Tommaso Costa

Background: Recent anti-amyloid trials have been interpreted as evidence of clinical benefit in Alzheimer's disease, although their relevance remains debated. Objective: To reinterpret pooled randomized evidence on anti-amyloid therapies within a Bayesian framework focused on benefit and harm. Methods: Pooled estimates from a recent Cochrane review were reanalyzed using posterior probability estimation with clinically relevant thresholds. Results: The probability that treatment effects reached clinically meaningful thresholds was generally low, whereas the probability of substantial harm, particularly amyloid-related imaging abnormalities, was high. Conclusions: Current randomized evidence suggests that statistically significant anti-amyloid effects may not translate into clinically meaningful benefit.

Across psychology, bold empirical claims often outpace the evidential support on which they rest. The replication crisis has shown that statistical significance alone provides little guidance about what should rationally be believed. To address this gap between results and rhetoric, we introduce the Bayesian audit—a conceptual and normative framework, rather than a new statistical method, for evaluating whether scientific claims are proportionate to the strength of their evidence. The audit proceeds by identifying the claim, specifying priors, translating the empirical evidence into a likelihood-based measure, updating to obtain posterior belief, testing sensitivity, and synthesizing proportional conclusions. Applied to a well-known case in social psychology—the “elderly priming” study by Bargh et al. (1996)—the audit reveals that the original finding corresponds to only modest Bayesian evidence (Bayes factor ≈ 3). Under reasonable priors (0.05–0.20), the posterior probability that the effect is genuine remains below 0.5, and replication attempts provide limited additional evidential impact at the level of individual studies. The exercise illustrates how strong theoretical language can emerge from weak evidential shifts and how Bayesian reasoning can realign scientific communication with inferential logic. The framework is particularly relevant for psychological science, where replication failures often reflect over-claiming rather than data absence, and where proportional reasoning can help restore coherence between evidence and belief.

Background: Alzheimer's disease (AD) and frontotemporal dementia (FTD) have distinct pathologies but frequently overlapping clinical presentations, making early and atypical differential diagnosis challenging. Blood-based biomarkers offer a minimally invasive alternative to cerebrospinal fluid and neuroimaging measures, yet their diagnostic performance—alone and in combination—remains to be fully established. Objective: To quantify the discriminative ability of plasma biomarkers for differentiating AD, FTD, and healthy controls (HC). Methods: We used a fully Bayesian classification framework, estimating Bayesian logistic regression models for all single, pairwise, and triplet combinations of six plasma biomarkers—phosphorylated tau at threonine 217 (pTau217), brain-derived tau (BD-Tau), neurofilament light chain (NfL), glial fibrillary acidic protein (GFAP), amyloid-β ₄₀ (Aβ ₄₀ ), and amyloid-β ₄₂ (Aβ ₄₂ )—in AD (n = 97), FTD (n = 255), and HC (n = 70). Models were fitted across three contrasts (AD versus HC, FTD versus HC, AD versus FTD) and evaluated via posterior distributions of cross-validated AUC, precision, recall, F1 score, and Brier score. Results: Across 41 candidate models, NfL was the top single biomarker (mean AUC = 0.85), achieving strong discrimination for FTD versus HC (AUC = 0.94). The best two-marker panel (pTau217 + NfL) improved AD versus HC (AUC = 0.96) and AD versus FTD (AUC = 0.90). Adding Aβ ₄₂ produced the highest-ranked triplet (AUC = 0.95) with modest, consistent gains. Posterior coefficients were biologically coherent (AD-specific pTau217 effects; severity-linked NfL), and calibration was satisfactory, with minor overconfidence only at extreme probabilities. Conclusions: A parsimonious pTau217 + NfL panel captures most diagnostic information in the full plasma profile, providing an accurate probabilistic classifier with interpretable uncertainty to support differential diagnosis and clinical triage in precision neurology.

