Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LSTM+Geo with xgBoost Filtering: A Novel Approach for Race and Ethnicity Imputation with Reduced Bias

S. Chalavadi, A. Pastor, T. Leitch

TL;DR

This work targets the problem of imputing race and ethnicity from incomplete self-reports, a task with important implications for disparities research and policy. It introduces LSTM+Geo, a neural model that integrates character-level name sequences with census-tract geolocation features, and an LSTM+Geo-XGBoost ensemble that applies a post-filter to correct residual errors. Across Florida and North Carolina voter data, LSTM+Geo outperforms traditional BISG and BIFSG, and reduces SES-linked misclassification biases, with an 88.7% holdout accuracy and lower White FPR compared to name-only approaches. The XGBoost post-filter further improves accuracy to 89.2% and reduces White misclassifications to 0.178, illustrating the value of combining neural sequence modeling with nonlinear ensemble corrections. These results position LSTM+Geo as a strong standalone imputation method and a powerful component in hybrid systems, though the authors caution against using these models for individual-record decisions and emphasize the need for broader validation across datasets and contexts.

Abstract

Accurate imputation of race and ethnicity (R&E) is crucial for analyzing disparities and informing policy. Methods like Bayesian Improved Surname Geocoding (BISG) are widely used but exhibit limitations, including systematic misclassification biases linked to socioeconomic status. This paper introduces LSTM+Geo, a novel approach enhancing Long Short-Term Memory (LSTM) networks with census tract geolocation information. Using a large voter dataset, we demonstrate that LSTM+Geo (88.7% accuracy) significantly outperforms standalone LSTM (86.4%) and Bayesian methods like BISG (82.9%) and BIFSG (86.8%) in accuracy and F1-score on a held-out validation set. LSTM+Geo reduces the rate at which non-White individuals are misclassified as White (White FPR 19.3%) compared to name-only LSTMs (White FPR 24.6%). While sophisticated ensemble methods incorporating XGBoost achieve the highest overall accuracy (up to 89.4%) and lowest White FPR (17.8%), LSTM+Geo offers strong standalone performance with improved bias characteristics compared to baseline models. Integrating LSTM+Geo into an XGBoost ensemble further boosts accuracy, highlighting its utility as both a standalone model and a component for advanced systems. We give a caution at the end regarding the appropriate use of these methods.

LSTM+Geo with xgBoost Filtering: A Novel Approach for Race and Ethnicity Imputation with Reduced Bias

TL;DR

This work targets the problem of imputing race and ethnicity from incomplete self-reports, a task with important implications for disparities research and policy. It introduces LSTM+Geo, a neural model that integrates character-level name sequences with census-tract geolocation features, and an LSTM+Geo-XGBoost ensemble that applies a post-filter to correct residual errors. Across Florida and North Carolina voter data, LSTM+Geo outperforms traditional BISG and BIFSG, and reduces SES-linked misclassification biases, with an 88.7% holdout accuracy and lower White FPR compared to name-only approaches. The XGBoost post-filter further improves accuracy to 89.2% and reduces White misclassifications to 0.178, illustrating the value of combining neural sequence modeling with nonlinear ensemble corrections. These results position LSTM+Geo as a strong standalone imputation method and a powerful component in hybrid systems, though the authors caution against using these models for individual-record decisions and emphasize the need for broader validation across datasets and contexts.

Abstract

Accurate imputation of race and ethnicity (R&E) is crucial for analyzing disparities and informing policy. Methods like Bayesian Improved Surname Geocoding (BISG) are widely used but exhibit limitations, including systematic misclassification biases linked to socioeconomic status. This paper introduces LSTM+Geo, a novel approach enhancing Long Short-Term Memory (LSTM) networks with census tract geolocation information. Using a large voter dataset, we demonstrate that LSTM+Geo (88.7% accuracy) significantly outperforms standalone LSTM (86.4%) and Bayesian methods like BISG (82.9%) and BIFSG (86.8%) in accuracy and F1-score on a held-out validation set. LSTM+Geo reduces the rate at which non-White individuals are misclassified as White (White FPR 19.3%) compared to name-only LSTMs (White FPR 24.6%). While sophisticated ensemble methods incorporating XGBoost achieve the highest overall accuracy (up to 89.4%) and lowest White FPR (17.8%), LSTM+Geo offers strong standalone performance with improved bias characteristics compared to baseline models. Integrating LSTM+Geo into an XGBoost ensemble further boosts accuracy, highlighting its utility as both a standalone model and a component for advanced systems. We give a caution at the end regarding the appropriate use of these methods.
Paper Structure (25 sections, 2 figures, 5 tables)

This paper contains 25 sections, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Existing Research: A Survey of Race and Ethnicity Imputation Methods
  3. Introduction
  4. Bayesian Improved Surname Geocoding (BISG)
  5. BIFSG – Adding First Name to BISG
  6. LSTM-Based Name Models (Neural Networks)
  7. raceBERT (Neural Networks)
  8. Hybrids: Geo Models with Boosted Decision Tree Post Filters (Argyle & Barber, SSA Filter Model, Zesat.ai ZRP)
  9. Proposed Models
  10. Data and Feature Engineering
  11. Geolocation Enrichment.
  12. LSTM+Geo Model
  13. Model Architecture and Features.
  14. XGBoost Post-Filtering Model
  15. Rationale.
  16. ...and 10 more sections

Figures (2)

  • Figure 1: Misclassification Rates by Census Tract Income. The graph examines how misclassification rates vary with income across different racial groups in several datasets. It standardizes the race predictions, groups income into relevant ranges, and computes misclassification rates for each group.
  • Figure 2: Misclassification Rates by Income Level and Race: Comparing baseline LSTM, LSTM+Geo, ZRP, and LSTM+Geo+XGBoost ensemble on nationwide PPP validation dataset. Results highlight effectiveness of geolocation and XGBoost filtering across socioeconomic contexts.