Measurement Error and Confounding Considerations in Longitudinal Mediation Models of Tertiary Interventions for Chronic Fatigue Syndrome: The PACE Trial

Introduction:  Studying tertiary prevention of illness using randomised clinical trials requires large monetary and time commitments. Such studies should therefore not simply answer questions about the effectiveness of interventions, but also about how interventions work.  Intervention mechanisms can be studied using mediation analysis, which allows testing of theoretical intervention models and refinement of interventions.  Mediation studies often use only single contemporaneous measures of the mediator and outcome limiting the conclusions that can be drawn.  Longitudinally measured mediators and outcomes, such as those in the Pacing, Graded Activity, and Cognitive Behaviour Therapy: A Randomised Evaluation trial of rehabilitative interventions for chronic fatigue syndrome (PACE, ISRCTN 54285094) allow for more realistic estimates of mediated effects.

Methods:  Longitudinal autoregressive structural equation models (SEM) accounting for measurement error and  unmeasured confounding (through correlated measurement errors) were used to study intervention effect mediation of cognitive behaviour therapy (CBT) and graded exercise therapy (GET) in PACE. Fear avoidance and physical function were used as example mediator and outcome; these were measured at baseline and three times post-randomisation as part of the trial design.  In using all the measured data, baseline and post-randomization measures of mediator and outcome were incorporated, which could be among the most important measured confounders of the mediator – outcome relationship. Model fit criteria, Wald tests and comparisons of parameter estimates were used.

Results:  Longitudinal SEM incorporating several measures of mediators and outcomes were more flexible than the single mediator/single outcome model, and gave what were likely more plausible estimates of mediated effects.  These models showed that taking account of measurement error and confounding was important. Constancy of mediator – outcome effects over time and across intervention groups increased precision.  For CBT and GET, 46% and 53% of the overall effect were mediated through fear avoidance. 

Conclusions:  More realistic longitudinal models of mediation showed that approximately half of the effect of each of CBT and GET on physical function was mediated through reducing avoidance of fearful situations.  Tertiary prevention trials should routinely address intervention mechanisms using mediation analysis.  Such trials should be designed to include multiple measurements of mediators and outcomes so that more realistic mediation models can be used.  Such models allow for measurement error and some type of unmeasured confounding, which is important.