Dealing with measurement error

Abstract

When estimating treatment effects in real-world data (RWD), it is common to assume that exposures, confounders, mediators and treatment-effect modifiers are measured accurately and similarly across data sources, (sub)populations and treatment groups. Although variable definitions and measurement methods can be standardized when designing prospective cohort studies, their quality and definitions can vary greatly when data are collected without a specific research aim. This situation typically arises in administrative databases and/or patient registries. In this chapter, we illustrate common examples of measurement error, we give a brief overview of the types of measurement error and explain how their presence may impact estimates of real-world effectiveness or safety. Subsequently, we discuss statistical methods to adjust for measurement error, and explain how they can be implemented when RWD from multiple studies are available. Examples will be used to illustrate the main methods.