Income assessment plays a critical role in credit decisions in microfinance institutions, particularly when applicants operate in informal or partially documented economic environments. In these contexts, loan officers frequently rely on subjective evaluations to estimate borrower income, which can introduce systematic disparities across demographic groups. This paper studies gender bias in the income evaluation stage of the credit process, focusing on whether female applicants receive systematically different income assessments than male applicants with comparable observable characteristics.
We first examine whether gender disparities exist in loan officer income assessments by employing regression analysis, propensity score matching, and Oaxaca-Blinder decomposition. Regression and propensity score matching allow us to estimate the gender gap in income evaluations conditional on observable borrower characteristics, while the Oaxaca-Blinder decomposition separates this gap into a component explained by differences in borrower attributes and an unexplained component potentially reflecting systematic disparities in how loan officers evaluate male and female applicants.
To formally characterize the nature of these disparities, we draw on the concept of counterfactual fairness. Under this framework, an income evaluation is considered fair if its outcome remains unchanged when the applicant's gender is set to a different value in the counterfactual world, holding all other factors constant. This allows us to isolate the share of the observed gender gap attributable to potentially discriminatory evaluation practices rather than genuine differences in borrower profiles, providing a principled basis for the corrective step that follows.
To mitigate these disparities, we propose a generative correction framework based on diffusion models that produces counterfactual income evaluations consistent with fairness constraints while preserving economically meaningful information. Diffusion models are particularly well-suited for this task as they learn the full conditional distribution of income assessments through an iterative denoising process, enabling more stable and faithful sample generation than alternative generative approaches. The diffusion-based approach generates fairness-adjusted counterfactual estimates that better capture the distributional complexity of income assessments across borrower profiles. We compare the proposed method against alternative generative approaches, including Variational Autoencoders (VAEs), evaluating their ability to produce realistic and fairness-consistent corrections.
The methodology is validated in two settings. In a controlled simulation environment, we introduce synthetic bias in the income evaluation process and assess each model’s ability to recover fair counterfactual outcomes. We then apply the framework to real microfinance credit data containing loan officer income assessments and borrower characteristics. Results show that diffusion models provide more stable and distributionally consistent corrections than alternative generative models while substantially reducing gender disparities in income evaluations.
These findings highlight the potential of generative models, particularly diffusion-based approaches, to operationalize counterfactual fairness in human-centered financial decision processes.