A Decision Tree Analysis on Drug Use and Health
Published:
ABSTRACT This report contains the analysis of factors that are associated with youth drug use, using decision tree models on survey data from the National Survey on Drug Use and Health. The dataset contains detailed information on various aspects of respondents’ lives, including demographics, youth experiences, and substantial drug use. The study explores three problem types: binary classification for marijuana use, multi-class classification for frequency of marijuana in past year, and regression for first ever use of marijuana. Decision trees and some ensemble methods are used to build predictive models for each problem type. The results suggest the impact of different demographic variables and youth experience variables on various substantial youth drug uses. The findings of this study have practical implications for public health interventions aimed at reducing youth drug use.
INTRODUCTION Our research aims to predict the determinants of youth drug use using decision trees and ensemble methods drawn from the National Survey on Drug Use and Health. The survey is demographically representative of non-institutionalized populations in the US; it offers a wealth of information regarding drug use patterns alongside pertinent details like age, gender, race/ethnicity mental health conditions. For close to five decades, SAMHSA has sponsored this annual study that features an aggregated dataset composed of approximately 300 variables cataloging around 70K respondents aged twelve or more.
The study at hand will concentrate on three variables associated with marijuana use, namely ‘mrjflag’ which denotes the respondent’s history of using marijuana; ‘mrjydays,’ indicating the number of days that the respondent used marijuana in the previous year, and ‘irmjage’ signifying the first age when they indulged in cannabis usage. The goal is to employ decision trees to predict whether an individual has ever consumed marijuana (binary classification), how many days they indulged in cannabis use during the last year (multi-class classification), and ultimately assess their initial age for consuming marijuana (regression).
Our analytical approach will encompass more than just decision trees. We are planning to employ ensemble methods such as random forests and boosting to increase our models’ accuracy. In this study, we intend to use decision trees and ensemble methods to analyze the dataset comprehensively, It is our aim to identify factors associated with youth drug use and provide insights that can be used to inform public health policies that aim to reduce teen drug use.
METHODOLOGY Data Cleaning: The dataset was filtered during the data cleaning process to remove observations with invalid or missing values. To be more specific, the dataset was filtered using the following criteria: 1.The values of 98 or 99 in the variable EDUSCHGRD2 were removed from the dataset. 2.The values of 94, 97, 98, or 99 in the variable eduskpcom were removed. 3.The observations with values of 3 or 4 in the variables imother and ifather were removed. 4.The observations with a value of 3 in the variable PDEN10 were removed. 5.Finally, any observations with missing values in the columns related to youth experience were removed.
Upon completion of the data cleaning process, the dataset underwent further analysis to identify three subsets categorized for different analysis techniques- binary classification, multiclass classification, and regression analysis. These segregated data sets were further used to create decision trees and ensemble models utilizing the relevant classification or regression approach.
Binary classification: In the binary classification, the “mrjflag” that is marijuana ever used (0=never, 1=ever) column is used as the response variable, while only the data from demographic and youth experience columns are used as predictor variables. The data is split into training and testing sets, and a decision tree is created using the “tree” function. Later using the testing data and the “predict” function the accuracy of the tree model is evaluated. I tried to then pruned the decision tree using cross-validation with the “cv.tree” function used best = 3 and eventually got exactly the same result as a normal decision tree with same accuracy after that used the the “importance” function to compute the important variable.
Multiclass classification: In the multi-class classification, the “mrjydays” that is the number of days of marijuana in past year column is used as the response variable, which is converted to a factor variable. The same predictor variables are used as in the binary classification (demographic and youth experience). The data is split into training and testing sets again, and a decision tree is created using the “tree” function. The variable importance is computed using the “importance” function. and finally, The accuracy of the tree model is evaluated using the testing data and the “predict” function. The variable importance is computed using the “importance” function.
Regression: In the regression classification, the “irmjage” marijuana age of first use column is the targeted variable, while only the data from demographic and youth experience columns are used as predictor variables same as binary and multi-class classifications. Data is cleaned where we used an observation 991 where marijuana is never used also checked for if there are any missing values. The data is split into training and testing sets, and a decision tree is created using the “tree” function. I tried to perform pruning but ended up getting the same result as normal decision tree so, three regression methods used are decision tree, bagging and random forest. The mean squared error is calculated for each method for the comparison of models. the “varImpPlot” function is used to plot the variable importance.
COMPUTATIONAL RESULTS
- BINARY CLASSIFICATION
Accuracy values of different models used for binary classification are as follows: • Decision Tree - 0.8811777 • Prune – 0.8811777 • Bagging – 0.8895899 • Random Forest – 0.8885384
According to the accuracy of the models, the best model to predict if marijuana is ever used or not is bagging with an accuracy of 0.8895899.
- MULTI-CLASS CLASSIFICATION
Accuracy values of different models used for multi-class classification are as follows: • Decision Tree – 0.8790747 • Prune – 0.8790747 • Bagging – 0.873817 • Random Forest – 0.8706625
According to the accuracy of the models, the best model to predict number of days marijuana was used in past year are Decision tree & prune as both of them are giving same accuracy of 0.8790747.
- REGRESSION
MSE values of different models used for multi-class classification are as follows: • Decision Tree – 2.076836 • Prune – 1.702468 • Bagging – 1.831604 • Random Forest – 1.616788
According to the MSE of the models, the best model to predict age that marijuana was used for the first time is random forest giving the least MSE of 1.616788.
DISCUSSION
- BINARY CLASSIFICATION
The first question being asked at the root node is whether the respondent approves or disapproves of marijuana use. If they strongly or somewhat disapprove, the algorithm moves down the left branch, and if they neither approve nor disapprove, it moves down the right branch.
