Mediation analysis in longitudinal intervention studies with an ordinal treatment-dependent confounder

In interventional health studies, causal mediation analysis can be employed to investigate mechanisms through which the intervention affects the targeted health outcome. Identifying direct and indirect (i.e. mediated) effects from empirical data, however, becomes complicated if the mediator-outcome association is confounded by a variable itself affected by the treatment. Here, we investigate identification of mediational effects under such post-treatment confounding in a setting with a longitudinal mediator, time-to-event outcome and a trichotomous ordinal treatment-dependent confounder. We show that if the intervention always affects the treatment-dependent confounder only in one direction (monotonicity), the mediational effects are identified up to a sensitivity parameter and derive their empirical non-parametric expressions. The monotonicity assumption can be assessed from empirical data, based on restrictions on the conditional distribution of the treatment-dependent confounder. We avoid pitfalls related to post-treatment conditioning by treating the mediator as a functional entity and defining the time-to-event outcome as a restricted disease-free time. In an empirical analysis, we use data from the Finnish Diabetes Prevention Study to assess the extent to which the effect of a lifestyle intervention on avoiding type 2 diabetes is mediated through weight reduction in a high-risk population, with other health-related changes acting as treatment-dependent confounders.