RefugeeMatch — Caseworker Decision Support Tool

A machine learning system that predicts whether newly-arrived refugee families will struggle to afford monthly living expenses, helping resettlement caseworkers identify families who need proactive support.

Author: Abbot Tubeine, Sattler College Course: BUS302 Advances in Data Dataset: 2022 Annual Survey of Refugees (ASR), distributed via ICPSR

---

What's in this project

File	Purpose
refugee_model.ipynb	Training notebook — data cleaning, feature selection, model training, optimization, model export
app.py	Streamlit caseworker dashboard — intake form, risk prediction, explanations, recommendations
requirements.txt	Python dependencies
asr_2022_data.dta	Source data (download from ICPSR; rename from 2022 ASR_Public_Use_File.dta)
refugee_model.pkl	Trained model bundle (generated by running the notebook)

---

How to run end-to-end

Step 1: Train the model

1. Get the 2022 ASR data:

   • Visit ICPSR project E207021V1
   • Create a free account and download the zip
   • Extract 2022 ASR_Public_Use_File.dta
   • Rename it to asr_2022_data.dta

2. Open refugee_model.ipynb in Jupyter or Google Colab.

3. Run all cells top to bottom. The final cells (Section 14) will save refugee_model.pkl — the bundle the Streamlit app needs.

Step 2: Run the Streamlit app locally

pip install -r requirements.txt
streamlit run app.py

Then open the URL shown (typically http://localhost:8501) in your browser.

Step 3 (optional): Deploy to the web

The fastest way to share the app:

1. Push this directory to a GitHub repo (don't push asr_2022_data.dta — it's licensed)
2. Go to share.streamlit.io
3. Sign in with GitHub and click "New app"
4. Point it at your repo and select app.py as the main file
5. Click Deploy

The app will be live at a public URL within a couple of minutes.

---

How the model works

Design constraint

The model uses only intake-collectable features — things a caseworker can know during the first conversation with a newly-arrived refugee. It does NOT use post-arrival information like current employment status, current English level, or current income, even though those would improve raw accuracy.

Features used (12)

• Demographics: household size, sex, age
• Background: lived in refugee camp, years in camp, marital status
• Skills: education before U.S. arrival, native language literacy, English on arrival
• Pre-arrival history: work status in home country (Employed / Self-employed / Not working / Other)
• Placement: region of resettlement, year of arrival

Features deliberately excluded

• Country of birth — excluded for fairness reasons (protected demographic; signal is largely captured by other features)
• All post-arrival outcomes — would not be available at intake time

Performance

• Test F1 score: ~0.30–0.40 (after optimization)
• Test ROC-AUC: ~0.70–0.75
• Threshold tuned for F1 maximization given the 87/13 class imbalance

---

Limitations

This is a decision-support tool, not a replacement for caseworker judgment:

1. Small sample size — ~1,400 usable rows after cleaning
2. Severe class imbalance — only ~13% of refugees in the data couldn't afford monthly expenses
3. Region-level geography only — public-use file suppresses cities, so we can't capture city-level economic variation
4. Single survey year (2022) — outcomes may shift across years
5. Self-reported target — "can pay expenses" is subjective
6. Survivorship bias — only refugees who stayed at their resettlement address were surveyed

All flagged cases should be reviewed by a qualified caseworker. The model is meant to surface families who may benefit from extra attention, not to make resource allocation decisions on its own.

---

Citation

Urban Institute. 2022 Annual Survey of Refugees. Inter-university Consortium for Political and Social Research [distributor], 2024-09-20. DOI: 10.3886/E207021V1