At its left node, the algorithm is asking whether any of the students in the respondent’s grade take marijuana. If most or all of them do not take marijuana, it moves down to it’s left branch, and if most or all of them do take marijuana, it moves down to its right branch.
At its left node, the algorithm is asking how their close friends feel about youth taking marijuana. If they strongly or somewhat disapprove of taking marijuana, then the youth are more likely to never used marijuana.
BAGGING:
BAGGING is my best model according to the accuracies.
According to the confusion matrix, 782 of youth who never used marijuana are classified correctly and 64 youth who uses marijuana are predicted correctly. 77 people who uses marijuana are misclassified as never used and 28 people that doesn’t use marijuana are misclassified as used. According to the plot, the most important variables that can predict if a person ever used marijuana or not are yflmjmo (how yth feels: peers using marijuana monthly), FRDMEVR2(rc-yth think: close frnds feel abt yth try marijuana), stndsmj(number of students in yth grade use marijuana), frdmjmon(rc-yth think: clse frnds feel abt yth use marijuana mon), YFLTMRJ2(rc-how yth feels: peers try marijuana).
- MULTI – CLASS CLASSIFICATION:
Example decision tree:
According to the accuracy of the models, the best model to predict number of days marijuana was used in past year are Decision tree & pruning.
1.DECISION TREE:
The confusion matrix shows the number of predicted values that match or do not match the actual values in the test dataset for a multi-class classification problem. The rows of the matrix represent the predicted values, and the columns represent the actual values. In this ![image]
case, the matrix is a 6x6 table, where the predicted values range from 1 to 6 classes and correspond to the frequency of marijuana use in the past year. The diagonal elements of the matrix represent the number of correct predictions, while the off-diagonal elements represent the number of incorrect predictions.
Looking at the matrix, we can see that the model correctly predicted the non-user category (level 6) for 823 cases. However, it performed poorly in predicting the other levels, particularly level 4 where it correctly predicted only 13 out of 24 cases. The model did not predict any cases in levels 1, 2, 3 or 5.
The overall accuracy of the model is 0.879, which means that it correctly classified 87.9% of the cases in the test data. However, the low performance in predicting the levels of marijuana use other than non-users suggests that the model may not be reliable for predicting levels of marijuana use.
- PRUNING:
The confusion matrix is same for pruning and decision tree. These both are best models in terms of accuracy. But let’s consider bagging to see it’s confusion matrix.
824 cases of level 6 are classified correctly and some cases in classes 1,2,3 are classified correctly. So, I think I prefer bagging.
According to the above plot, the most important variables that can be helpful in classifying the number of days of marijuana use in past year are PRMJEVR2(rc-yth think: parents feel abt yth try marijuana), FRDMEVR2(rc-yth think: close frnds feel abt yth try marijuana), YOSELL2(RC-youth sold illegal drugs), frdmjmon(rc-yth think: clse frnds feel abt yth use marijuana mon), YFLTMRJ2(rc-how yth feels: peers try marijuana).
3.REGRESSION: Example decision tree:
According to the MSE of the models, the best model to predict age that marijuana was used for the first time is random forest.
1.RANDOM FOREST:
According to the above plot, the top variables that are used to predict the age of first use of marijuana is EDUSCHGRD2(what grade in now/will be in), YTHACT2(rc-yth participated in youth activities), PRGDJOB2(rc-parents tell yth had done good job in pst yr), PRPROUD2(rc-parents tell yth proud of things done in pst yr) and GRPCNSL2(rc-participated in program to help substance use).
CONCLUSION
This study has examined the factors associated with youth drug use using decision tree models and ensemble methods applied to data from the National Survey on Drug Use and Health. The findings of this study have practical implications for public health interventions aimed at reducing youth drug use.
The results showed that certain demographic and youth experience variables are significant predictors of youth drug use, especially for marijuana use. The decision tree models produced accurate predictions for binary classification of marijuana use, multi-class classification for frequency of marijuana use in the past year, and regression for first-ever use of marijuana. The ensemble methods, including bagging and random forests, improved the accuracy of the models, which indicates the importance of using multiple techniques to create models for predicting substance use among young people.
or binary classification, the bagging model had the highest accuracy of 0.889, indicating that it is the best model to predict if a person has ever used marijuana or not. The decision tree and pruning models had the highest accuracy of 0.879 for multi-class classification, making them the best models to predict the number of days of marijuana use in the past year. But for better classification without misclassifications towards majority class, we can consider bagging. For regression, the random forest model had the lowest MSE of 1.617, indicating that it is the best model to predict the age at which a person first used marijuana.
Based on the confusion matrices and feature importance plots, we have identified the variables that are most important in predicting the use of marijuana. For binary classification, the most important variables were how youth feels about their peers using marijuana monthly, how close friends feel about youth trying marijuana, and how students in youth grade use marijuana. For multi-class classification, the most important variables were how youth thinks parents feel about youth trying marijuana, how close friends feel about youth trying marijuana, and whether youth has ever tried smoking cigarettes. For regression, the most important variables were how youth perceives risk in using marijuana, how many days a week youth exercises, and how many days in the past week youth felt depressed.
Overall, our models provide insights into the factors that influence the use of marijuana among youth and can help policymakers and healthcare professionals to identify and address the underlying risk factors.
BIBLIOGRAPHY
Substance Abuse and Mental Health Services Administration. National Survey on Drug Use and Health. Retrieved from https://www.samhsa.gov/data/data-we-collect/nsduh-national-survey-drug-use-and-health
Substance Abuse and Mental Health Services Administration. NSDUH Codebook: Survey Questions, Responses, and Coding. Retrieved from https://www.samhsa.gov/data/report/nsduh-codebook-detailed-instructions-using-2018-nsduh-public-use-file