Signal Detection & Management · Section 7.3
~7 min read · The Drug Safety Coach — Global PV Career Course
Key points
The four most common disproportionality metrics
| Metric | What it measures | Typical signal threshold |
|---|---|---|
| PRR (Proportional Reporting Ratio) | How much more often an event is reported for a drug compared to all other drugs in the database | PRR ≥ 2, with ≥ 3 cases and a chi-squared ≥ 4 (varies by organisation) |
| ROR (Reporting Odds Ratio) | The odds of an event being reported for a drug vs. the odds of it being reported for all other drugs — mathematically related to PRR, generally considered more statistically robust for rare events | Lower bound of the 95% CI ≥ 1 |
| EBGM (Empirical Bayes Geometric Mean) | A Bayesian-shrunk estimate of the reporting ratio, used by the FDA/FAERS — pulls small-count estimates toward the overall average to reduce false signals from rare pairs | EB05 (lower 90% interval bound) ≥ 2, typical threshold |
| IC (Information Component) | A Bayesian metric on a log2 scale, used by WHO-UMC/VigiBase — conceptually similar to EBGM’s shrinkage approach | Lower bound of the IC 95% credibility interval ≥ 0 |
Full text
Disproportionality analysis is the statistical engine behind large-scale signal detection, and all four of its most common metrics — PRR, ROR, EBGM, and IC — are answering a version of the same underlying question: is this specific drug-event pair showing up in the database more often than you’d statistically expect, given how frequently that drug appears overall and how frequently that event appears overall, independent of each other? A pair reported far more often than chance alone would predict is a candidate signal; a pair reported at roughly the rate you’d expect is background noise.
PRR (Proportional Reporting Ratio) and ROR (Reporting Odds Ratio) are the two simpler, frequentist versions of this calculation, and they’re closely mathematically related — ROR is generally considered somewhat more statistically robust, particularly for rare events, which is part of why it’s widely used alongside or instead of PRR. Both, though, share a specific weakness: with a small number of cases, the calculated ratio can be dramatically high purely by chance, because the underlying math doesn’t distinguish between "3 cases, elevated ratio" and "300 cases, the same elevated ratio" in terms of the point estimate alone.
EBGM (Empirical Bayes Geometric Mean), used by the FDA’s FAERS system, and IC (Information Component), used by WHO-UMC’s VigiBase, both address that weakness directly through Bayesian shrinkage: rather than trusting a small-count ratio at face value, these methods statistically pull the estimate toward the overall database average until enough cases accumulate to justify trusting the raw ratio. A drug-event pair with 3 cases and a wildly elevated PRR will show a much more conservative, pulled-toward-average EBGM or IC value — which is exactly the behaviour you want from a metric meant to filter noise from genuine signal.
None of these four metrics, or their associated thresholds, is universally "correct" — different regulators, different databases, and different organisations have settled on different primary metrics and different specific thresholds for what counts as signal-worthy, which is precisely why Lesson 7.9’s comparison of VigiBase, EudraVigilance, and FAERS matters: understanding which metric a given database favours, and why, is part of correctly interpreting a signal that shows up (or doesn’t show up) in one system versus another.
Note
A drug-event pair with only 3 reported cases can still produce a very high PRR or ROR point estimate — the math doesn’t know the difference between 3 cases and 300 in terms of the ratio itself. This is exactly why the confidence interval matters as much as the point estimate, and exactly why EBGM and IC exist: to pull small-count estimates back toward the database average until there’s enough evidence to trust them.
Quick check
Test yourself before moving on — no pressure, just click an answer.
1. Why do EBGM and IC apply Bayesian shrinkage, while PRR and ROR don’t?
2. Why does a disproportionality point estimate need to be interpreted alongside its confidence interval, not on its own?