Case Study · Real-World Evidence · Pharmacovigilance

Sex-Stratified Pharmacovigilance Signal Detection for GLP-1 Therapies Using FAERS

Disproportionality analysis across 18.5M+ harmonized FDA adverse event records

Methods-forward, hypothesis-generating signal detection. Results describe patterns in adverse event reporting. They do not establish causality, incidence, or prevalence.

Status
Completed
Records Harmonized
18.5M+
Analysis Window
Q1 2021 – Q4 2024
Role
Analyst

Overview

Executive Summary

  • Harmonized and analyzed more than 18.5 million FAERS records to detect pharmacovigilance signals for GLP-1 receptor agonists and tirzepatide across a four-year window (Q1 2021 through Q4 2024).

  • Applied sex-stratified disproportionality analysis using Reporting Odds Ratio with 95% confidence intervals and Proportional Reporting Ratio as a sensitivity analysis, with formal comparison of female and male signal profiles.

  • All results are hypothesis-generating signals from a spontaneous reporting system. They reflect patterns in adverse event reporting, not causal relationships or population-level incidence.

Central Question

Research Question

Do GLP-1 receptor agonists and tirzepatide exhibit differential adverse event signal profiles by sex, and do these patterns vary across drug class, adverse event category, and analysis quarter?

Background

Project Context

Motivation

GLP-1 receptor agonists have experienced unprecedented prescribing growth since the approval of semaglutide for weight management. With millions of patients initiating treatment across diverse demographic groups, post-market pharmacovigilance has become essential for characterizing real-world safety profiles that clinical trial populations may not fully represent. Sex-stratified analysis is particularly important given documented differences in pharmacokinetics, body composition, and treatment indication between female and male populations.

Background

The FDA Adverse Event Reporting System (FAERS) is the largest spontaneous adverse event database for marketed drugs in the United States. While FAERS cannot establish causality or calculate incidence rates, disproportionality analysis methods identify patterns in reporting that warrant further investigation. Reporting Odds Ratio and Proportional Reporting Ratio are established methods in regulatory pharmacovigilance, used by agencies including the FDA and EMA for signal detection and hypothesis generation.

Contributions

My Role

Analyst

End-to-end responsibility for database architecture, data harmonization pipeline, methodology design, and analysis execution.

  • Designed and built the PostgreSQL schema for FAERS quarterly file ingestion across all core tables

  • Developed the data harmonization pipeline for drug name standardization and MedDRA term validation

  • Implemented FDA CASEID-based deduplication within and across quarterly files

  • Mapped brand and generic name variants to five target compounds including the GLP-1/GIP dual agonist

  • Constructed sex-stratified 2x2 contingency tables for each drug-adverse event pair

  • Computed Reporting Odds Ratio with 95% confidence intervals and PRR as a sensitivity analysis

  • Applied multiple-testing adjustment to control false discovery rate across the full signal matrix

  • Conducted stimulated-reporting checks using quarterly temporal trend analysis

  • Applied Standardised MedDRA Queries for validated adverse event groupings

Inputs

Data Sources & Scale

Records Harmonized

18.5M+

Analysis Window

Q1 2021 – Q4 2024

GLP-1 Cases

158K+

Signals Detected

665

Exposures analyzed

SemaglutideLiraglutideDulaglutideExenatideTirzepatide
FAERSSpontaneous adverse event reporting system

FDA Adverse Event Reporting System

Period

Q1 2021 through Q4 2024

Scale

18.5M+ records harmonized

Tables used

DEMODRUGREACOUTCRPSRTHERINDI

Primary data source. Spontaneous reporting system. Results represent reporting signals, not population incidence.

MedDRAMedical terminology dictionary

Medical Dictionary for Regulatory Activities

Period

Current version at time of analysis

Scale

Standardised MedDRA Queries applied for validated AE groupings

Used for preferred-term validation and SMQ-level adverse event groupings where appropriate.

Pipeline

Data Engineering Workflow

Primary data source. Spontaneous reporting system. Results represent reporting signals, not population incidence.

IngestRaw FAERS database

Load FAERS quarterly ASCII files into PostgreSQL across DEMO, DRUG, REAC, OUTC, RPSR, THER, and INDI tables.

PostgreSQLSQL
HarmonizeHarmonized records

Standardize drug names to generic terms, normalize date formats, and validate MedDRA preferred terms against the current dictionary.

