Assessing the Impact of Non-Random Measurement Error on Inference: A Sensitivity Analysis Approach

Abstract

Many commonly used data sources in the social sciences suffer from non-random measurement error, understood as mis-measurement of a variable that is systematically related to another variable. We argue that studies relying on potentially suspect data should take the threat this poses to inference seriously and address it routinely in a principled manner. In this article, we aid researchers in this task by introducing a sensitivity analysis approach to non-random measurement error. The method can be used for any type of data or statistical model, is simple to execute, and straightforward to communicate. This makes it possible for researchers to routinely report the robustness of their inference to the presence of non-random measurement error. We demonstrate the sensitivity analysis approach by applying it to two recent studies.