Drug exposure misclassification in pharmacoepidemiology: Sources and relative impact

Abstract

Background: Drug exposure assessment based on dispensing data can be misclassified when patients do not adhere to their therapy or when information about over-the-counter drugs is not captured in the study database. Previous research has considered hypothetical sensitivity and specificity values, whereas this study aims to assess the impact of literature-based real values of exposure misclassification.Methods: A synthetic cohort study was constructed based on the proportion of exposure theoretically captured in a database (range 0.5-1.0) and the level of adherence (0.5-1.0). Three scenarios were explored: nondifferential misclassification, differential misclassification (misclassifications dependent on an unmeasured risk factor doubling the outcome risk), and nondifferential misclassification in a comparative effectiveness study (RR A and RR B both 2.0 compared to nonuse, RR A-B 1.0).Results: For the scenarios with nondifferential misclassification, 25% nonadherence or 25% uncaptured exposure changed the RR from 2.0 to 1.75, and 1.95, respectively.Applying different proportions of nonadherence or uncaptured use (20% vs. 40%) for subgroups with and without the risk factor, an RR of 0.95 was observed in the absence of a true effect (i.e., true RR = 1). In the comparative effectiveness study, no effect on RR was seen for different proportions of uncaptured exposure; however, different levels of nonadherence for the drugs (20% vs. 40%) led to an underestimation of RR A-B (0.89).Discussion: All scenarios led to biased estimates, but the magnitude of the bias differed across scenarios. When testing the robustness of findings of pharmacoepidemiologic studies, we recommend using realistic values of nonadherence and uncaptured exposure based on real-world data.

A real-world study on SGLT2 inhibitors and diabetic kidney disease progression

“…The systematic differences in baseline characteristics between treatment-exposed and unexposed groups departed from what we usually observed in RCTs. We do not know if patients adhered to the assigned treatment or received medications from other healthcare institutions, leading to misclassification of drug exposure [ 42 ]. Essentially we commenced the prescription of SGLT2is in early 2016.…”

To clarify the safety profile of paracetamol for home-care patients with COVID-19: a real-world cohort study, with nested case–control analysis, in primary care

“…Third, prescriptions of paracetamol for COVID-19 might be underestimated because it is not fully reimbursed. Nevertheless, the exposure misclassification due to paracetamol use as an OTC drug [ 28 ] might have caused biased results towards the null. When we conducted probabilistic sensitivity analysis [ 27 ], the bias-adjusted OR were consistent with those obtained for the primary analysis.…”

“…First, to consider the burden of misclassification for COVID-19 diagnoses, were re-ran the primary analyses by limiting cases to those classified as “confirmed”. Second, we carried out a probabilistic sensitivity analysis [ 27 ] to evaluate the effect on the results of exposure misclassification due to paracetamol use as OTC medication [ 28 ]. We assumed that younger individuals were featured by frequent use of OTC medications and less severe infection.…”

Primary nonadherence to drugs prescribed by general practitioners: A Dutch database study

“…Yet the impact of 10% exposure misclassification due to nonadherence and the difference in the level of PNA between subgroups are likely limited. For example, a simulation study of the impact of exposure misclassification on effect estimates revealed that a 10% nondifferential nonadherence could cause an approximate 10% bias toward the null effect, and the impact of differential misclassification was also limited 55 …”

“…For example, a simulation study of the impact of exposure misclassification on effect estimates revealed that a 10% nondifferential nonadherence could cause an approximate 10% bias toward the null effect, and the impact of differential misclassification was also limited. 55 …”