A Causal Perspective on Label Bias

Vishwali Mhasawade, Alexander D'Amour, Stephen R. Pfohl

Abstract: Predictive models developed with machine learning techniques are commonly used to inform decision making and resource allocation in high-stakes contexts, such as healthcare and public health. One means through which this practice may propagate equity-related harms is when the data used for model development or evaluation exhibits label bias. In such cases, the target of prediction is a proxy label of a construct of interest that may be difficult or impossible to measure, while the relationship between the proxy and the construct of interest differs systematically across subgroups. Label bias can be especially challenging to identify and mitigate in practice because consequential fairness violations are masked when the model is evaluated with respect to the proxy label. In this work, we aim to develop further formal understanding of label bias to inform the development of approaches for the identification and mitigation of it. To do so, we present desiderata for unbiased and biased proxy labels, introduce candidate causal graphical criteria for label bias, and consider the extent to which proxy labels can be used to reason about fairness with respect to a true construct of interest. We validate our findings with a simulation study and experiments with synthetic health insurance data used in the context of a care management system.