greenadntan/child-welfare-outcomes-institutional-vs-homebased

Child Welfare Outcomes: Institutional Care vs Home-Based Professional Nearness (HBPN)

🎯 Overview

Synthetic longitudinal dataset comparing child welfare outcomes between institutional/residential care and home-based professional nearness (HBPN) — a family-based intervention model where trained professionals provide intensive support within a home setting.

This dataset preserves the well-documented "twice the success rate" finding from decades of child welfare research: children in home-based professional care achieve approximately 2× the placement success rate of those in institutional settings, at substantially lower cost.

⚠️ This is synthetic data generated to match published research distributions. It should not be used as a substitute for real clinical data, but is suitable for policy modeling, cost-effectiveness analysis, ML method development, and educational purposes.

📊 Key Findings (Built Into the Data)

ICER (Incremental Cost-Effectiveness Ratio)

HBPN dominates institutional care: it is simultaneously cheaper AND more effective. The negative ICER of -$670,834 per additional success (95% CI: [-$732,571, -$620,069]) indicates a cost-saving intervention.

📚 Literature Calibration

This synthetic data is calibrated to findings from major studies:

📁 Dataset Structure

4 Configurations (Subsets)

outcomes — Child-level final outcomes (5,000 rows × 24 columns)

The primary analysis table with one row per child.

trajectories — Quarterly longitudinal records (60,000 rows)

12 quarterly measurement points per child over 3 years.

costs — Monthly cost breakdown (180,000 rows)

36 monthly cost records per child with component breakdown.

summary — Aggregate program statistics (2 rows)

Quick-reference aggregates per care type.

🤖 Included Models

Two XGBoost models are included in the models/ directory:

1. Success Classifier

• —Task: Predict placement success (binary) from child features + care type
• —Performance: ROC AUC = 0.674 (test), 5-fold CV AUC = 0.681 ± 0.016
• —Top feature: care_type (49% importance) → care modality is the strongest predictor

2. Cost-Effectiveness Regressor

• —Task: Predict cost per wellbeing point
• —Performance: R² = 0.980, MAE = $272
• —Top feature: care_type (93% importance) → modality dominates cost-effectiveness

Treatment Effect Heterogeneity

The models reveal that HBPN benefits are broadly distributed:

• —By age: 3-6 (+2.7%), 7-10 (+3.0%), 11-16 (+3.1%)
• —By adversity: Low (+3.0%), Medium (+3.1%), High (+2.5%)

💡 Usage

from datasets import load_dataset

# Load specific subset
outcomes = load_dataset("greenadntan/child-welfare-outcomes-institutional-vs-homebased", "outcomes")
trajectories = load_dataset("greenadntan/child-welfare-outcomes-institutional-vs-homebased", "trajectories")
costs = load_dataset("greenadntan/child-welfare-outcomes-institutional-vs-homebased", "costs")

# Quick analysis
import pandas as pd
df = outcomes["train"].to_pandas()
print(df.groupby("care_type")["success"].mean())
# HBPN             0.638
# institutional    0.323

🔬 Methodology

Synthetic Data Generation

1. 1.Demographics: Sampled from US child welfare demographic distributions
2. 2.Trajectories: Sigmoid growth curves with individual noise, regression events, and disruption modeling
3. 3.Costs: Log-normal monthly costs with component breakdown, step-down modeling for HBPN, and crisis cost spikes
4. 4.Outcomes: Logistic model calibrated to hit target success rates (32%/64%) while preserving individual-level covariate effects

Cost-Effectiveness Analysis

• —ICER: Incremental Cost-Effectiveness Ratio with 5,000-iteration bootstrap CIs
• —CEA Plane: HBPN falls in the southeast (dominant) quadrant — lower cost, higher effect

⚖️ Limitations & Ethics

• —Synthetic data only — distributions match published research, but individual records are simulated
• —US-centric cost structures — international contexts may differ significantly
• —Simplified model — real child welfare involves complex multi-system interactions not captured here
• —Selection bias not modeled — real-world placement decisions are non-random; this data assumes comparable populations
• —Not for clinical decisions — this data supports policy modeling and method development, not individual-level care planning

📄 Citation

If you use this dataset, please cite both the dataset and the foundational research:

@dataset{child_welfare_outcomes_2026,
  title={Child Welfare Outcomes: Institutional Care vs Home-Based Professional Nearness},
  author={greenadntan},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/greenadntan/child-welfare-outcomes-institutional-vs-homebased}
}

Foundational References

• —Nelson, C. A., Fox, N. A., & Zeanah, C. H. (2014). Romania's Abandoned Children. Harvard University Press.
• —Chamberlain, P. (2003). Treating Chronic Juvenile Offenders. American Psychological Association.
• —Dozier, M., et al. (2014). Effects of foster care on young children's language learning. Child Development, 85(4).
• —Doyle, J. J., & Peters, H. E. (2007). The market for foster care. Review of Economics of the Household, 5(4).