Applying causal mediation methods to clinical trial data: What can we learn about why our interventions (don't) work?

Many randomized controlled trials (RCTs) of psychosocial interventions for low back pain (LBP) have been found to have only small effects on disability outcomes. Investigations of the specific mechanisms that may lead to an improvement in outcome have therefore been called for. We present an application of the causal inference approach to mediation analysis using the example of a cluster RCT in a primary care population with (sub)acute LBP randomized to either usual GP care (n = 171) or a minimal psychosocial intervention (n = 143). Mediation analysis explored the causal pathway between treatment allocation and disability at 3 months by considering pain catastrophizing, fear-avoidance beliefs, distress and receiving and following advice as potential mediators, all measured at 6 weeks. We have attempted to explain this approach to mediation analysis in a step-by-step manner to help clinical researchers apply this method more easily. In unadjusted mediation analyses, fear-avoidance beliefs were identified as a mediator of treatment on disability, with an indirect effect of -0.30 (95% CI: -0.86, -0.03), although this relationship was found to be non-significant after adjusting for age, gender and baseline scores. This finding supports the trial authors' hypothesis that while fear-avoidance beliefs are important, this intervention may not have targeted them strongly enough to lead to change. The use of mediation analysis to identify what factors may be part of the causal pathway between intervention and outcome, regardless of whether the intervention was successful, can provide useful information and insight into how to improve future interventions. This study presents a step-by-step approach to mediation analysis using the causal inference framework to investigate why a psychosocial intervention for LBP was unsuccessful. Fear-avoidance beliefs were found to mediate the relationship between treatment and disability, although not when controlling for baseline score