Reports highlight new Alzheimer's disease treatments using anti-amyloid-β immunotherapy, but we see major concerns. The trials supporting lecanemab and donanemab approvals have methodological flaws, and the benefits may be smaller than the minimal clinically important difference—or absent—since patients with poor tolerance were excluded from efficacy analyses. Moreover, treatment increases amyloid-related imaging abnormalities, suggesting local tissue damage, and is linked to brain volume loss. These issues raise doubts about whether regulators are adequately balancing risks and benefits compared to academic critics.

Background: Neuroimaging studies have consistently revealed neuroanatomical abnormalities in individuals with bipolar disorder (BD), major depressive disorder (MDD), and schizophrenia (SZ). However, it remains unknown whether and to what extent disorder-selective gray matter variations occur in these prominent psychiatric disorders. This study conducted a meta-analysis of 25 years of published voxel-based morphometry (VBM) research to assess the presence of selective and robust neuroanatomical substrates of gray matter variation in BD, MDD, and SZ. Methods: Peer-reviewed experiments encompassing subjects with target disorders were systematically searched in the MEDLINE database. Additionally, peer-reviewed data on 30 other psychiatric disorders and 65 neurological diseases were obtained from the BrainMap database. Experiments reporting whole-brain group comparisons between patients and healthy controls were included if they identified significant reductions in gray matter morphometry. Results: The data were analyzed using the Bayes fACtor mOdeliNg algorithm. A total of 1,021 VBM experiments were included, comprising 29,540 patients and 28,177 healthy controls. Primary analyses of psychiatric data revealed strong evidence of gray matter reduction in the right middle temporal gyrus for BD and the posterior dorsal anterior cingulate cortex for SZ ( P ≥ 95% selectivity). The robustness of these findings was confirmed using the fail-safe method tailored to the neuroimaging meta-analytic environment. No selective findings were observed in additional analyses that included neurological diseases. Conclusions: Taken together, these findings offer a framework that underscores the significance of diagnosis-selective neural substrates in psychopathology, a new perspective that could inform distinct pathophysiological processes and assist in diagnosis and treatment.

Few sentences in the philosophy of probability have generated more confusion than Bruno de Finetti’s famous claim that “probability does not exist”. The line is routinely interpreted as a denial of probability itself, while de Finetti intended only to reject the idea that probability is an objective property of the world. A recent essay by David Spiegelhalter in Nature has reintroduced this ambiguity through an instrumentalist “as-if” formulation. This short note clarifies the distinction between de Finetti’s anti-metaphysical stance, Spiegelhalter’s pragmatic instrumentalism, and E. T. Jaynes’s conception of probability as a form of epistemic objectivity. Recognising these differences helps avoid persistent misunderstandings about what ‘non-existence’ means in probabilistic contexts.

Cognitive reserve (CR) refers to the capacity of the brain to sustain cognitive performance despite age-related changes or pathophysiological conditions. Task-based functional magnetic resonance imaging (tb-fMRI) has been instrumental in exploring its neural correlates. CR is traditionally assessed through socio-behavioral proxies such as education, intelligence quotient, or composite indices. However, these proxies vary considerably in definition and application, posing challenges regarding their validity and consistency across experiments. This systematic review and meta-analysis examined whether these proxies are associated with consistent brain activation patterns in healthy adults. A literature search identified 12 eligible tb-fMRI experiments (n = 802 participants) reporting whole-brain CR-related activation. A coordinate-based meta-analysis using permutation of subject images—signed differential mapping (PSI-SDM) was conducted to assess consistent activation patterns across experiments, complemented by meta-regression analyses to examine whether differences in proxy type accounted for inter-study variability. The PSI-SDM meta-analysis yielded no significant clusters of activation. A qualitative synthesis of individual experiments further highlighted a lack of topographical consistency. These findings indicate that widely used socio-behavioral proxies are not associated with detectable, reproducible convergent patterns in tb-fMRI, revealing a measurement gap in current CR research. This null convergence likely reflects methodological heterogeneity and conceptual inconsistency in the literature, rather than the absence of CR-related neural mechanisms.

