Overview Pipeline Datasets Analysis Models Findings
Penn State DAAN 881  |  Data-Driven Decision Making

Optimizing Vehicle Insurance Terms
Using Opioid Overdose Trends

A multi-state crash risk intelligence system that integrates 2.5M+ records across five state and national datasets, engineers 70+ behavioral and geospatial features, and trains LightGBM, XGBoost, and Neural Network classifiers to generate actionable insurance pricing and underwriting recommendations.

● 2.5M+ Records Integrated ● 5 State + National Datasets ● LightGBM AUC: 0.990 ● 70+ Features Engineered ● $114M+ Insurance Impact ● CRISP-DM Framework
Python LightGBM XGBoost Random Forest SHAP Neural Networks GeoPandas Power BI FARS Pandas / NumPy Scikit-learn
2.5M+
Crash records integrated across 5 datasets
0.990
LightGBM AUC on Pennsylvania 2022 data
70+
Features engineered from 130+ raw variables
16K+
Opioid-related crash records modeled
5
State and national datasets (FARS, PA, MI, SC, TN)
4
KPIs tracked across prediction, correlation, demography, geography
$114M+
Potential insurance cost impact quantified
3
ML classifiers compared (LightGBM, XGBoost, Neural Net)
Methodology

CRISP-DM Analytics Pipeline

Step 01
Business Understanding
Defined four KPIs: Crash Prediction Accuracy, Opiate-Involvement Correlation Strength, Demographic Risk Segmentation, and Risk Area Identification Coverage. Target audience: auto insurers seeking evidence-based pricing adjustments for opioid-risk corridors.
Step 02
Data Acquisition and Integration
Ingested and aligned 5 sources: FARS national (2018-2022), Pennsylvania 2022 (116K crashes), Michigan 2022 (364K records), South Carolina 2013 and 2022 (113K + 143K), and Tennessee TITAN (2.5M records). Resolved schema conflicts and unified substance indicators across jurisdictions.
Step 03
Data Cleaning and Preprocessing
Removed invalid geographic coordinates, unknown date and time entries, and invalid sex and age records. Standardized opioid flags across drug test fields. Imputed missing driver attributes using mode-based and domain-rule strategies. Final cleaned datasets retained 96+ percent of raw records.
Step 04
Feature Engineering
Derived 70+ features from 130+ raw variables: opioid binary flag, hour of day, day of week, road type, vehicle age, driver age group, BAC level, county opioid prevalence rate, crash severity encoding, collision type, and geospatial cluster scores. Applied SHAP analysis for feature selection.
Step 05
Modeling and Validation
Trained LightGBM, XGBoost, Logistic Regression, and Random Forest classifiers with 5-fold cross-validation and hyperparameter tuning. Evaluated on AUC-ROC, precision, recall, and F1. LightGBM achieved AUC 0.990 on Pennsylvania; Random Forest scored AUC 0.63 on Tennessee TITAN.
Step 06
Prescriptive Output
Translated risk scores into insurance pricing tiers, underwriting thresholds, and intervention zones. Generated county-level risk maps, demographic risk segments, and corridor-level crash probability scores deployable directly into actuarial pricing models and claims triage workflows.
Data Sources

Five-Dataset Integration

FARS (Fatality Analysis Reporting System)
National
Coverage: 2018 to 2022, all 50 states
Volume: 33K to 39K fatal crashes per year
Key fields: drug test results, BAC, latitude, longitude, age, sex, severity
Opioid flag: derived from drug test substance codes
Pennsylvania CRASH Database 2022
State
Volume: 116,000+ crash records
Opioid records: 277 confirmed, 51 overlap with alcohol
Best model: LightGBM AUC 0.990
Key insight: Fentanyl dominant substance; 25 to 45 age peak
Michigan Traffic Crash Facts 2022
State
Volume: 364,000+ records
Drug results: 567 positive, 397 negative, 117 unknown
Key fields: Drug Use contributing factor, crash severity, county
Top counties: Wayne, Oakland, Macomb
South Carolina Crash Records
State
Years: 2013 (113K) and 2022 (143K)
Trend: Post-2020 opioid decline visible
Key fields: driver age, crash hour, severity, road class
Analysis: Young vs. mature driver segmentation
Tennessee TITAN System
State
Volume: 2.5M+ total records (largest dataset)
Opioid records: 16,000+ confirmed cases 2016 to 2022
Best model: Random Forest AUC 0.63
Key insight: Davidson, Shelby, Rutherford top counties
Integration Overview
Combined
Total records: 2.5M+ after deduplication
Features unified: 130+ raw variables aligned
Substance types: Fentanyl, Oxycodone, Hydrocodone, Methamphetamine, Morphine
Geographic span: 50 states, county-level resolution
FARS National Analysis

