Disproportionality Analysis for Pharmacovigilance Signal Detection

82 concepts63 papers1 proposal

When a new drug is approved, its safety profile is far from complete. Clinical trials are highly controlled and typically include at most a few thousand patients—making them statistically blind to rare adverse drug events (ADEs) or toxicities that only emerge across diverse, real-world populations. To catch these hidden risks, regulatory bodies rely on spontaneous reporting systems (SRSs) like the FDA Adverse Event Reporting System (FAERS), where clinicians and patients voluntarily submit reports of adverse events.

However, these massive databases present a fundamental statistical challenge: they lack a denominator. We know how many people took a drug and experienced a specific side effect, but we do not know how many people took the drug in total. Without this baseline exposure rate, traditional epidemiological metrics like absolute incidence cannot be calculated.

Disproportionality analysis (DPA) solves this problem. By treating the entire database as a pseudo-control group, DPA detects safety signals by identifying when a specific drug-event combination occurs significantly more often than random chance would dictate.


The Case-Noncase Framework

To understand why disproportionality analysis works, imagine a massive repository of millions of adverse event reports. Each report is a snapshot containing at least one drug and at least one adverse event.

Instead of looking for absolute rates, DPA structures the database into a $2 \times 2$ contingency table for a specific drug of interest and a specific event of interest:

Event of Interest All Other Events
Target Drug $A$ (Target drug + Target event) $B$ (Target drug + Other events)
All Other Drugs $C$ (Other drugs + Target event) $D$ (Other drugs + Other events)

This structure sets up a case-noncase design. If the target drug has no relationship with the target event, the proportion of reports for that event among patients taking the target drug ($\frac{A}{A+B}$) should be roughly equal to the proportion of reports for that event among patients taking any other drug ($\frac{C}{C+D}$).

When the observed frequency significantly exceeds this background expectation, a statistical "safety signal" is generated, alerting researchers to a potential causal link that warrants clinical investigation. As shown by Zhong et al. in their safety profile of sapropterin dihydrochloride and Du et al. in their study of trifluridine/tipiracil, this simple comparative framework can extract meaningful safety signals from millions of highly heterogeneous reports.


Quantifying Over-Reporting: ROR and PRR

To mathematically evaluate the contingency table, pharmacovigilance relies on distinct statistical indices. The two most common frequentist metrics are the Proportional Reporting Ratio (PRR) and the Reporting Odds Ratio (ROR).

Proportional Reporting Ratio (PRR)

The PRR directly compares the proportion of a specific event within the target drug's reports to its proportion within the rest of the database:

$$\text{PRR} = \frac{A / (A+B)}{C / (C+D)}$$

A PRR of 2 means that the adverse event is reported twice as frequently for the target drug as it is for all other medications in the database.

Reporting Odds Ratio (ROR)

The ROR borrows from case-control logic, calculating the odds of an event occurring in the exposed group versus the unexposed group:

$$\text{ROR} = \frac{A / B}{C / D} = \frac{A \times D}{B \times C}$$

As noted by Ma et al. in an international analysis of immune checkpoint inhibitor-related myocarditis, these frequentist measures offer a highly interpretable way to quantify relative over-reporting.

To prevent false positives caused by small sample sizes (e.g., if $A=1$ and $B=1$), a signal is typically only considered valid if the lower bound of the 95% confidence interval for the ROR or PRR remains greater than 1, and a minimum threshold of reports (often $A \ge 3$) is met.

Bayesian Alternatives

When report counts are very low, frequentist metrics become unstable. To address this, researchers like Zhang et al. leverage Bayesian metrics such as the Information Component (IC) in their analysis of multiple myeloma targeted therapies. The IC uses a Bayesian neural network to calculate the log-transformed ratio of the observed joint probability of a drug-event pair versus its expected probability under the assumption of independence. The Bayesian approach applies a "shrinkage" effect, pulling unstable estimates toward zero when data is scarce, thereby reducing noise.


From Single Drugs to Complex Interactions

While classic DPA evaluates one drug and one event, modern clinical realities require analyzing combinations of therapies.

Drug-Drug Interactions (DDIs)

If two drugs are taken together, do they cause a rate of adverse events that exceeds the sum of their individual risks? Ren et al. adapted DPA to study immune checkpoint inhibitors combined with proton pump inhibitors, evaluating whether the combination group demonstrated a disproportionate increase in toxicities compared to monotherapy baselines. Similarly, Xia et al. applied ROR calculations to isolate the heightened cardiac risks unique to dual immune checkpoint inhibitor combinations.

High-Order Mining

To scale this to combinations of three or more medications, traditional $2 \times 2$ tables become unwieldy. As highlighted by Ning et al. in The International Conference on Intelligent Biology and Medicine 2019, researchers utilize frequent itemset mining techniques. Instead of testing pre-specified pairs, these algorithms scan the database to find complex, multi-drug combinations that exhibit statistical "lift"—meaning the co-occurrence of the entire set of drugs and an event is significantly higher than expected based on the independent frequencies of its subsets.


Overcoming the Blind Spots of Spontaneous Data

Disproportionality analysis is highly effective, but it is fundamentally constrained by the biases inherent to voluntary reporting. It measures reporting behavior, not necessarily true clinical incidence.

Confounding by Indication

The most persistent challenge in DPA is separating a drug's side effects from the symptoms of the underlying disease it is meant to treat. For example, if a drug treats depression, reports of suicidal ideation will be disproportionately high in the database for that drug, even if the drug does not cause the ideation.

Reps et al. explored this limitation when comparing data-mining algorithms across longitudinal observational databases and electronic healthcare databases, demonstrating that automated tools frequently flag these clinical "blind spots" as false-positive safety signals.

To mitigate this, researchers use creative structural adjustments:

  • Self-Controlled Baselines: Liu et al. developed a dual-metric approach in their study of Pravastatin, calculating one ratio that compares a patient's post-exposure event rate to their own pre-exposure baseline, and a second ratio comparing it to the general population. If the event frequency spikes only after the prescription is filled, the likelihood of a true drug-induced reaction increases.
  • Likelihood Modeling: In highly unstructured environments like social media, Yates et al. introduced a Bayesian framework to map drug, symptom, and condition mentions on Twitter. By explicitly modeling the statistical likelihood of a symptom relative to the underlying condition, their framework isolates true adverse reactions when a symptom correlates strongly with a medication but weakly with the treated disease itself.

By shifting the analytical focus from absolute counts to relative proportions, disproportionality analysis turns the chaotic, uncalibrated nature of spontaneous reporting into a structured, predictive engine for global drug safety.

Key Foundational Work

Detect adverse drug reactions for drug Atorvastatin (2012) — contributed 2 concept elements to this direction.
Detect Adverse Drug Reactions for Drug Aspirin (2012) — contributed 2 concept elements to this direction.
Investigating the Detection of Adverse Drug Events in a UK General Practice Electronic Health-Care Database (2011) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 64 papers
  2. Version 1 · 2026-10-04 · 51 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_disproportionality_analysis_9ccec1e1,
  title = {disproportionality analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/disproportionality-analysis/rdr_9ccec1e18a0045e0a8ebc7a8164b3136},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_disproportionality_analysis_9ccec1e1_v2,
  title = {disproportionality analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/disproportionality-analysis/rdr_9ccec1e18a0045e0a8ebc7a8164b3136/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake