Gender is prohibited from use in decision-making in many countries. This does not necessarily benefit females in situations of automated algorithmic decisions, e.g. when a credit scoring model is used as a decision tool for loan granting. By analysing a unique proprietary dataset on car loans from an EU bank, this paper shows that Gender as a variable in a credit scoring model is statistically significant. Its removal does not alter the predictive accuracy of the model, yet the proportions of accepted women/men depend on whether Gender is included. The paper explores the association between predictors in the model with Gender, to demonstrate the omitted variable bias and how other variables proxy for Gender. It points to inconsistencies of the existing regulations in the context of automated decision-making.
|Journal||Journal of the Royal Statistical Society: Series A|
|Early online date||1 Aug 2019|
|Publication status||Published - 25 Oct 2019|
- credit scoring
- algorithmic decision-making
- statistical discrimination