Introduction

The use of hard drugs remains a public health concern in the United States and globally, despite improved education and outreach efforts. Between 2005 and 2015, nearly half of all U.S. prisoners were imprisoned due to drug-related crimes (National Center for Drug Abuse Statistics, 2024), and substance use disorders (SUDs) comprise a source of significant social and economic burden on not only patients, but also on their families and on society (Daley, 2013). Understanding the relationship between drug use, particularly hard drug use, and various demographic and personal features allows national and international public health agencies to prioritize their education and outreach programs in communities where these programs may have the highest impact.

Previous work has revealed relationships between individual socioeconomic status and drug use (Patrick et al., 2012). However, socioeconomic status only provides one view into how an individual’s circumstances are associated with their likelihood of drug use. We reasoned that an individual’s personality, which is likely to also be influenced by the individual’s environment, especially during childhood, may be predictive of drug use. Using personality measurements as a predictor of drug use can provide public health experts additional information about the interplay between an individual’s circumstances and their likelihood of drug use independent of the known relationship between drug use and socioeconomic status.

To this end, we used a publicly available dataset sourced from an online survey of individuals’ drug use habits for a variety of drugs, in addition to measurements of their personality traits along several axes: extraversion, openness, neuroticism, impulsivity, and sensation-seeking (Fehrman et al., 2017). This dataset provides an opportunity to investigate the relationship between these personality features and individuals’ drug use. Further, because the dataset also contains demographic information, we were interested in whether or not the potential personality-drug use relationships changed by the individual’s demographic background. As an example, we felt that it could be possible for high-extraversion individuals of a particular demographic background (for example, college-educated individuals) to display the opposite drug usage patterns compared to high-extraversion individuals of a different background, due to varying cultural norms.

Thus, our analysis of this dataset aims to answer four main research questions:

  1. Are the personality variables measured in this dataset predictive of drug use? If so, which variables are associated with drug use, and how are they associated?
  2. Is there a significant difference between education level for the usage of hard drugs and its association with personality?
  3. Is there a significant difference between genders for the usage of hard drugs and its association with personality?
  4. Is there a significant difference between different ethnicities for the usage of hard drugs and its association with personality?

Description of the Dataset

Our dataset contains information on drug consumption, legal and illegal, for 1885 respondents of varying levels of education, ethnicity, age, and gender. Each row in the dataset corresponds with one respondent. The columns in the dataset fall into two main categories: information about the respondent and information about the respondent’s drug use. The information about the respondent includes the following attributes: age, gender level of education, country of origin, ethnicity, neuroticism score, extraversion score, openness to experience score, agreeableness score, conscientiousness score, impulsiveness score, and sensation seeking score. The drug consumption columns having the following scale:

The drug consumption columns themselves are as follows: alcohol consumption, amphetamines consumption, nitrite consumption, benzodiazepine consumption, caffeine consumption, marijuana consumption, chocolate consumption, cocaine consumption, crack cocaine consumption, ecstasy consumption, heroin consumption, ketamine consumption, legal highs consumption, LSD consumption, methadone consumption, magic mushroom consumption, nicotine consumption, consumption of the fictitious drug Semeron as a control, and volatile substance abuse consumption. The dataset is found at the following link: https://www.kaggle.com/datasets/obeykhadija/drug-consumptions-uci.

Results

Research Question 1: Are the personality variables predictive of drug use, and if so, which one(s)?

First, we wished to assess whether or not the personality variables were predictive of drug use. We reasoned that projecting each individual onto a reduced-dimensionality space on the basis of their personality variable measurements, then visualizing all individuals by their drug use status, would allow us to discern whether or not these personality variables as a whole were associated with drug use. If there was no relationship, then individuals who used and did not use drugs would be randomly scattered in the space, and if there was a relationship, then there would be a clearer pattern.

However, defining who counted as a “drug user” was a challenge. We chose to focus on users of “hard drugs,” which are all drugs in the dataset excluding chocolate, alcohol, and caffeine. These “hard drugs” include cocaine, methamphetamine, and others, serving our goal of separating more psychoactive substances from less psychoactive substances. Survey respondents were asked about their frequency of using each drug, from having never used the drug, having used the drug more than a decade ago, up to having used the drug within the past day. For the purposes of this research question, we binarized the response based on whether or not that individual had ever used a hard drug. We also removed all individuals who said that they had ever used “Semeron,” a fictitious drug introduced in the survey for the purposes of identifying over-claiming or low-quality survey responses.

