🟠 Moderate Evidence
- When probability gets inverted: the prosecutor’s fallacy in drug safety
- Simpson’s paradox: when stratification reveals hidden confounding
- Artificial intelligence as a safeguard against statistical reasoning error
- Implications for pharmacovigilance and real-world drug monitoring
- Frequently asked questions
Regulators and clinicians routinely misinterpret causality evidence when evaluating drug safety, a pattern rooted in fundamental statistical reasoning errors that can delay life-saving approvals or mask genuine harms. A new analysis published in Pharmaceutical Research (2026) traces how the prosecutor’s fallacy—confusing the probability of evidence given innocence with the probability of innocence given evidence—and Simpson’s paradox shape pharmaceutical decision-making, and explores how artificial intelligence systems might help escape these analytical traps.
Key takeaways
- Prosecutor’s fallacy and Simpson’s paradox distort interpretation of drug safety signals, leading to flawed causality judgments
- Regulators often reverse conditional probabilities, concluding a drug is safe when statistical evidence merely shows safety is plausible
- AI-assisted causal inference models may reduce human cognitive bias in pharmacovigilance, though validation remains incomplete
Study at a Glance
| Source | Pharmaceutical Research |
| Study type | Analytical review with case examples |
| Focus | Statistical reasoning errors in pharmacovigilance |
| Key concept | Prosecutor’s fallacy, Simpson’s paradox, AI applications |
| Relevance | Regulatory decision-making, causality assessment, drug safety |
Two Statistical Traps in Drug Safety Assessment
How prosecutor’s fallacy and Simpson’s paradox lead to opposite safety conclusions from identical datasets
Source: Pharmaceutical Research, 2026 | Georgian Medical Journal News
When probability gets inverted: the prosecutor’s fallacy in drug safety
The prosecutor’s fallacy occurs when regulators conflate P(evidence | drug is safe) with P(drug is safe | evidence observed). According to the analysis in Pharmaceutical Research, this error is endemic in pharmacovigilance reports submitted to agencies like the FDA. A safety signal might have a low probability under a true null hypothesis—meaning the observed adverse events are unlikely if the drug is truly safe—yet regulators incorrectly conclude the drug must be safe because the signal itself is rare. This logical inversion can delay recognition of genuine harms when signal rarity is mistaken for evidence of safety.
The authors cite examples from real pharmacovigilance where a drug with a documented but statistically “unexpected” adverse event profile was approved without adequate subgroup analysis, only later requiring label changes or market restrictions. The fallacy persists partly because conditional probability reasoning demands cognitive effort that competing pressures—time-to-approval, political pressure, manufacturer lobbying—actively discourage.
Simpson’s paradox: when stratification reveals hidden confounding
Simpson’s paradox presents an even subtler trap. A drug may appear safer in the overall population but more harmful within age-stratified or dose-stratified subgroups—or vice versa. The Pharmaceutical Research analysis demonstrates how unequal subgroup sizes and correlated confounders (e.g., older patients receiving higher doses) can flip safety conclusions depending on which stratification is chosen. Regulatory files rarely document systematic subgroup causal analysis before submission, leaving paradoxes buried in appendices until harm emerges post-market.
This is particularly acute in populations with multiple comorbidities—elderly patients, those on polypharmacy regimens, patients in low-income settings with limited access to medical monitoring. A drug labeled “safe” in controlled trials may conceal a dose-dependent or age-stratified harm signal that emerges only after wider deployment. The source article notes that current EMA and FDA pharmacovigilance guidelines do not mandate formal causal discovery checks for Simpson’s paradox, leaving this analytical gap unaddressed.
Artificial intelligence as a safeguard against statistical reasoning error
The authors propose that machine learning systems trained on causal inference frameworks—Bayesian networks, directed acyclic graphs, and counterfactual reasoning—could systematically audit drug safety submissions for logical fallacies before regulatory decisions. Unlike human reviewers, AI systems do not fatigue or succumb to availability bias; they can mechanically decompose complex conditional probability statements and flag reversals or unstated assumptions.
However, the analysis is candid about AI’s current limitations. Causal inference models remain sensitive to unmeasured confounding, and their output is only as trustworthy as the data fed into them. A system trained exclusively on FDA approvals will inherit the FDA’s own analytical biases. The authors recommend AI deployment as a second opinion tool—flagging suspicious reasoning patterns for human expert review—rather than as an autonomous decision-maker. This approach has gained traction in some regulatory pilot programs, including exploratory work at the European Medicines Agency, though large-scale validation studies remain pending.
Implications for pharmacovigilance and real-world drug monitoring
The article underscores a critical infrastructure gap: health policy makers have not mandated formal causal logic auditing in drug approval workflows. Most regulatory submissions rely on narrative arguments and p-values, leaving the conditional probability statements implicit and unchecked. Even when causality assessment tools exist—such as the Naranjo algorithm for individual case reports—they are applied inconsistently and are vulnerable to the same fallacies discussed here when scaled to population-level data.
Statistical reasoning errors—particularly prosecutor’s fallacy and Simpson’s paradox—systematically distort drug safety causality judgments and remain inadequately audited in current regulatory workflows, a gap that AI-assisted causal inference may help address.
— Authors, Pharmaceutical Research (2026)
What this means
Frequently asked questions
What is the prosecutor’s fallacy, and why does it matter for drug approval?
The prosecutor’s fallacy is the logical error of treating P(evidence | innocence) as equivalent to P(innocence | evidence). In drug safety, it means confusing the rarity of a safety signal with proof that a drug is safe. A genuinely harmful drug might produce an unexpected adverse event pattern that seems statistically improbable under a true null hypothesis; regulators then incorrectly conclude the drug is safe because the signal is unusual. This can delay or prevent recognition of real harms.
How does Simpson’s paradox hide drug safety risks?
Simpson’s paradox occurs when a trend (e.g., lower harm rate) in combined data reverses when the data are stratified by subgroups (e.g., by age or dose). A drug might appear safe overall because younger patients receiving low doses have few adverse events, but older patients on higher doses might have substantially elevated harm rates. If regulators only examine aggregate data, the subgroup risk goes undetected. This is especially dangerous in populations with high heterogeneity in age, renal function, or concomitant medications.
Can AI fully replace human judgment in drug safety assessment?
No. According to the analysis, AI is best used as a second opinion tool to flag logical fallacies and hidden confounding. Machine learning causal inference systems cannot detect unmeasured confounding and inherit biases from training data. They should enhance human expert review, not replace it. Regulatory agencies like the EMA are piloting AI-assisted review, but validation studies are still underway.
As pharmaceutical complexity increases and real-world data streams become richer, the stakes of analytical misjudgment grow correspondingly higher. The authors call for a cultural shift in pharmacovigilance: from post-hoc narrative reasoning toward prospectively designed causal analysis, supported by AI quality control but ultimately accountable to transparent, peer-reviewed methods. Regulators, drug manufacturers, and academic researchers all have roles in building this safer epistemic foundation.
Source: Analytic Misjudgment of Drug Safety Evidence and Causality: From the Prosecutor’s Fallacy and Simpson’s Paradox to Artificial Intelligence, Pharmaceutical Research (2026)
Was this article helpful?
Disclaimer. This article is health journalism intended for general information and education. It is not medical advice and is not a substitute for professional diagnosis or treatment. Always consult a qualified healthcare provider about your individual circumstances. Full disclaimer →
Related Coverage
Medically reviewed by Prof. Giorgi Pkhakadze, MD, MPH, PhD. Spotted an error? Contact the editorial team.
