Predicting Adolescent Drug Use with Machine Learning

Author: Jamie Dowat

Contents

• src contains source code to run the final_notebook.ipynb.
• final_notebook.ipynb contains a detailed description and demonstration of the modeling process for this project, as well as the full list of source material (modeling and otherwise).
• capstone_presentation.pdf is a slide deck containing major highlights of the project.

---

Business Problem

The drug problem in the United States is widespread and growing:

Mortality rates from drug use disorders increased nationally and in every county between 1980 and 2014, but the relative increase varied from 8.2% to 8369.7%.

---

Adolescents are especially vulnerable. After a year since first use of drugs: * 11.2% of adolescents had prescription opioid use disorder versus 6.9% of young adults. * 13.9% of adolescents had prescription stimulant use disorder versus 3.9% of young adults. * 11.2% of adolescents had prescription tranquilizer use disorder versus 4.7% of young adults.

---

Purpose of Project:

To develop/tune machine learning classifiers in the hopes of more deeply understanding factors that correlate with drug use.

For potential deployement the final model from this project has potential to serve as a "barometer" for high schools, who can implement the survey anonymously and gain some high-level data about the state of their school and be able to more precisely analyze trends amongst their students and measure their prevention efforts against any changes seen in the model.

DATA: Monitoring the Future

Monitoring the Future began in 1975 as an attempt to collect and study drug use and feelings toward drug use amongst adolescents (8th-12th).

Data consists of 101,658 survey reponses from 12th graders from 133 schools nationwide.

For more information on this study, click here.

To download the dataset used in this project, click here.

Questions in this dataset are limited to the following:

• Demographic information
• Drug Use Information

---

MODELING:

The following models were implemented, along with Boruta feature selection, Phi-K and Bayesian Optimization.

• Logistic Regression
• Decision Trees
• Random Forests
• XGBoost

Models were validated with StratifiedKFolds, dividing the training set into 10 folds.

The final model produced an accuracy of 80.6% and a recall of 72% (Test Set)

Recall was given extra attention as a metric, since false negatives (predicting no drug use when drug use is present) are predictions that would potentially cause the most harm.

Feature Importances:

In terms of real-life applicability, Logistic Regression ended up being the best choice due to its superior performance and high level of interpretability.

TOP 3 Postive Features:

• YEAR=2019 - 1.122
• YEAR=2015 - 0.714
• Consumed Alcohol 40+ times in the past year - 0.655

TOP 3 Negative Features:

• Consumed NO Alcohol in the past year - -1.044
• Never Smoked Cigarette - -1
• Grew up on a Farm - -0.577

Final Model: Confusion Matrix

---

├── img
....├── amenbrain.png
....├── fsm.png
....├── mtf.png
....├── nida.png
....├── samhsa.png
....└── stopdrugs.png
├── src
....├── __init__.py
....├── eda.py
....└── modeling.py
├── capstone_presentation.pdf
├── environment.yml
├── final_notebook.ipynb
└── README.md