KC-OverdoseModels2025/TweedieGLM_Census-Cluster-Velocity

Tweedie GLM — Fentanyl Overdose Rate Forecasting

Author: Ansh Gupta

Overview

A Tweedie Generalized Linear Model (GLM) for forecasting census tract-level fentanyl/opioid overdose rates in King County, Washington. The model integrates American Community Survey (ACS) census data, Group-Based Trajectory Modeling (GBTM) cluster assignments, and a year-over-year overdose rate velocity feature to predict future overdose rates per 100,000 population.

Performance (5-Fold Cross-Validation)

Metric	Value
R²	0.861
MAE	8.50 per 100,000
Spearman ρ	0.859
Top-10 Overlap	100.0%
BPR K=10	100.0%
Top-20 Overlap	95.0%
BPR K=20	98.7%
Top-50 Overlap	66.0%
BPR K=50	89.2%
Mean BPR (K=5–200)	94.9%

Features

Feature	Description
Med_HHD_Inc_Thousands_ACS__	Median household income (thousands)
pct_Renter_Occp_HU_ACS__	% renter-occupied housing units
pct_Vacant_Units_ACS__	% vacant housing units
pct_Not_HS_Grad_ACS__	% without high school diploma
pct_College_ACS__	% with college degree
pct_NH_Blk_alone_ACS__	% non-Hispanic Black population
Pct_No_Health_Ins_CALCULATED_ACS__	% without health insurance
Med_HHD_Inc_Thousands_ACS___Neighbor_Avg	Average median income of 3 nearest neighbor tracts
pct_Vacant_Units_ACS___Neighbor_Avg	Average vacancy rate of 3 nearest neighbor tracts
pct_Renter_Occp_HU_ACS___Neighbor_Avg	Average renter rate of 3 nearest neighbor tracts
Cluster_1	GBTM trajectory cluster 1 indicator
Cluster_2	GBTM trajectory cluster 2 indicator
Cluster_3	GBTM trajectory cluster 3 indicator
Velocity_1Yr	Year-over-year change in overdose rate (Rate_2024 − Rate_2023)

Model Specification

Parameter	Value
Family	Tweedie
Link Function	Log
Variance Power	1.5
Velocity Formula	Rate_2024 − Rate_2023

Validation

• Temporal: Out-of-time forecast (trained on ≤2024 data, predicting 2025 TTM rates)
• Cross-validation: 5-fold CV — every tract predicted by a model never trained on it
• Geographic: Cross-regional transfer validated on Los Angeles County, CA

Training Data

• Region: King County, Washington (463 census tracts)
• Target: Overdose death rate per 100,000 (2025 trailing twelve months)
• Census data: ACS 5-year estimates (2019–2023)

Usage

import pickle
import statsmodels.api as sm
import pandas as pd

# Load model
with open("tweedie_model.pkl", "rb") as f:
    result = pickle.load(f)

# View coefficients and rate ratios
print(result.summary())

# Predict on new data
predictions = result.predict(sm.add_constant(your_features_df))

Files

File	Description
tweedie_model.pkl	Fitted statsmodels GLM result object
coefficients.json	All coefficients, rate ratios, and p-values (readable JSON)
model_config.json	Model configuration (family, link, var_power, features)
model_summary.txt	Full statsmodels summary table
training_data.csv	Full training dataset (463 tracts)

License

MIT