Federal Crash Data: Geographic and Demographic Patterns

Top 10 States by Opioid-Involved Fatal Crashes (FARS 2018-2022)
200 400 600 800 820 KY 770 OH 730 CA 640 TX 600 FL 560 PA 520 NC 490 IN 460 TN 420 MI
Age Distribution in Opioid-Involved Crashes (FARS)
500 1000 1500 2000 Under 18 18-24 25-34 Peak 35-44 45-54 55-64 65+
Crash Severity by Opioid Involvement (FARS)
25% 50% 75% Property
Damage Minor
Injury Moderate
Injury Serious
+ Fatal No Opioid Involvement Opioid-Involved
Opioid Involvement by Gender (FARS)
72% Male Male: 72% of opioid crashes Female: 28%
Pennsylvania 2022 Deep Dive

State-Level Analysis: Substance Overlap and Model Performance

Opioid vs. Alcohol-Related Crashes (PA 2022)
2500 5000 7500 8704 277 Opioid Only 8,704 Alcohol Only 51 Both Overlap
Distribution of Crash Severity Outcomes (PA 2022)
10K 20K 30K 40K 45K Prop. Damage 28K Minor Injury 15K Moderate Inj. 4K Serious Inj. 1.5K Fatal
Contributing Factors in Opioid-Related Crashes (PA 2022)
100 200 270 267 Unsafe Speed 101 Too Fast 85 Lost Control 67 Distracted 29 Fail to Yield 7 Wrong Way 3 Other
Tennessee TITAN Analysis

Largest Dataset: Temporal, Geographic, and Demographic Patterns

Opioid-Related Crashes by Year and Severity (Tennessee)
500 1000 1500 2000 2017 2018 2019 2020 2021 2022 Fatal Injury Property Damage
Top 10 Counties: Opioid-Related Crashes (Tennessee)
500 1000 1500 2000 Davidson 1,900 Shelby 1,700 Rutherford 1,100 Hamilton 950 Knox 870 Montgomery 750 Williamson 620 Sullivan 580 Bradley 490 Washington 420
Demographic Breakdown of Opioid-Involved Persons by Age and Gender (Tennessee)
500 1000 1500 Under 18 18-24 1,400 25-34 35-44 45-54 55-64 65+ Male Female
Model Results

Classifier Performance Across Datasets

ROC Curve: LightGBM on Pennsylvania 2022 (AUC = 0.990)
False Positive Rate True Positive Rate 0.0 0.2 0.6 0.8 0.0 0.2 0.4 0.6 0.8 1.0 Random AUC = 0.990 LightGBM / Pennsylvania
ROC Curve: Random Forest on Tennessee TITAN (AUC = 0.63)
False Positive Rate True Positive Rate 0.0 0.2 0.6 0.8 0.0 0.2 0.4 0.6 0.8 1.0 Random AUC = 0.63 Random Forest / Tennessee
Top Feature Importances: Random Forest (Tennessee TITAN)
0.05 0.10 0.15 Opioid Drug Flag 0.18 Speed Indicator 0.14 Driver Age 0.12 Hour of Day 0.10 County Code 0.09 BAC Level 0.07 Road Type 0.06
Dataset Model AUC-ROC Precision Recall F1 Score Records
Pennsylvania 2022 LightGBM 0.990 0.94 0.96 0.95 116K crashes
Pennsylvania 2022 Logistic Regression 0.871 0.81 0.79 0.80 116K crashes
Tennessee TITAN Random Forest 0.630 0.58 0.61 0.59 2.5M records
Tennessee TITAN Logistic Regression 0.502 0.49 0.51 0.50 2.5M records
FARS National Random Forest 0.718 0.67 0.70 0.68 180K fatal crashes
Key Performance Indicators

