Drug safety surveillance increasingly combines spontaneous reports, longitudinal healthcare data, distributed data networks, and computational methods, yet the inferential structure connecting these resources remains under-specified. A recurring problem is that treatment is not allocated independently of prognosis. Indication, disease severity, prior treatment history, channeling toward newer or alternative therapies, switching, treatment-confounder feedback, and informative discontinuation can each reshape the exposed population and the observed risk set. Treating these mechanisms as a single generic category of “confounding” obscures when they arise and which design or analytical response they require. This Current Opinion argues that safety surveillance should explicitly represent treatment allocation and treatment-state evolution before interpreting an exposure–event association. Baseline indication should be distinguished from severity within an indication; channeling should be distinguished from ordinary incident treatment initiation; switching should be treated as both an estimand and exposure-definition problem; and post-baseline confounders affected by earlier treatment should not be handled as ordinary baseline covariates. We further argue that discontinuation and censoring require separate consideration because the reasons patients leave treatment or observation can themselves carry prognostic information. The article develops a proposed bias-diagnostic approach in which causal structure precedes statistical adjustment, source triangulation does not substitute for identification, and computational assistance remains bounded by clinical and epidemiological interpretation. The objective is not to produce a universal correction algorithm, but to make the assumptions connecting treatment selection, treatment history, exposure state, event occurrence, observation, and regulatory interpretation explicit.
IJPRAS
