Disproportionality analysis has been the backbone of spontaneous reporting surveillance for more than thirty years. The Reporting Odds Ratio appeared in the pharmacovigilance literature in the mid-1990s as a computationally tractable adaptation of the case-control odds ratio, and the Proportional Reporting Ratio followed shortly after as the UK's Yellow Card scheme needed a method that didn't require exact ICSR counts in the comparison group. Both methods worked well within the constraints they were designed for: relatively clean case populations, drugs used in reasonably well-defined therapeutic indications, and a regulatory environment where quarterly signal detection cycles were the norm.
The question worth examining in 2025 is not whether those methods have value — they clearly do — but whether the assumptions embedded in them still hold for the pharmacovigilance problems that matter most right now.
What Disproportionality Analysis Actually Measures
It's worth being precise about what ROR and PRR compute before discussing their limits. The ROR for drug D and adverse event E is the odds of seeing event E in reports mentioning drug D, divided by the odds of seeing event E in reports not mentioning drug D. The PRR is the proportion of reports for drug D that contain event E, divided by the proportion of all other drug reports that contain event E. Both are measures of reporting disproportionality — they tell you whether event E is reported at a higher-than-expected rate among reports that mention drug D.
Two things this is not: it is not an incidence rate, and it is not a causal estimate. The ICSR denominator is "all reports containing some mention of drug D," not "all patients who took drug D." That distinction matters enormously for drugs with high voluntary reporting rates, for drugs that have been under regulatory scrutiny (generating stimulated reporting), and for drugs used in populations who are already experiencing high adverse event burdens from other causes.
The BCPNN (Bayesian Confidence Propagation Neural Network), developed at Uppsala Monitoring Centre and deployed in WHO's VigiBase, and MGPS (Multi-item Gamma Poisson Shrinker), used extensively within FDA's AERS predecessor system, addressed some of the small-count instability problems in PRR and ROR by applying shrinkage estimation. They are methodologically more sophisticated but rest on the same fundamental assumption: that drug-event pairs can be evaluated as independent units.
The Independence Assumption in a Polypharmacy World
The independence assumption — that the reporting behavior for drug A is meaningfully separable from the co-medications drug A is typically used with — was defensible when most patients took one or two drugs for clearly distinct indications. For a large fraction of today's post-market drug population, that assumption is materially violated.
Consider a drug used primarily in second-line treatment of a metastatic solid tumor. The patients who take this drug are almost universally on multiple prior and concurrent therapies. Their FAERS reports are not reports of Drug X taken alone — they are reports of Drug X taken alongside antiemetics, supportive care agents, sometimes prior lines of chemotherapy still clearing from the system, and possibly investigational co-medications. When you compute ROR for Drug X against a specific adverse event using the standard 2×2 contingency table, the "reports not mentioning Drug X" group contains a very different polypharmacy mix than the "reports mentioning Drug X" group. The disproportionality you're measuring is partly a signal and partly a confounder from differential polypharmacy exposure.
This isn't a flaw that better calibration of the threshold fixes. It's a structural issue with what the denominator contains.
Time-Series Insensitivity
Standard disproportionality analysis is point-in-time: you take a snapshot of the cumulative FAERS database (or a rolling window) and compute your statistics. This design means the method is not particularly sensitive to emerging signals that are growing — it detects elevated disproportionality, not increasing disproportionality. A signal that crosses PRR > 2 in quarter 3 of a given year looks the same to the algorithm whether it's been stable at 1.9 for five years or whether it went from 0.8 to 2.1 in twelve months.
For signal prioritization, the distinction matters. An emerging signal with an upward trend in reporting rate may warrant faster escalation to a formal assessment than a long-stable borderline signal, even if the current PRR values are similar. Sequential analysis methods like SPRT (Sequential Probability Ratio Test) or MaxSPRT were developed specifically to address this — they test whether the observed evidence is sufficient to trigger a decision at the current time step without exhausting the data. These methods are more computationally intensive and require careful calibration of the Type I error boundary, but they provide temporal sensitivity that standard disproportionality lacks.