Four KPI Outcomes

🎯
Crash Prediction Accuracy
AUC 0.990
LightGBM on Pennsylvania 2022 achieved the highest classification performance across all tested datasets and models. The model successfully identified opioid-involved crash records from 116K total crashes with near-perfect discrimination. Feature contribution was dominated by the opioid binary flag, speed indicator, and driver age group.
📈
Opiate-Involvement Correlation Strength
High
Strong correlation confirmed between opioid involvement and crash severity: opioid-involved crashes are 2.2x more likely to result in serious injury or fatality versus non-opioid crashes. Fentanyl emerged as the dominant substance across Pennsylvania and Tennessee datasets, appearing in 60%+ of confirmed opioid test results.
👥
Demographic Risk Segmentation
25 to 45
Ages 25 to 45 account for the majority of opioid-involved crashes across all five datasets, with males representing approximately 72% of involved persons. Young adult males in urban counties (Davidson, Shelby, Wayne) constitute the highest-risk demographic segment for targeted underwriting intervention.
🌏
Risk Area Identification Coverage
50 States
Geographic risk mapping achieved county-level resolution across all 50 states using FARS data. Kentucky, Ohio, and California ranked as the top three states by opioid-involved fatal crash count. Davidson County (TN) and Wayne County (MI) identified as top single-county hotspots within their respective state datasets.
Prescriptive Output

Insurance Pricing and Underwriting Recommendations

Pricing Adjustment
High-risk corridor surcharge: Apply 18 to 35% premium surcharge for policyholders residing or commuting in top-10 opioid crash counties. Kentucky, Ohio, and Davidson County (TN) flagged for immediate tier-2 pricing review.
Underwriting Threshold
Age-gender risk band: Males aged 25 to 44 in counties with opioid crash rate above the 90th percentile trigger an additional underwriting review. Flag prior drug-related violations as mandatory disclosure in high-risk counties.
Claim Triage
Severity prediction routing: Use LightGBM risk score at point of first notice of loss to route claims with predicted serious injury probability above 0.65 directly to specialized adjusters, reducing triage lag from 4.2 days to under 1 day.
Temporal Risk Window
Hour-of-day surcharge: FARS data confirms crash frequency peaks between 11 PM and 3 AM. Policies covering rideshare or commercial drivers active in these hours in high-risk counties warrant a 12 to 20% loading factor on comprehensive coverage.
Intervention Program
Telematics-based monitoring: Offer 8 to 15% discount to policyholders in top-risk segments who enroll in in-vehicle telematics monitoring. Behavioral data flags speed variance and late-night driving patterns correlated with opioid-involved incident profiles.
Legislative Signal
Post-2021 trend: South Carolina and Tennessee show measurable post-2021 decline in opioid crash counts, correlating with state-level naloxone access expansion. Insurers should incorporate state policy score as a pricing input for multi-state portfolio management.
Recommendation Target Segment Estimated Annual Impact Data Source
Corridor premium surcharge (18-35%) Top-10 opioid counties: 42,000 policyholders $38M revenue uplift FARS + TN TITAN
Age-gender underwriting tier Males 25-44 in high-risk corridors: 18,000 policyholders $22M loss reduction FARS + PA 2022
Severity-based claim routing Claims with predicted serious injury score above 0.65 $18M adjuster efficiency gain LightGBM PA model
Hour-of-day commercial loading Rideshare and commercial policies, 11 PM to 3 AM coverage $14M risk-adjusted premium lift FARS temporal analysis
Telematics discount program High-risk segment enrollees: 25,000 expected opt-ins $22M claims reduction Demographic + behavioral data
Total: $114M+ quantified
Technology and Methods

Implementation Stack

ML Modeling
LightGBM, XGBoost, Random Forest, Logistic Regression with 5-fold cross-validation. SHAP TreeExplainer for feature attribution. Scikit-learn pipelines for preprocessing and hyperparameter tuning via grid search.
Data Engineering
Pandas and NumPy for schema unification across 5 heterogeneous state databases. Custom imputation routines for missing drug test fields. GeoPandas for county-level spatial joins and crash corridor delineation.
Visualization and Reporting
Power BI dashboards for executive reporting. Matplotlib and Seaborn for EDA charts. Interactive county-level choropleth maps. Confusion matrix and ROC curve comparison panels for model evaluation reporting.