample size calculation provided.
Yes, actual sample size is described.
Yes, description of how missing data were handled.
No sample size calculation provided.
Yes, actual sample size is described.
Yes, description of how missing data were handled.
No sample size calculation provided.
Yes, actual sample size is described.
Yes, description of how missing data were handled.
No sample size calculation provided.
Yes, described actual sample size.
n/a -no missing data.
No sample size calculation provided.
Yes, described actual sample size.
No description of how missing data were handled. 3.3 Does the study report all confounders considered and adjusted for in the analysis?
Yes, all baseline covariates considered and adjusted for via propensity scores are reported.
No confounders were considered or adjusted for.
No confounders were considered or adjusted for.
Yes, baseline values of the mediator and outcome variables that were considered and adjusted for in the analysis as covariates are reported.
No confounders were considered or adjusted for.
No confounders were considered or adjusted for.
3.4 Does the study report the model building procedure and the final form of all models used in the analysis? Do they report the goodness-of-fit of these models?
No, model building procedures and final form of all models in the analysis are not described.
No goodness-of-fit metric is provided.
No, model building procedures and final form of all models in the analysis are not described.
No goodness-of-fit metric is provided.
No, model building procedures and final form of all models in the analysis are not described.
No goodness-of-fit metric is provided.
No, model building procedures and final form of all models in the analysis are not described.
No goodness-of-fit metric is provided.
No, model building procedures and final form of all models in the analysis are not described.
No goodness-of-fit metric is provided.
No, model building procedures and final form of all models in the analysis are not described.
No goodness-of-fit metric is provided. 3.5 Does the study report the point estimates and the confidence intervals (CIs) of the different direct, indirect and total treatment effects?
Reported: point estimates of the direct and total effect. 3.6 Does the study report the methods and results of all sensitivity and other additional analyses (in the main paper or appendices)?
N/a -no sensitivity or other analyses are conducted.
Yes, the study reports methods and results of additional analyses (timelagged analysis). n/a for sensitivity analysis as none conducted.
N/a -no sensitivity or other analyses are conducted.
N/a -no sensitivity or other analyses are conducted.
N/a -no sensitivity or other analyses are conducted.
Yes, the study reports methods and results of additional analyses.
n/a for sensitivity analysis as none conducted.
3.7 Does the study discuss the validity of all causal assumptions underlying the analysis (in the main paper or appendices)?
No -Does not discuss assumption of temporal ordering.
Yes -Acknowledges the plausibility of bias due to unmeasured confounders. But, does not justify all measured confounders are adequate to fully adjust for confounding. Also does not follow with a sensitivity analysis to test the plausibility of the assumption.
Yes -Acknowledges the requirement for temporal ordering and conducts a timelagged analyses to justify the validity of the assumption of temporal ordering No -Does not acknowledge or justify the plausibility of the assumption of no unmeasured confounding. Does not follow with a sensitivity analysis to test the plausibility of the assumption.
Yes -Acknowledges the limitation that temporal order between mediator and outcome was not established. Does not conduct analyses to justify the validity of the assumption of temporal ordering.
Yes -Acknowledges the limitation that confounders were not measured or assessed. Does not conduct a sensitivity analysis to test the plausibility of the assumption of no unmeasured confounders.
Yes -Acknowledges the requirement for temporal ordering and conducts a timelagged mediation model to justify the validity of the assumption of temporal ordering No -Does not acknowledge or justify the plausibility of the assumption of no unmeasured confounding. Does not follow with a sensitivity analysis to test the plausibility of the assumption.
Yes -Acknowledges the requirement for temporal ordering.
Does not conduct analyses to justify the validity of the assumption of temporal ordering.
No -Does not acknowledge or justify the plausibility of the assumption of no unmeasured confounding. Does not follow with a sensitivity analysis to test the plausibility of the assumption.
Yes -Acknowledges the requirement for temporal ordering and conducts hierarchical regression analyses to justify the validity of the assumption of temporal ordering.
No -Does not acknowledge or justify the plausibility of the assumption of no unmeasured confounding. Does not follow with a sensitivity analysis to test the plausibility of the assumption.