Background: p-values are ubiquitous in scientific research, yet they fundamentally fail to quantify the strength of evidence for or against competing hypotheses. This limitation is particularly problematic in neuroimaging meta-analyses, where researchers need to assess how strongly the available data support specific and spatially consistent patterns of brain activation across studies. Methods: In this work, we present a practical approach that transforms p-values into their corresponding upper bounds on the Bayes factor, which quantify the maximum plausible evidence in favor of the alternative hypothesis given the observed data. The method is illustrated within the framework of Activation Likelihood Estimation, the most widely used coordinate-based meta-analytic technique in neuroimaging and applied to a reference dataset comprising 73 finger-tapping experiments. Results: The results show that effects traditionally classified as statistically significant using the canonical Activation Likelihood Estimation framework actually span a wide range of evidential strengths, with Bayes factor bounds varying approximately from 46 to 410. This finding reveals substantial heterogeneity in weight of evidence that is concealed by conventional threshold-based inference. Conclusion: By enabling the construction of voxel-wise maps of evidential strength while remaining fully compatible with existing analysis pipelines, the proposed approach helps to avoid common misinterpretations of p-values and improves the interpretability and reliability of neuroimaging meta-analytic conclusions. It therefore provides a conservative, Bayesian-inspired complement to standard significance maps.

The classical conception of falsification presents scientific theories as entities that are decisively refuted when their predictions fail. This picture has long been challenged by both philosophical analysis and scientific practice, yet the relationship between Popper's eliminative view of theory testing and Bayesian model comparison remains insufficiently articulated. This paper develops a unified account in which falsification is reinterpreted as a Bayesian process of model elimination. A theory is not rejected because it contradicts an observation in a logical sense; it is eliminated because it assigns vanishing integrated probability to the data in comparison with an alternative model. This reinterpretation resolves the difficulties raised by the Duhem-Quine thesis, clarifies the status of auxiliary hypotheses, and explains why ad hoc modifications reduce rather than increase theoretical credibility. The analysis is illustrated through two classical episodes in celestial mechanics, the discovery of Neptune and the anomalous precession of Mercury. In the Neptune case, an auxiliary hypothesis internal to Newtonian gravity dramatically increases the marginal likelihood of the theory, preserving it from apparent refutation. In the Mercury case, no permissible auxiliary modification can rescue the Newtonian model, while general relativity assigns high probability to the anomaly without adjustable parameters. The resulting posterior collapse provides a quantitative realisation of Popper's eliminative criterion. Bayesian model comparison therefore supplies the mathematical structure that Popper's philosophy lacked and offers a coherent account of scientific theory change as a process of successive eliminations within a space of competing models.

Background: Observational studies have consistently reported large reductions in COVID-19 risk among vaccinated individuals. However, critics have raised concerns that unmeasured confounding may entirely explain these associations. Methods: We combined the classical Cornfield inequality with a Monte Carlo sensitivity analysis to evaluate whether unmeasured confounding alone could plausibly account for the observed effectiveness of COVID-19 vaccines. The Cornfield inequality provides a lower bound on the strength of confounding required to explain a given association. The Monte Carlo analysis simulates uncertainty over possible confounder–exposure and confounder–outcome relationships by drawing from weakly informative prior distributions, allowing us to estimate the frequency with which such confounding would be sufficient. Results: For an observed risk ratio of 0.08—consistent with early estimates for the Pfizer-BioNTech vaccine—the confounder would need to be both highly imbalanced (e.g., 10 times more prevalent among vaccinated individuals) and strongly protective (e.g., reducing disease risk by 99%). Simulation results showed that, under the specified assumptions, fewer than 2% of draws satisfied this condition. Even in the more moderate case of a risk ratio of 0.25 (e.g., AstraZeneca), the proportion remained below 6%. Conclusions: Our findings suggest that while residual confounding may attenuate effect estimates, it is statistically and epidemiologically implausible that unmeasured confounding alone could fully account for the magnitude of observed vaccine effectiveness. This framework combines the falsificatory logic of Cornfield bounds with the flexibility of simulation-based sensitivity analysis, providing a transparent tool for evaluating confounding-based explanations in observational research.