Author: Rajveer Singh Pall
Institution: Gyan Ganga Institute of Technology and Sciences
Status: Work in progress — arXiv preprint targeted Month 5
This repository extends Persistent Racial Disparities in U.S. Mortgage Approval: Evidence from 42 Million Applications, 2020–2024 — a companion paper documenting an average within-lender racial approval penalty of 10.6 percentage points using regression discontinuity, difference-in-differences, and DFL decomposition.
That paper answers: Is there a racial approval gap?
This project answers: For whom is it largest — and why?
Using Causal Forests (Wager & Athey, 2018) and Double Machine Learning (Chernozhukov et al., 2018), we estimate the Conditional Average Treatment Effect (CATE) of being a Black applicant on mortgage approval probability across the full distribution of 42 million applicants. We then map how the approval penalty varies across income decile, loan-to-value ratio, lender tier, and loan purpose — identifying the applicant profiles most severely affected by institutional discretion.
- Is the 10.6 pp average penalty uniform, or does it mask severe heterogeneity?
- Is discrimination largest for high-income applicants (suggesting animus) or lower-income, higher-LTV applicants (suggesting structural risk aversion)?
- Is the 2.0 pp PMI boundary effect (at 80% LTV) concentrated in specific applicant subgroups?
- Did the 2022 Fed tightening cycle widen the gap uniformly, or did it amplify existing disparities for already-disadvantaged applicants?
Repo 1 (published paper): HMDA-Racial-Disparities
Contains 16 notebooks covering data cleaning, DFL decomposition, within-lender fixed effects, LTV-RDD, tightening-DiD, permutation inference, and all manuscript tables and figures.
This repository (Repo 2) uses the processed panel files from Repo 1 as inputs.
DML handles high-dimensional covariates (100+ features including income, LTV, loan purpose, property type, occupancy type, lender size, geographic dummies) by:
- Residualising the treatment (Black indicator) on all covariates using LightGBM
- Residualising the outcome (approval) on all covariates using LightGBM
- Regressing outcome residuals on treatment residuals
5-fold cross-fitting eliminates regularisation bias. Standard errors clustered at the lender (LEI) level.
Grows 500 trees that split on covariates maximising treatment effect heterogeneity (not prediction accuracy). Honest splitting — training and estimation samples are separate within each tree. Produces a CATE estimate for every individual applicant.
shap.TreeExplainer on the fitted Causal Forest identifies which covariates drive heterogeneity in treatment effects, producing the personalised disparity map.
CATE-HMDA-Heterogeneous-Effects/
│
├── README.md
├── requirements.txt
├── environment.yml
├── .gitignore
│
├── notebooks/
│ ├── NB17_feature_engineering.ipynb
│ ├── NB18_overlap_diagnostics.ipynb
│ ├── NB19_double_ml_baseline.ipynb
│ ├── NB20_nonparam_dml_cate.ipynb
│ ├── NB21_causal_forest_cate.ipynb
│ ├── NB22_shap_attribution.ipynb
│ ├── NB23_disparity_map.ipynb
│ ├── NB24_subgroup_rdd.ipynb
│ ├── NB25_subgroup_did.ipynb
│ └── NB26_paper_figures.ipynb
│
├── outputs/
│ ├── tables/ # CSV and LaTeX tables (generated)
│ └── figures/ # PNG and PDF figures (generated)
│
├── data/
│ └── README.md # Data download instructions (no data committed)
│
└── paper/
└── CATE_HMDA_draft.tex
Source: CFPB HMDA Public Loan-Level Datasets
Years: 2020–2024 | Observations: 42,323,519 | Lenders: 5,500+
This repository does not contain data files. Raw HMDA CSVs are publicly available at the link above. Processed panel CSVs are generated by Repo 1 (notebooks 01 and 02).
| File | Source | Size (approx) |
|---|---|---|
hmda_2020.csv through hmda_2024.csv |
CFPB HMDA portal | ~2–4 GB each |
processed/panel_2020.csv through panel_2024.csv |
Generated by Repo 1, NB01 | ~500 MB each |
# 1. Clone this repository
git clone https://github.com/YOUR_USERNAME/CATE-HMDA-Heterogeneous-Effects.git
cd CATE-HMDA-Heterogeneous-Effects
# 2. Create conda environment
conda env create -f environment.yml
conda activate cate_hmda
# 3. Register the kernel for Jupyter
python -m ipykernel install --user --name cate_hmda --display-name "CATE-HMDA"
# 4. Place data files in data/ (see Data section above)
# 5. Run notebooks in order: NB17 → NB18 → ... → NB26| Notebook | Description | Runtime (est.) | Key output |
|---|---|---|---|
| NB17 | Feature engineering | 20–40 min | features_panel.parquet |
| NB18 | Overlap diagnostics | 10–20 min | Propensity score plots, trimmed sample |
| NB19 | Double ML baseline (LinearDML) | 20–40 min | ATE ≈ −10.6 pp (sanity check) |
| NB20 | NonParamDML CATE | 30–60 min | First CATE distribution plot |
| NB21 | CausalForestDML | 60–120 min | Individual CATE estimates |
| NB22 | SHAP attribution | 10–20 min | Heterogeneity drivers |
| NB23 | Disparity map | 5–10 min | Income × LTV × lender-tier heatmap |
| NB24 | Subgroup RDD | 20–40 min | PMI boundary effect by subgroup |
| NB25 | Subgroup DiD | 20–40 min | Tightening cycle effect by CATE quartile |
| NB26 | Paper figures | 10–20 min | All manuscript figures |
Total runtime: approximately 3.5–7 hours (hardware: i7-13650HX, 16GB RAM)
- Wager, S., & Athey, S. (2018). Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 113(523), 1228–1242.
- Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1), C1–C68.
- Bartlett, R., Morse, A., Stanton, R., & Wallace, N. (2022). Consumer-lending discrimination in the FinTech era. Journal of Financial Economics, 143(1), 30–56.
- Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. NeurIPS, 30.
If you use this code, please cite both this repository and the companion paper:
@misc{pall2026cate,
author = {Pall, Rajveer Singh},
title = {Heterogeneous Treatment Effects in Mortgage Lending},
year = {2026},
url = {https://github.com/YOUR_USERNAME/CATE-HMDA-Heterogeneous-Effects}
}Code: MIT License
Data: Subject to CFPB HMDA public data terms