Table 3 Examination of Identifiability Assumptions for Causal Mediation Analysis
References | No unmeasured exposure-outcome confounders | No unmeasured mediator-outcome confounders | No unmeasured exposure-mediator confounders | No mediator-outcome confounder affected by the exposure | ||||
|---|---|---|---|---|---|---|---|---|
Acknowledged assumption | Empirical analyses or sensitivity analyses | Acknowledged assumption | Empirical analyses or sensitivity analyses | Acknowledged assumption | Empirical analyses or sensitivity analyses | Acknowledged assumption | Empirical analyses or sensitivity analyses | |
Studies estimating controlled direct effects only | ||||||||
Banack et al. [26] | ✓ | Not reported | ✓ | Unmeasured confounder cardiorespiratory-fitness Estimates of the direct effect of cardiorespiratory fitness on mortality from well-established literature. No literature on estimates of prevalence differences of unmeasured confounder—so a range of 10–90 % was considered | Not applicable | |||
Mendola et al. [33] | ✓ | Not reported | ✓ | Unmeasured confounder maternal infection Estimates of the direct effect of maternal infection on neonatal outcome ranged from 2 to 10. Prevalence differences of unmeasured confounder—so a range of 1–99 % was considered. Whether this was done because no literature was available on which to base the sensitivity analyses was not reported | Not applicable | |||
Messerlian et al. [34] | ✓ | It is unclear if they were addressing this concern although additional pre-specified stratum- specific with different reference categories and exposure groups were used for sensitivity analyses | ✓ | Stratified analyses “triangulated” those derived from marginal structural models. It is unclear if they were addressing this concern | Not applicable | |||
Rao et al. [36] | ✓ | Unmeasured confounder situation that unmeasured confounders could be correlated with exposure, mediator, and outcome were considered. Using parameters, such as γ (conditional increase in risk for oral cancer), P1 (prevalence in smokers/chewers/drinkers), and P2 (prevalence among non-smokers/non-chewers/non-drinkers) were specified. The bias introduced by unmeasured confounders that may entirely invalidate the controlled direct effect was calculated | ✓ | Unmeasured confounder considered with the exposure-outcome relationship | Not applicable | |||
Studies estimating natural direct and indirect effects | ||||||||
D’Amelio et al. [27] | Randomized controlled trial-not applicable | ✓a | Not reported | Randomized controlled trial-not applicable | ✓a | No sensitivity analyses, but adjusted for biomarkers that were unbalanced between the two treatment groups at baseline | ||
Freeman et al. [28] | Randomized controlled trial-not applicable | ✓ | No sensitivity analyses, but adjusted for baseline confounders; can’t rule out | Randomized controlled trial-not applicable | ✓ | Not reported | ||
Jackson et al. [29] | ✓ | Showed risk factors by antipsychotic group | ✓ | No sensitivity analyses, but adjusted for many risk factors; cannot rule out residual confounding | ✓ | No sensitivity analysis, but residual confounding (i.e. delirium) at baseline that could bias the total and indirect effects upwards was acknowledged | ✓ | No sensitivity analyses, but conducted stratified analyses by mediators to provide qualitative evidence for whether or not the association between mediator and mortality is modified by antipsychotic type |
Louwies et al. [31] | X | No sensitivity analyses, but adjusted for confounders in Table 1, except day of the week | X | Not reported | X | Not reported | X | Not reported |
Lu et al. [32] | ✓ | Excluded first 3 years of follow-up to reduce the influence of baseline confounders Restricted the analysis to never-smokers to better control for confounding by smoking | ✓ | Unmeasured confounder Common cause of metabolic mediators and coronary heart disease (e.g. family history, genetic factors, residual confounding due to measurement error in diet and physical activity). Sensitivity analyses done with two scenarios: (1) mild confounding (increased hazard ratio by factor of 1.1 and prevalence 20 % for normal weight/25 % for overweight/obese); and (2) strong confounding (increased hazard ratio by factor of 1.8 and prevalence of 45 % for normal weight and 40 % for overweight/obese) | ✓ | Restricted the analysis to never-smokers to better control for confounding by smoking | ✓ | Not reported |
Raghavan et al. [35] | X | Not reported | X | No sensitivity analyses, but mediation analysis was conducted with all three metabolic mediators (CIR, HOMA-IR and MSS) together | X | No sensitivity analyses, but mediation analysis was conducted with all three metabolic mediators (CIR, HOMA-IR and MSS) together | X | Not reported |
Song et al. [37] | ✓ | No sensitivity analysis, but included all the covariates that may confound the relationship | ✓ | No sensitivity analysis, but included all the covariates that may confound the relationship | ✓ | No sensitivity analysis, but included all the covariates that may confound the relationship | ✓ | Sensitivity analysis was conducted through excluding BMI, a mediator-outcome confounder that is possibly affected by the exposure (low birth weight) |
Xie et al. [38] | X | Not reported | X | Not reported | X | Not reported | X | Not reported |
Effects not identified | ||||||||
Kositsawat et al. [30] | X | Not reported | X | Not reported | X | Not reported | X | Not reported |