SQLPython
DeduplicateDeduplicated case base

Apply FDA CASEID-based deduplication within and across quarterly files. Retain the most recent follow-up report for each case.

SQL
Define Exposures158K+ GLP-1 exposure cases

Identify cases with primary or secondary suspect GLP-1 RA or tirzepatide. Map all brand and generic name variants to the five target compounds.

SQLR
Sex-StratifyStratified contingency tables

Partition exposure cases into female and male subcohorts. Construct separate 2x2 contingency tables for each drug-adverse event pair within each sex stratum.

RSQL
Compute SignalsSignal estimates with confidence intervals

Calculate Reporting Odds Ratio with 95% confidence intervals as the primary measure, and Proportional Reporting Ratio as a sensitivity analysis. Apply multiple-testing adjustment.

R
Quality ChecksQuality-validated signal set

Conduct stimulated-reporting checks via quarterly trend analysis, reporter-type comparisons, and SMQ-level validation. Flag signals coinciding with known external events.

RPython

Methods

Analytical Methodology

Expand each method for description and purpose.

Analysis Outputs

Selected Output Previews

Charts below are rendered from actual analysis outputs. Axes, scales, and data values reflect results from the full Q1 2021 through Q4 2024 FAERS run. Results are disproportionality signals only and do not establish causality or population-level incidence.

Female vs Male ROR ComparisonAnalysis output

BH-Adjusted Signals · Jointly Evaluable PTs · Q1 2021 – Q4 2024

-2-1012-2-1012log₁₀ ROR (female stratum)log₁₀ ROR (male stratum)Equal disproportionalityFemale-dominant signalMale-dominant signalSignal in females onlySignal in males onlyShared signal

Each point is one type of adverse event, plotted by how often it was reported in female patients (x-axis) versus male patients (y-axis). Points above the dashed diagonal appear more in males; points below appear more in females. Colored points are statistically significant after multiple-comparison adjustment. Hover over any point to see the adverse event name and exact values.

Quarterly Reporting VolumeAnalysis output

Q1 2021 – Q4 2024 · FAERS Deduplicated Cases · Reporter Composition

Health professionalConsumerOther / unknownConsumer %

Each bar shows the number of GLP-1 adverse event reports received by the FDA per quarter, broken down by who submitted them. The chartreuse line tracks the share from consumer reporters (patients and caregivers). Spikes near labeled approval dates reflect stimulated reporting, a pattern where media attention temporarily inflates submission counts rather than indicating new safety concerns.

Pre-Specified Clinical Group SignalsAnalysis output

Pre-Specified Groups · ROR with 95% CI · Female vs Male

GastrointestinalPancreatitisGallbladder/BiliaryHypoglycemiaRenalPsychiatricCardiovascular0.512358Reporting Odds Ratio (95% CI)FemaleMale

Each row is an adverse event category. The dot shows how much more (or less) often that category was reported with GLP-1 therapies compared to all other drugs in FAERS. Values to the right of 1.0 (the dashed line) indicate elevated reporting. Pink represents female patients; teal represents male patients. Gastrointestinal and pancreatitis signals are elevated across both sexes; cardiovascular events are underrepresented. Hover over any point for exact figures.

Rigor

Data Quality & Methodological Safeguards

  • FDA CASEID-based deduplication removes within-quarter and cross-quarter duplicate reports, retaining the most recent follow-up for each case

  • Drug name harmonization validates brand and generic name variants against a curated reference list covering all five target compounds

  • MedDRA preferred term validation against the current MedDRA hierarchy ensures consistent adverse event classification

  • Primary suspect and secondary suspect drug roles are analyzed separately to assess role-sensitivity of signals

  • Reporter source comparisons assess whether consumer-reported and healthcare-provider-reported signals are directionally consistent

  • Quarterly trend analysis flags reporting periods that coincide with label updates, regulatory communications, or high media coverage events

  • Minimum case count threshold applied before computing disproportionality measures to exclude unstable low-frequency estimates

Responsible Interpretation

Limitations

Understanding these limitations is essential for correctly interpreting results.

Spontaneous reporting system

FAERS captures adverse events reported voluntarily by patients, healthcare providers, and manufacturers. Cases are not systematically ascertained. The database represents reporting patterns, not true adverse event incidence in treated populations.

Mitigation

Results are framed and interpreted as disproportionality signals only. No claims about incidence, prevalence, or absolute risk are made.

