Signal detection is how pharmacovigilance teams spot a possible new safety issue hidden inside large volumes of case data. The most common statistical approach is disproportionality analysis: looking for drug-event pairs that are reported more often than you would expect by chance. Four methods dominate the field — PRR, ROR, BCPNN/IC, and EBGM. This guide explains each without the heavy maths.
The foundation: the 2x2 contingency table
Every disproportionality method starts from the same idea. For a given drug and a given event, you build a 2x2 table counting four things: cases with both the drug and the event (a), the drug but not the event (b), the event but not the drug (c), and neither (d). The methods differ in how they turn those counts into a signal score.
PRR — Proportional Reporting Ratio
PRR compares the proportion of a specific event among reports for the drug of interest against that proportion among all other drugs. A PRR above 1 means the event is reported disproportionately often for that drug. A common signal threshold is PRR of at least 2, with at least 3 cases and a supporting chi-square value. PRR is simple, transparent, and widely used by regulators such as the MHRA — but it can be unstable when case counts are very low.
ROR — Reporting Odds Ratio
ROR is closely related to PRR but uses odds rather than proportions (a*d / b*c). It behaves similarly to PRR in most situations and is favoured in some academic and EudraVigilance-style analyses because it fits neatly with logistic-regression approaches that adjust for confounders. Like PRR, it is a frequentist method and shares the same sensitivity to small numbers.
BCPNN / IC — Bayesian Confidence Propagation Neural Network
BCPNN, used by the WHO's Uppsala Monitoring Centre in VigiBase, produces the Information Component (IC). It is a Bayesian method, which means it 'shrinks' estimates toward neutral when data are sparse. The practical benefit is stability: BCPNN is far less likely to throw a false signal off one or two reports than PRR or ROR. The IC and its credibility interval are reported; an IC lower bound above 0 is the usual flag.
EBGM — Empirical Bayes Geometric Mean
EBGM comes from the Multi-item Gamma Poisson Shrinker (MGPS) method associated with the FDA. Like BCPNN it is Bayesian and shrinks unstable estimates, but it also handles higher-order associations (drug-drug-event interactions). The EB05 — the lower bound of the EBGM confidence interval — is a popular conservative signal threshold, often flagged when EB05 is at or above 2.
Side-by-side comparison
| Method | Type | Statistic | Strength | Limitation |
|---|---|---|---|---|
| PRR | Frequentist | Ratio of proportions | Simple, transparent, regulator-friendly | Unstable at low counts |
| ROR | Frequentist | Reporting odds ratio | Fits regression / confounder adjustment | Unstable at low counts |
| BCPNN | Bayesian | Information Component (IC) | Stable with sparse data | Less intuitive to explain |
| EBGM | Bayesian | EBGM / EB05 | Stable; handles interactions | More computationally heavy |
Which method should you use?
There is no single 'best' method — most mature programmes run more than one and look for agreement. Frequentist methods (PRR, ROR) are easy to communicate and good first filters. Bayesian methods (BCPNN, EBGM) are more robust when data are sparse, which is common for newer products or rare events. The right choice depends on your data volume, regulatory expectations, and how conservative you need to be.
A note on case volume
Disproportionality statistics are only meaningful where case volume supports them. For small datasets, a flagged 'signal' may be noise — which is why qualitative review and clinical judgement always sit alongside the numbers.
Whichever methods you run, the statistic is only the starting point. A detected signal must be triaged, validated, and assessed by qualified personnel before any action. Platforms like PVgenix Signal Detection provide PRR, ROR, BCPNN/IC and EBGM with configurable thresholds, masking detection, and a managed signal lifecycle so the statistics feed a controlled, auditable review process rather than standing alone.
Frequently asked questions
It is a statistical approach that looks for drug-event pairs reported more often than expected by chance, using a 2x2 contingency table. PRR, ROR, BCPNN, and EBGM are the common methods.
PRR is a simple frequentist ratio that can be unstable when case counts are low. EBGM is a Bayesian method that shrinks unstable estimates toward neutral, making it more robust for sparse data, and it can handle higher-order interactions.
There is no single best method. Many programmes run both frequentist (PRR, ROR) and Bayesian (BCPNN, EBGM) methods and look for agreement, choosing thresholds based on data volume and regulatory expectations.