City Research Online

Abstract

Excluding protected attributes from insurance pricing models does not guarantee the absence of discrimination, as remaining covariates may still act as proxies. This paper develops a differential framework for measuring proxy discrimination in fitted pricing models. The approach decomposes the sensitivity of prices to a covariate into components reflecting the direct impact on predictions and the proxying of a protected attribute. In applications with predominantly categorical features, we use Multiple Correspondence Analysis (MCA) to obtain a continuous latent representation that supports differential sensitivity analysis. In that setting, we decompose discrete price changes due to policyholder profile perturbations into direct and proxy effects, using Aumann-Shapley attributions. Two empirical illustrations on insurance claims datasets reveal weak but detectable proxy effects related to gender. The proposed approach provides a diagnostic framework for assessing the presence and magnitude of proxy-discriminatory effects at both the individual and portfolio level, which is applicable to differentiable pricing models, without requiring model refitting.

Publication Type:	Monograph (Working Paper)
Publisher Keywords:	Proxy discrimination, insurance pricing, fairness, sensitivity analysis, Multiple Correspondence Analysis
Subjects:	H Social Sciences > HF Commerce
Departments:	Bayes Business School Bayes Business School > Faculty of Actuarial Science & Insurance
SWORD Depositor:	Symplectic Administrator

Export

Downloads