Nikola Panajotovikj

// PhD Researcher | AI and Drug Safety

February 1, 2026

Signal detection in pharmacovigilance

Pharmacovigilance databases are full of patterns, but not every pattern is a signal and not every signal is a risk. One of the central statistical problems in post-marketing drug safety is deciding when a drug–event pair is reported more often than expected.12

This is where disproportionality analysis comes in. Methods such as the proportional reporting ratio (PRR), reporting odds ratio (ROR), and information component (IC) are designed to screen large spontaneous reporting databases for unusual reporting patterns.32 They do not prove causality. What they do is help pharmacovigilance teams prioritize which drug–event combinations deserve closer clinical review.1

At a high level, all of these methods ask a similar question: is this adverse drug reaction being reported with this drug more often than we would expect based on background reporting in the database?3 That framing is simple, but it is powerful. It allows thousands or millions of individual case safety reports to be turned into a ranked list of statistical leads.

The basic intuition

Imagine a spontaneous reporting database such as FAERS, EudraVigilance, or VigiBase. For any drug and any adverse event term, you can build a 2×2 table:

  • reports with the drug and the event
  • reports with the drug and other events
  • reports with other drugs and the event
  • reports with other drugs and other events

From that table, disproportionality metrics estimate whether the reporting frequency is unusually high.42

This matters because modern pharmacovigilance does not begin with certainty. It begins with screening. Statistical signal detection helps narrow the search space so expert reviewers can focus on the combinations most likely to reflect emerging safety concerns.15

PRR, ROR, and IC: same goal, different mathematics

The PRR compares the proportion of a specific event among reports for one drug with the proportion of the same event among reports for all other drugs.34 If the event is relatively enriched for that drug, the PRR rises.

The ROR uses the same 2×2 table, but expresses the association as an odds ratio.2 In practice, PRR and ROR often point in similar directions, though their statistical behavior and interpretability differ slightly.

The IC, widely associated with Bayesian signal detection at the Uppsala Monitoring Centre, expresses disproportionality on a logarithmic scale and includes shrinkage toward the null when data are sparse.67 That shrinkage is important, because spontaneous reporting systems contain many rare combinations where naïve ratios can look unstable or exaggerated.

A useful way to think about these methods is that they are not competing philosophies so much as different lenses on the same problem. They all compare observed to expected reporting, but they differ in how they stabilize estimates, quantify uncertainty, and behave when counts are low.28

Why pharmacovigilance needs shrinkage methods

One of the recurring challenges in spontaneous reporting data is noise. Some drug–event pairs have very few reports. Others are affected by publicity, stimulated reporting, channeling, indication effects, masking, or differences in how reporters describe events.19

That is why data mining algorithms with empirical Bayes shrinkage became so influential. A major example is the Multi-item Gamma Poisson Shrinker (MGPS), which estimates disproportionality while pulling unstable low-count estimates back toward the overall background.108

In other words, MGPS is built for the real world of pharmacovigilance: huge databases, sparse cells, and a constant tension between sensitivity and false positives.

MGPS is often summarized through the empirical Bayes geometric mean (EBGM) and its interval estimates.108 Compared with simpler frequentist measures, this can make it especially attractive when screening very large reporting systems where low counts are common and overinterpretation is a real risk.

A signal is not a verdict

Disproportionality methods detect reporting signals, not confirmed adverse drug reactions.1 A high PRR or EBGM does not automatically mean the drug caused the event. It may reflect confounding by indication, co-medication, reporting practices, or database artifacts.9

That is why regulatory guidance treats statistical signal detection as part of a broader workflow rather than a stand-alone answer.15 Signals need medical assessment, case review, biological plausibility, temporality, literature review, and often follow-up using other study designs.

In practice, these methods are best understood as disciplined hypothesis generators.

Why these methods still matter

For all their limitations, disproportionality metrics remain foundational because they solve a real operational problem: how to screen enormous volumes of post-marketing safety data in a reproducible way.37

They are also remarkably adaptable. The same statistical logic can be extended beyond simple drug–event pairs to subgroup analyses, drug–drug interaction screening, indication-specific safety questions, and comparative performance studies between methods.811

For anyone working in drug safety, it is worth understanding both the appeal and the limits of these tools. They are fast, scalable, and often informative — but they are only the opening move in signal evaluation, not the final word.

The real craft of pharmacovigilance lies in knowing how to move from disproportionality to judgment.

  1. European Medicines Agency. Guideline on good pharmacovigilance practices (GVP) Module IX – Signal management (Rev 1). EMA, 2017. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. van Puijenbroek EP, Bate A, Leufkens HGM, Lindquist M, Orre R, Egberts ACG. A comparison of measures of disproportionality for signal detection in spontaneous reporting systems for adverse drug reactions. Pharmacoepidemiology and Drug Safety. 2002;11(1):3–10. ↩ ↩2 ↩3 ↩4 ↩5

  3. European Medicines Agency. Guideline on the use of statistical signal detection methods in the Eudravigilance Data Analysis System. EMA/106464/2006 rev. 1. ↩ ↩2 ↩3 ↩4

  4. Evans SJW, Waller PC, Davis S. Use of proportional reporting ratios (PRRs) for signal generation from spontaneous adverse drug reaction reports. Pharmacoepidemiology and Drug Safety. 2001;10(6):483–486. ↩ ↩2

  5. European Medicines Agency. EudraVigilance User Manual: Marketing Authorisation Holders. EMA, 2021. ↩ ↩2

  6. Bate A, Lindquist M, Edwards IR, et al. A Bayesian neural network method for adverse drug reaction signal generation. European Journal of Clinical Pharmacology. 1998;54(4):315–321. ↩

  7. Uppsala Monitoring Centre. Pharmacovigilance in Perspective. WHO Collaborating Centre for International Drug Monitoring. ↩ ↩2

  8. Almenoff JS, Pattishall EN, Gibbs TG, DuMouchel W, Evans SJW, Yuen N. Comparative performance of two quantitative safety signalling methods: MGPS and PRR. Drug Safety. 2006;29(10):875–887. ↩ ↩2 ↩3 ↩4

  9. European Medicines Agency. GVP Module IX Addendum I: Methodological aspects of signal detection from spontaneous reports of suspected adverse reactions. EMA, 2017. ↩ ↩2

  10. DuMouchel W. Bayesian data mining in large frequency tables, with an application to the FDA spontaneous reporting system. The American Statistician. 1999;53(3):177–190. ↩ ↩2

  11. Truong B, Zhou Y, et al. Screening for clinically relevant drug-drug interactions using disproportionality analyses, logistic regression, and MGPS. Frontiers in Pharmacology. 2023. ↩

Back to blog