Using this modified data, we chose to perform multidimensional scaling (MDS) on the measured personality variables down to two variables, since MDS would seek to preserve the inter-individual distances within this personality space. We then colored each individual by whether they have ever used a hard drug (Fig. 1).

Fig. 1: MDS reduction suggests that personality variables are associated with hard drug usage. The continuous personality variable measurements for each variable were transformed via multidimensional scaling (MDS), and the two MDS axes (MDS 1 and MDS 2) are plotted here. Each point represents an individual’s response, colored by whether or not that individual has ever used a hard drug. Lower MDS 1 is associated with a higher probability of ever using a hard drug.

While the second MDS dimension (MDS 2) did not appear associated with drug use, MDS 1 seems to be predictive of hard drug use. Lower MDS 1 values tend to correlate with a higher probability of ever having used a hard drug, while higher MDS 1 values tend to correlate with a lower probability of hard drug use. As a result, our MDS analysis (Fig. 1) suggests that these personality variables do have a relationship with hard drug use.

We then wondered which specific personality variables were associated with hard drug use. Using the same MDS reduction, we plotted two personality variables, openness and impulsivity, rather than hard drug usage (Fig. 2).

Fig. 2: Increased openness and impulsivity are associated with lower MDS 1, which is associated with a higher probability of using hard drugs. Using the same MDS reduction in Fig. 1, we colored by individuals’ openness and impulsivity measurements, finding that both of these personality variables were associated with lower MDS 1. Given the negative association between MDS 1 and hard drug usage in Fig. 1, there may be a correlation between openness, impulsivity, and hard drug usage.

We identified fairly robust associations between both openness and MDS 1 and impulsivity and MDS 1. In both cases, higher openness or higher impulsivity was associated with lower MDS 1, which itself was associated with higher drug use, leading us to believe that there may be a relationship between increased openness, increased impulsivity, and the probability that an individual has ever used a hard drug. Based on similar MDS analysis, other personality variables were also associated with hard drug use, such as sensation-seeking (MDS plot not shown due to restrictions on number of similar plots, but please see the analysis in the section Research Question 2).

While our MDS reduction in Fig. 1 had suggested that MDS 1 is (anti-)correlated with drug use, we wondered if we could identify clusters of individuals with similar personality variable measurements who all were hard drug users, or who all were not hard drug users. Our MDS reduction was very noisy, precluding visual identification of such clusters. We therefore used an agglomerative clustering algorithm to perform hierarchical clustering of all individuals based on all of their personality variables. We then chose to divide the dendrogram into two clusters, since we were primarily interested in the use (or not) of hard drugs (Fig. 3).

Fig. 3: Hierarchical clustering dendrogram of individuals by their personality variable measurements; leaves colored by drug use. We performed agglomerative hierarchical clustering using Euclidean distance and complete linkage on individuals’ personality variable measurements, then colored individuals (leaves, bottom) by their hard drug use (black, never used a hard drug; red, used at least one hard drug at some point). Since we were interested in binary hard drug use, we separated the responses into two clusters, leading to one cluster (left) with lower hard drug use probability than the other (right).

Broadly, the clustering results do suggest that we are able to identify two major clusters, one in which there is a lower probability of hard drug use (left cluster, fewer individuals who have used hard drugs), and one in which there is a higher probability of hard drug use (right cluster).

Taken together, these results suggest that the personality variables, as a whole, are predictive of hard drug use. MDS reductions of the personality variables, as well as hierarchical clustering results, identify subpopulations of the data with increased likelihood of having ever used a hard drug. More specifically, increased openness and increased impulsivity seem to be two major drivers of this relationship, with both increased openness and increased impulsivity likely being associated with increased hard drug usage.

Research Question 2: Is there a significant difference between education level for the usage of hard drugs and its association with personality?

After looking at how ethnicity may play a role in the association between drug usage and education, we now look at how education levels may play a role in that association. We will first create a scatterplot of personalities vs total drug usage (addition of all drug ratings), representing the number of drugs tried per individual. This scatter plot will be accompanied with a LOESS regression line to see the general relationship of personality and total drug usage score by education level.

Fig. 4: LOESS regression of total drug use score by personality scores.

Each graph represents the relationship between total drug usage score and personality scores with education level differentiated by color. Along with each regression line representing the trend of different education levels, the dotted black line displays the general trend as it includes all observations from all education levels.

