Beyond total treatment effects in randomised controlled trials: Baseline measurement of intermediate outcomes needed to reduce confounding in mediation investigations

Abstract

Background:Random allocation avoids confounding bias when estimating the average treatment effect. For continuous outcomes measured at post-treatment as well as prior to randomisation (baseline), analyses based on (A) post-treatment outcome alone, (B) change scores over the treatment phase or (C) conditioning on baseline values (analysis of covariance) provide unbiased estimators of the average treatment effect. The decision to include baseline values of the clinical outcome in the analysis is based on precision arguments, with analysis of covariance known to be most precise. Investigators increasingly carry out explanatory analyses to decompose total treatment effects into components that are mediated by an intermediate continuous outcome and a non-mediated part. Traditional mediation analysis might be performed based on (A) post-treatment values of the intermediate and clinical outcomes alone, (B) respective change scores or (C) conditioning on baseline measures of both intermediate and clinical outcomes.Methods:Using causal diagrams and Monte Carlo simulation, we investigated the performance of the three competing mediation approaches. We considered a data generating model that included three possible confounding processes involving baseline variables: The first two processes modelled baseline measures of the clinical variable or the intermediate variable as common causes of post-treatment measures of these two variables. The third process allowed the two baseline variables themselves to be correlated due to past common causes. We compared the analysis models implied by the competing mediation approaches with this data generating model to hypothesise likely biases in estimators, and tested these in a simulation study. We applied the methods to a randomised trial of pragmatic rehabilitation in patients with chronic fatigue syndrome, which examined the role of limiting activities as a mediator.Results:Estimates of causal mediation effects derived by approach (A) will be biased if one of the three processes involving baseline measures of intermediate or clinical outcomes is operating. Necessary assumptions for the change score approach (B) to provide unbiased estimates under either process include the independence of baseline measures and change scores of the intermediate variable. Finally, estimates provided by the analysis of covariance approach (C) were found to be unbiased under all the three processes considered here. When applied to the example, there was evidence of mediation under all methods but the estimate of the indirect effect depended on the approach used with the proportion mediated varying from 57% to 86%.Conclusion:Trialists planning mediation analyses should measure baseline values of putative mediators as well as of continuous clinical outcomes. An analysis of covariance approach is recommended to avoid potential biases due to confounding processes involving baseline measures of intermediate or clinical outcomes, and not simply for increased precision.

Less Medication Use in Inpatients With Severe Mental Illness Receiving a Multidisciplinary Lifestyle Enhancing Treatment. The MULTI Study III

“…Moreover, by analyzing possible mediation by total activity, we took a first step to gaining more insight into the contribution of different elements of MULTI to the observed improvements. Finally, by including baseline measures in this analysis, our mediation analysis is mathematically in line with a covariance approach, which was the recommended method for mediation analyses (58, 59).…”

Optimising AVATAR therapy for people who hear distressing voices: study protocol for the AVATAR2 multi-centre randomised controlled trial

“…Causal mediation analysis will be based on parametric regression models [ 50 ]. For each mediator separately, this involves estimating a linear model for each mediator with random allocation, baseline outcome, baseline mediator, site and characterisation as covariates, and separately estimating a linear model for each outcome with the mediator, random allocation, baseline outcome, baseline mediator, site and characterisation as covariates.…”

Digitally supported CBT to reduce paranoia and improve reasoning for people with schizophrenia-spectrum psychosis: the SlowMo RCT

“…The analysis used causal mediation analysis based on parametric regression models. 54 This involved estimating a linear model for each mediator with random assignment, baseline outcome, baseline mediator, site and paranoia cut-off point at baseline as covariates, and separately estimating a linear model for each outcome with the mediator, group assignment, baseline outcome, baseline mediator, site and paranoid cut-off point as covariates. The effect of group assignment on the mediator is multiplied by the effect of the mediator on outcome to estimate the indirect effect, and the effect of SlowMo on outcome in the model including mediator is an estimate of the direct effect.…”