1.29M
loan records
Under review · Grade 10
Algorithmic Credit Scoring and Disability-Correlated Exclusion: Evidence from a Large-Scale Credit Dataset with Implications for Microfinance
Does a credit-scoring model quietly turn away profiles that look like disability? On 1.29 million loans, yes, by 2.8 points.
Submitted to Andover Economics Review
- loan records
- 1.29M
- loan records
- approval gap for the proxy group
- 2.8 pts
- approval gap for the proxy group
- false-denial gap
- 2.0 pts
- false-denial gap
- AUC after removing the gap, nearly unchanged
- 0.7050 → 0.7048
- AUC after removing the gap, nearly unchanged
01 Why I did it
The question.
At Prayatna I kept asking what happens to the children I coach when they grow up. If one of them wanted to start a business, would a loan become another obstacle? That question became this paper. The answer suggests disability status should be collected voluntarily and confidentially, used for fairness audits and never for the lending decision.
02 Abstract
What the paper says.
A proxy group was constructed from 1,292,175 Lending Club records using characteristics reported in India's NSS Report No. 583, scaled to the 2.2 percent prevalence of disability in India. The proxy is not a measure of disability. For the proxy group the baseline model showed a statistically significant approval gap of 2.8 percentage points and a false-denial gap of 2.0 points (p < .001), unaffected by reweighting. Closing the gap did not lower the control group's approvals; it overcorrected the proxy group above the control while AUC moved from 0.7050 to 0.7048.
03 Method and results
How, and what came out.
- Logistic regression proxy model over 1.29 million loan records
- Proxy built from NSS Report No. 583 characteristics, scaled to 2.2 percent prevalence
- Reweighting test to rule out sampling imbalance
- Threshold re-engineering and AUC comparison before and after
2.8 pts
approval gap for the proxy group
2.0 pts
false-denial gap
0.7050 → 0.7048
AUC after removing the gap, nearly unchanged