Notoriety and stimulated reporting bias

High-profile drugs attract disproportionate reporting independent of true pharmacological effects. GLP-1 receptor agonists received extensive media attention during the analysis window, which may artificially inflate reporting volumes and signal counts.

Mitigation

Stimulated-reporting checks using quarterly trend analysis identify and flag periods of elevated notoriety-driven reporting. Signals coinciding with known external events are interpreted with additional caution.

Absence of denominator data

FAERS does not include prescription volume, patient-years of exposure, or treated population size. Incidence rates and absolute risk estimates cannot be calculated from this data source alone.

Mitigation

Analysis is explicitly framed as disproportionality analysis. No incidence figures are presented or implied.

Variable missingness

Sex, age, weight, and other demographic variables are not uniformly reported across FAERS submissions. Substantial missingness in sex data requires careful handling in stratified analyses.

Mitigation

Missingness rates are reported per variable. Sex-stratified analyses exclude cases with missing sex or apply sensitivity analyses to assess the impact of exclusions.

Confounding by indication

GLP-1 receptor agonists are approved for both type 2 diabetes and obesity. These populations have different baseline comorbidity profiles, which can confound adverse event comparisons across drugs within the class.

Mitigation

INDI table analysis provides indication-level context where available. Drug-specific results are interpreted in light of each compound's primary approved indication.

Drug name harmonization complexity

Mapping brand name variants, biosimilar entries, and misspellings to the correct generic compound requires a curated reference list. Misclassification could lead to case misattribution.

Mitigation

A manually curated brand-to-generic mapping is maintained and versioned. Ambiguous entries are reviewed individually before inclusion.

Stack

Technology Stack

Database

PostgreSQL

Query language

SQL

Statistical analysis

R

Data pipeline

Python

Infrastructure

Docker

Version control

GitGitHub

Reflection

Lessons Learned

  • FAERS deduplication is substantially more complex than the official documentation suggests. Cross-quarter deduplication using CASEID requires careful logic to retain the most informative follow-up report for each case while discarding redundant submissions.

  • Sex missingness in FAERS is substantial for GLP-1 analyses. A documented sensitivity analysis protocol for missing sex data must be defined before interpreting stratified results.

  • Stimulated reporting from media coverage creates temporal artifacts in GLP-1 signal data. GLP-1 drugs received sustained public attention during the analysis window, making notoriety checks an essential step rather than an optional validation.

  • Drug name harmonization for GLP-1 compounds requires extensive brand-to-generic mapping. Multiple brand names, route-specific formulations, and emerging biosimilar entries all add classification complexity.

  • Empirical Bayes shrinkage improves estimate stability for low-frequency drug-adverse event pairs but requires careful parameter selection and should be validated against the ROR-based primary analysis.

Forward

Next Steps

  • Extend the analysis window to pre-2021 data to establish historical signal baselines before the GLP-1 prescribing surge began

  • Incorporate indication analysis using the INDI table to control for diabetes-versus-obesity confounding in drug comparisons

  • Implement Empirical Bayes Geometric Mean as a primary analysis alternative to complement the ROR-based approach

  • Compare detected signals against label-listed adverse events as a validation benchmark for method calibration

  • Explore integration with WHO-UMC VigiBase for international cross-validation of key signals detected in the US FAERS dataset

Interested in this work?

Whether you have questions about the methodology, want to discuss a collaboration, or are curious about applying similar approaches in your context, I would be glad to hear from you.

More work

Analytical Product · Real-World EvidenceCompleted

Real-World Evidence Studio

An analytical product concept that guides researchers from a structured clinical question through cohort definition, SQL generation, diagnostics, and interpretable evidence outputs. Designed to make real-world evidence generation more transparent and reproducible. Currently in active development.

Methods

Cohort designPropensity score methodsCausal inferenceOMOP Common Data Model

Technologies

PythonSQLROMOP CDM
Consulting · Life Sciences StrategyCompleted

Healthcare AI Strategy and Business Case

Business case development and strategic analysis for a conversational AI platform in early pregnancy care, conducted through the Penn Graduate Consulting Club. Work included economic modeling across care delivery pathways, payer and customer segment analysis, and scenario and sensitivity frameworks. Engagement details are confidential.

Methods

Economic modelingCare pathway analysisPayer and customer segment analysisSensitivity analysis

Technologies

ExcelFinancial modeling