Adverse drug reaction or innocent bystander? A systematic comparison of statistical discovery methods for spontaneous reporting systems

Abstract

Purpose: Spontaneous reporting systems (SRSs) are used to discover previously unknown relationships between drugs and adverse drug reactions (ADRs). A plethora of statistical methods have been proposed over the years to identify these drug-ADR pairs. The objective of this study is to compare a wide variety of methods in their ability to detect these signals, especially when their detection is complicated by the presence of innocent bystanders (drugs that are mistaken to be associated with the ADR, since they are prescribed together with the drug that is the ADR's actual cause). Methods: Twelve methods, 24 measures in total, ranging from simple disproportionality measures (eg, the reporting odds ratio), hypothesis tests (eg, test of the Poisson mean), Bayesian shrinkage estimates (eg, the Bayesian confidence propagation neural network, BCPNN) to sparse regression (LASSO), are compared in their ability to detect drug-ADR pairs in a large number of simulated SRSs with varying numbers of innocent bystanders and effect sizes. The area under the precision-recall curve is used to assess the measures' performance.Results: Hypothesis tests (especially the test of the Poisson mean) perform best when the associations are weak and there is little to no confounding by other drugs.When the level of confounding increases and/or the effect sizes become larger, Bayesian shrinkage methods should be preferred. The LASSO proves to be the most robust against the innocent bystander effect. Conclusions:There is no absolute "winner". Which method to use for a particular SRS depends on the effect sizes and the level of confounding present in the data. K E Y W O R D S innocent bystander effect, pharmacoepidemiology, pharmacovigilance, side effect, surveillance

An investigation on the association between sodium glucose co‐transporter 2 inhibitors use and acute pancreatitis: A VigiBase study

“…Metformin, known for its potential to cause pancreatitis, 12 but also the 1st line treatment in T2D, so therefore expected to be used together with gliflozins, 1,33 was co‐suspect or interacting drug in 58 reports. The association between gliflozins and metformin was also found in all but one of the literature case reports of AP to gliflozins 22–24,26–28 . One possible mechanism proposed by other authors for AP cases associated with gliflozins and metformin could involve dehydration and lactic acidosis (due to metformin), leading to pancreatitis 19 …”

“…In addition, because diabetes is a potential risk‐factor for pancreatitis, we performed a sensitivity analysis in restricting the database to the diabetic therapeutic area (i.e., considering only medicines used in diabetes) as the background, to take into account the increased baseline risk of pancreatitis. Furthermore, we excluded from the database all reactions with co‐suspected drugs to limit the innocent bystander effect with gliflozins 28 …”

Adverse events associated with potential drugs for COVID-19: a case study from real-world data

“…We excluded non-FDA-approved drugs: favipiravir, umifenivir and camostat. The information component (IC) was calculated by comparing the observed and expected drug–AE associations in a Bayesian confidence propagation neural network [ 7 , 8 ]. These AEs reported in the full FAERS database were used as the background data.…”

“…IC025 is defined as the lower tail of a 95% credibility interval for the IC, and IC025 > 0 is considered as statistically significant [ 8 ]. The calculation of IC and reporting odds ratio (ROR) was conducted by R package ‘pvm’ [ 7 ]. ROR025 > 1 is considered as statistically significant.…”

Adverse Events Related to Off-Label Drugs Using Spontaneous Adverse Event Reporting Systems

“…In our study, we used four different data mining algorithm for signal detection using the FAERS data because the detectability depends on the effect sizes and the level of confounding data present in the data. 57 There is no gold standard for this type of algorithm, and various disproportionality analysis methods arrive at different concordance under different circumstances. 58 On the other hand, the disproportional analysis could not be conducted using KIDS-KD data in our study because the whole reports were not provided according to the policy of KIDS.…”