Aggregation Bias Explained

Aggregation bias appears when a single model or decision rule is imposed on groups with meaningfully different relationships between inputs and outcomes. Pooling data can simplify implementation and improve average fit, but a shared function may fit no subgroup well if their patterns differ. This differs from representation bias, where a group or condition is missing or sparsely sampled. Both can occur together: a poorly represented group also makes it harder to detect that the pooled relationship is unsuitable. Suresh and Guttag describe aggregation bias as a lifecycle source of harm, with examples including models that do not account for differences among subpopulations. The concern is not that every group must have a separate model. Rather, teams should test whether the assumptions shared by one model hold for the relevant groups and decision context. A medical model trained across children and adults, for example, may fail if physiology, baseline risk, or the clinical meaning of a measurement differs by age. A language model can likewise interpret community-specific expressions incorrectly when it applies one pooled mapping. Averages conceal this problem. A model can score well overall while being miscalibrated or inaccurate for subgroups. Disaggregated evaluation should inspect relevant error, calibration, and utility measures, with enough examples to support conclusions. If patterns truly differ, options include group-specific models, conditional features, hierarchical models, recalibration, or different decision thresholds where lawful and appropriate. Each option has tradeoffs: separate models may have sparse data, create operational complexity, or encode group categories in ways that raise ethical and legal questions. Define groups based on relevant mechanisms rather than convenience, and consult domain experts about why relationships may differ. Validate any adaptation on separate data and monitor changes in deployment. A pooled model is justified only when evidence supports its assumptions for the population and purpose, not merely because it is easier to deploy.