We're not saying every PV team needs to replace their ROR/PRR workflow with sequential analysis tomorrow. We're saying that if your portfolio includes drugs in rapidly growing use populations — GLP-1 receptor agonists in obesity, CDK4/6 inhibitors in breast cancer, newer PCSK9 inhibitors — the temporal dynamics of emerging signals in those populations deserve more than a quarterly snapshot comparison.
Signal Dilution and Market Share Bias
FAERS reporting rates are not uniform across drugs in the same class. A newly approved agent enters the market with high reporter attention — prescribers and patients are more vigilant, spontaneous reporting is stimulated by the novelty, and the medical press covers adverse events more actively than it will five years later. An older agent in the same class, prescribed to far more patients, may have much lower reporting rates simply because it's familiar and its adverse event profile is considered established.
When you compute ROR for a new agent against an older comparator drug, you are comparing signal rates against a background that includes both well-reported and poorly-reported drugs. The "all reports not mentioning Drug X" denominator is dominated numerically by high-volume established agents with systematically lower reporting rates. This structurally inflates ROR for newer drugs and suppresses it for older drugs — not because of any real difference in pharmacological risk, but because of differential reporting behavior.
Researchers have proposed adjustments for this — using drug-specific exposure estimates from prescription databases to normalize reporting rates, stratifying the background by time-on-market, or using subgroup-specific comparators instead of the entire FAERS population. Each approach adds complexity and brings its own assumptions. None is standard practice in most PV teams' quarterly workflows.
What Automation Changes and What It Doesn't
The automation wave in pharmacovigilance has accelerated in the past few years, and some of that acceleration has been operationally useful: faster ICSR intake and processing, more consistent MedDRA coding, reduced manual data entry errors, and more systematic management of the aggregate report deduplication problem. These are real improvements.
What automation hasn't changed — and what most current PV automation platforms haven't substantively addressed — is the analytical method underlying signal detection. Automating the computation of PRR across 50,000 drug-event pairs is faster than doing it manually, but you're automating a calculation that has the same methodological properties it had when done in a spreadsheet. Speed doesn't change what the statistic measures or what its blind spots are.
The more interesting automation question is whether the additional computational capacity can be used to run analytical approaches that were previously impractical: temporal trend analysis across quarterly releases, population-stratified disproportionality that separates polypharmacy-heavy from monotherapy-dominant populations, or combination analysis that models drug co-occurrence patterns rather than treating each drug as an independent signal source.
Some of these approaches are computationally expensive relative to standard ROR. A graph-based analysis of drug co-occurrence patterns across 20 million FAERS reports is not something a two-person PV team does in Excel on a Tuesday afternoon. But for organizations where the polypharmacy gap in pairwise analysis represents a real safety risk — particularly in oncology, infectious disease, and cardiovascular therapeutic areas — the computational cost is worth examining against the cost of systematically missing interaction signals.
Calibrating Expectations for What Comes Next
Disproportionality analysis will remain part of the pharmacovigilance toolkit for the foreseeable future. The regulatory guidance documents that govern aggregate safety reporting — ICH E2C(R2) for Periodic Benefit-Risk Evaluation Reports, ICH E2F for Development Safety Update Reports — reference quantitative signal detection without mandating specific methods, which gives PV teams methodological flexibility. ROR and PRR satisfy the quantitative requirement, they're defensible to regulators, and every qualified safety scientist understands their interpretation.
The useful framing for 2025 is that disproportionality analysis is necessary but not sufficient for modern post-market safety surveillance. It is excellent at detecting elevated reporting of drug-event pairs in well-populated reporting populations with reasonably clean monotherapy exposure. It is structurally limited for polypharmacy interaction detection, temporal trend sensitivity, and populations where differential reporting rates distort the apparent signal.
PV teams that understand both the strengths and the structural limits of their core method are better positioned to know when to escalate for deeper analysis, when to be skeptical of an apparent signal that the method's assumptions don't support, and what complementary data sources — prescription databases, electronic health record linkages, spontaneous reporting from other regions — can fill the gaps that any single quantitative method leaves open.