Model-Based Sensitivity Analysis of Mediation in Tertiary Prevention: The PACE Trial of Interventions for Chronic Fatigue Syndrome

Introduction: We can study how interventions work using mediation analysis.  Addressing mediation hypotheses can lead to a greater understanding of intervention mechanisms, and the potential to refine interventions. Mediation analyses often focus on variables measured only at a single time point, but such cross-sectional analysis does not respect the implied temporal ordering of a mediation model. Tertiary prevention trials generally provide multiple measures of variables. Repeated measures of mediators and outcomes allow for the application of different types of longitudinal mediation models, such as autoregressive, latent growth and latent change models.  These models can provide flexibility in modelling trajectories, incorporate measurement error and allow for some types of unmeasured confounding. Using different types of models to address mediation hypotheses could be thought of as a type of sensitivity analysis.

Methods:  Longitudinal mediation of the effects of the cognitive behaviour therapy (CBT) and graded exercise therapy (GET) interventions evaluated in the Pacing, Graded Activity, and Cognitive Behavior Therapy: A Randomized Evaluation (PACE) trial of rehabilitative interventions for chronic fatigue syndrome (ISRCTN 54285094) was studied. Fear avoidance was used as an example mediator, and physical functioning as an example outcome. Three types of longitudinal mediation structural equation models were fitted to the PACE data: a simplex model (an autoregressive model allowing for measurement error), a latent growth model, and a latent change model. The models incorporated correlated measurement errors to allow for concurrent unmeasured confounding between the mediator and outcome. Information criteria and informal comparisons of the Root Mean Square Error of Approximation between models were used. The parameter estimates and their 95% CI were plotted and informally compared.

Results: All three model types indicated that fear avoidance mediated the effects of CBT and GET on physical functioning. The most appropriate model by fit statistics was a more flexible variant of the latent change model.  This model provided for accumulation of the mediated effect early after randomization.  This reflected the trend on average in the PACE data, and is likely a common pattern in clinical trials of complex interventions.

Conclusions: In the absence of strong theory or prior evidence, and when longitudinal data are available, robustness of results can be increased by the application of different classes of mediation models. Here, the different models led to the same conclusion, which was that fear avoidance was a mediator of the effect of tertiary prevention interventions for chronic fatigue syndrome. Latent change models provided the best fit for the PACE data, possibly because these models allow for non-uniform change between different time points. Tertiary prevention trials need to consider mediation hypotheses at the design stage and ensure that sufficient measures are taken in order to apply the models described.