Correcting what cannot be corrected: rethinking publication bias analysis methods in clinical meta-analyses

Methods to assess and adjust for publication bias are often presented as tools to correct distorted evidence in meta-analyses. However, statistical adjustment cannot recover information that was selectively generated, reported, or disseminated. Clinical evidence syntheses frequently rely on small or selective sets of trials and are characterized by substantial heterogeneity, multiple outcomes and time points, and complex dissemination pathways. Publication bias analysis methods are thus prone to over-interpretation and may yield conflicting conclusions. Therefore, they should be understood as an inferential process that links detection, model-based adjustment, and interpretation under explicit and unverifiable assumptions. We review classical methods to detect publication bias, including funnel plots, tests of small-study effects, and P-value-based approaches, and demonstrate their essential role as stress tests of model adequacy rather than as definitive detectors of publication bias. We then examine widely used methods to adjust for publication bias, such as trim-and-fill, selection models, regression-based approaches relating the effect size to study precision, and the Bayesian approach, clarifying their key assumptions and typical failure modes. Using a worked example, we illustrate how applying different publication bias adjustment methods to the same evidence base can yield divergent adjusted effects, emphasizing their assumption dependence. We additionally identify common misuses, propose a framework for evaluations, and discuss emerging challenges related to preprints, umbrella reviews, and AI-assisted evidence synthesis. This review thus aims to help align the strength of clinical conclusions with the robustness or fragility of the underlying data, with direct implications for authors, reviewers, and editors.