Bias in Demographic Models

Artificial Intelligence (AI) and Machine Learning (ML) have made tremendous advancements in recent decades. AI/ML models have been used in demographic research to gain insights for specific populations and research focuses. While these advanced models are certainly capable of providing novel and in-depth analysis, challenges related to bias and fairness remain a major issue. To address this issue of bias identification and mitigation, AI/ML models must be designed with fairness and trustworthiness as a core component of the model. Towards the fairness and trustworthiness of AI/ML models, explainable AI (XAI) has garnered interest in filling the gaps where traditional AI/ML models fall short. Explainability plays a central role in ensuring the fairness and trustworthiness of AI/ML models. In this project, we highlight the use of XAI to identify bias within AI/ML models and the datasets used for these models.

Detailed Information

• Model name: Language-Income Classification
• Model author(s) and affiliations (inquiries can be sent to inquiries@xd.gov):
  • Atul Rawal, Ph.D (xD - Office of Deputy Director)
  • Sandy L Dietrich, Ph.D (Social, Economic, and Housing Statistics Division)
  • James McCoy (xD - Office of Deputy Director)
• Model acquisition/development method: Internally built by Atul Rawal, Sandy L Dietrich & James McCoy

• Division(s)using the model: xD & SEHSD
• Intended application(s) and stakeholder(s) of the model: Research studies for language equity in the US by xD & SEHSD

• Current model version and release date: V1, released on 11/3/2023
• Changes made since the last release. (If any): N/A
• License for use: N/A
• Type of model (Classification or Regression): Classification

• Type of algorithm used: Multiple ( RF, LR, GBR, LGBM, XGB, CatBoost & CNN)

• Source(s) of the training data: IPUMS data repository for ACS data
• Data collection/ generation method: Data downloaded form IPUMS for 2015 - 2019 period
• Number of variables in this dataset: 16
• Number of entries in this dataset: 500,000
• Percent of data chosen as a training, testing and validation set: 80% training & 20% testing

• Metrics used to rate model performance:
  • Accuracy
  • Precision
  • Recall
  • F1-score
• Factors that limit the model's performance (Example: Limited dataset, Number of Nulls/NAs) (If any): N/A

• Inclusion of information related to individuals or human populations in the training/testing/validation dataset: Yes, sensitive attributes such as age, sex, race and ethnicity
• Degree of risk of human judgement injecting bias within the workflow: N/A
• Methods used to minimize bias from human judgement: N/A
• Potential biases found in the training dataset from collection methods, sample size, representation, etc.: Bias in sample distribution
• Testing/evaluation performed to look for bias in the workflow of the model: Equal distribution of the languages spoken at home
• Degree of model explainability/transparency: Post-hoc explainability via SHAP

• Model/dataset compliance with existing laws and regulations (Including privacy protection regulations): Yes, compliance with both Title 13 & 26 regulations