Objective:

The article explains how causal diagrams clarify exposures, estimands, and adjustment choices in evidence synthesis, and how g-methods can address time-varying confounding when evaluating monitoring strategies.

Approach:

• Causal diagrams: Directed acyclic graphs make assumptions about causal relationships explicit and help identify variables to adjust for or leave unadjusted when estimating a specified effect.
• Define the exposure and estimand: Monitoring assignment, actual receipt of monitoring, and assignment to a monitoring-guided strategy are distinct exposures and causal questions.

Key Findings:

• Nagy et al. pooled 20 estimates from 19 randomised and observational studies, reporting an odds ratio of 0.66 and a prediction interval of 0.18 to 2.40.
• The pooled estimate combines potentially different interventions and estimands, so it cannot establish whether variation reflects settings, interventions, or study designs.
• In PAC-Man, catheter-derived information changed management within two hours in 80% of patients, but the trial found no survival benefit; the study population was a general intensive care population, not patients specifically in shock.
• Marginal structural models use inverse-probability weights for treatment and censoring based on covariate history; their validity depends on assumptions including correct model specification and no unmeasured confounding.

Interpretation:

The article argues that diagrams should be specified before evidence synthesis to clarify the exposure, estimand, confounders, mediators, and censoring process. Diagrams state assumptions rather than verify them, and g-methods do not remove the need for adequate data or their identifying assumptions.

Limitations:

• Causal diagrams depend on substantive assumptions and cannot confirm that those assumptions are correct.
• G-methods require no unmeasured confounding at each time point, positivity, consistency, and correctly specified treatment and censoring models.
• Longitudinal g-methods require individual-level histories of treatment, covariates, and censoring; published summary estimates cannot recreate these data.

Conclusion:

Causal diagrams make the assumptions and causal questions underlying evidence synthesis explicit, while g-methods may address time-varying confounding when their identifying assumptions and data requirements are met.