Likelihood Ratio Test-Based Method for Signal Detection in Drug Classes Using FDA's AERS Database

Abstract

In 1968 the Food and Drug Administration (FDA) established the Adverse Event Reporting System (AERS) database containing data on adverse events (AEs) reported by patients, health care providers, and other sources through a spontaneous reporting system. FDA uses AERS for monitoring the safety of the drugs on the market after approval. Most statistical methods that are available in the literature to analyze large postmarket drug safety data for identifying drug-event combinations with disproportionately high frequencies are designed to explore signals of a single drug-AE combination, but not signals including a drug class or a group of AEs simultaneously. Those methods are also not designed to control type I error and are subject to high false discovery rates. In this paper, we first briefly review a recently developed method, known as the likelihood ratio test (LRT)-based method, which has been demonstrated to control the family-wise type I error and false discovery rates. By introducing a concept of weight matrix for the drugs (or for AEs), we then extend the LRT method for detecting signals including a class of drugs (or AEs) in addition to detecting signals of single drug (or AE). A simplified Bayesian method is also proposed and compared with LRT method. The proposed methods are applied to study the signal patterns of drug classes, namely, the gadolinium drug class for magnetic resonance imaging (MRI) and statins for hypercholesterolemia, over different time periods, using the datasets with only suspect drugs and with both suspect and concomitant drugs from the AERS database. The signals detected by the statistical methods can be confirmed by signals detected across different databases, existing medical evidence from research or regulatory resources, prospective biological studies, and also through simulation as illustrated in the application.

An evaluation of statistical approaches to postmarketing surveillance

“…That is, “The EudraVigilance Expert Working Group noted that thresholds commonly used to detect signals in spontaneous data are a trade‐off between two conflicting goals: either generating too many false positive signals if the threshold is too low or missing true signals if this threshold is too high. Given this trade‐off, it is important to identify and calibrate methods to strike a reasonable balance between these two parameters.” Although this evaluation was very extensive, it did not include a more recently developed approach using a likelihood ratio test (LRT)‐based method and the FDR‐based extensions of the aforementioned methods. In addition, comparisons among these methods are based on using data from different databases.…”

“…We refer to these outcome measures as the AE‐based and drug‐based outcome measures, respectively. In their evaluation of the LRT method, Huang et al used the AE‐based method but demonstrated good performance in several contexts . Specifically, the LRT method controls both the type I error and the FDR well while retaining good power and sensitivity, and Huang et al compared its performance with the PRR and BCPNN methods.…”

Translational Biomedical Informatics and Pharmacometrics Approaches in the Drug Interactions Research

“…By treating a drug combination as a new drug, the disproportionality measurements can be obtained, accordingly. For instance, Huang et al . introduced an extended LRT method that can be used for detecting signals for multiple drugs (or ADEs) in a drug class (or in an ADE group).…”

Revisiting the reported signal of acute pancreatitis with rasburicase: an object lesson in pharmacovigilance

“…Applying the latter methods to a reference set of adverse events with 220 drugs in four pharmaceutical companies, one national and two international safety databases, they concluded: 'The choice of disproportionality statistic does not appreciably affect the achievable range of SD performance and so this can be primarily based on ease of implementation, interpretation and minimization of computing resources'. Even with new varieties of DA, such as p-plots, Ƴ-multinomial modeling and likelihood-based methods [Johnson et al 2012;Huang et al 2013;, the latest data do not support a religious devotion to a single method.…”