. 2023 Jul 14;4(7):100790. doi: 10.1016/j.patter.2023.100790

Table 1.

A path forward for practitioners to help diagnose and mitigate the different causes of bias

Cause	Effect	Diagnosis	Mitigation
Label noise, label biases, selection biases	estimator bias, uninformative performance estimates	domain expertise, analyze label correlation with proxy variables,³⁵ gather higher-fidelity labels²²^,³¹	use other target variables,³¹^,³⁵ bias-robust learning techniques⁴⁴^,⁴⁵^,⁴⁶
Concept shift: differences in $p (y ∣ x)$ between groups	estimator bias	investigate effects of group balancing and model stratification²²^,⁴⁷	use stratified model,²²^,⁴⁷ gather additional features
Low model expressivity, differences in $p (x)$ between groups	estimator bias	investigate effects of group balancing²²^,⁴⁷ and increasing model expressivity	increase model expressivity
Underrepresentation and highly expressive model	high estimator variance	epistemic uncertainty quantification,¹⁵^,¹⁶ analysis of sample size-performance relationship per group¹³^,⁴⁸	gather more samples,⁴⁸^,⁴⁹ decrease model expressivity, regularize
High task difficulty	high irreducible error	aleatoric uncertainty quantification,¹⁵^,¹⁶ analysis of sample size-performance relationship per group¹³^,⁴⁸	gather additional or alternative features,³¹^,⁵⁰^,⁵¹ reformulate prediction task or target population

While these can help diagnose and mitigate bias in practice, they do not come with guarantees, and improved diagnostics and mitigation remain an open research problem. The list of potential causes of performance differences is not exhaustive.