Oscore(Openness to experience), Impulsive, and SS(sensation seeking) displays a strong positive trend for most education levels. In general, individuals with more openness to experience, stronger impulsivity, and seeking more sensation tend to have a higher level of drug usage. However, there are outliers. For the Oscore plot, those who left school at 18 and before 16 have highest drug usage scores between 0 to 1 then slowly decrease. For Impulsive plot, the slope is different for each type of group where PhD and those who left school before 16 exhibit high drug usage scores for lowest level of impulsiveness, which initially decreases then continuously increases again. On the other hand, other groups logarithmically or exponentially increase. Overall, when excluding some outliers, openness to experience, impulsiveness, and sensation seeking display a positive trend consistently across different educational levels.

Ascore(Agreeableness) shows exponential decay for all education groups, indicating individuals with higher agreeableness tend to have lower drug usage (except for those that left before 16 where they show weak increase).

For Escore, groups with higher levels of education(PhD, Masters, Bachelor) and low extraversion tend to have lower scores for drug usage, along with slower increase in drug usage when extraversion is higher. Other groups display parabolic relationships with those with average extraversion scores have the lowest drug usage while those with extreme scores (very extraverted or very introverted) have high drug usage. This can be seen in Cscore (consciousness) Groups with higher levels of education(PhD, Masters, Bachelor) have exponential decaying relationships while other groups have fluctuating trend (generally increasing level of drug usage from -2 to 0, decreasing level from 0 to 1, and increasing after 1). Similarly, for Nscore (neuroticism), higher education groups trend to have flatter trend, displaying small increase as Nscore increases. On the other hand, lower groups have more volatile trends.

In order to look more closely at the differences of association between personalities and drug usage by education levels, we have derived a nonlinear regression model. We have chosen to use the General Additive Model to model the relationship. Due to LOESS regression’s complexity, it was computationally impractical to include all variables while splines and other polynomial or logistic regression similarly cannot the complexity of our data due to various variables (furthermore, from GCV and AIC, GAM provided the smallest value).

## 
## Method: GCV   Optimizer: magic
## Smoothing parameter selection converged after 7 iterations.
## The RMS GCV score gradient at convergence was 0.0003545908 .
## The Hessian was positive definite.
## Model rank =  94 / 94 
## 
## Basis dimension (k) checking results. Low p-value (k-index<1) may
## indicate that k is too low, especially if edf is close to k'.
## 
##                 k'   edf k-index p-value    
## s(Nscore)    11.00  2.45    0.93  <2e-16 ***
## s(Escore)    11.00  6.55    0.89  <2e-16 ***
## s(Oscore)    11.00  2.76    0.88  <2e-16 ***
## s(AScore)    11.00  1.96    0.94    0.01 ** 
## s(Cscore)    11.00  3.01    0.91  <2e-16 ***
## s(Impulsive)  9.00  1.52    0.92  <2e-16 ***
## s(SS)         9.00  3.24    0.89  <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Fig. 5: GAM model fitting diagnostic plots (top) and diagnostic output (bottom).

From the fitting output (bottom of Fig 5), we have used gam.check() in order to check the aptness of our model. The Model rank is 94/94, showing that all effective parameters have been used in this model. K’ represents the number of basis functions while edf represents the degree of non-linearity of the curve (1EDFk’) which is true for all smoothed terms). Although k(basis dimension) is low, we have increased its value yet still exists due to the low unique values for SS and Impulsive (10~11 values). Furthermore, k-index is generally near 1, meaning k value is sufficient enough.

The QQ plot shows that the points are generally aligned around the expected distribution(red line) with few outliers at each end, showing that the data is approximately normally distributed. For the Residual vs linear predictor plot, the residuals are generally scattered randomly, with a slight shape of a funnel, showing possible violation of homoscedasticity. However, transformations of our variables complicate our interpretation, especially when extracting the spline functions to compare individual predictions of drug score for each personality score, and our focus is not on standard error. For the histogram of residuals, it is normally distributed with a slight skew to the left, showing no significant violations. For the response vs fitted value plot, it shows general positive linear trend and some significant variance.

Fig. 6: Smoothing functions of the GAM model for each personality score.

When graphing the GAM model, each graph represents the partial effect of each personality’s smoothed term. Nscore(neuroticism), Cscore(conscientiousness), and Impulsiveness have flat lines, showing weak nonlinear relationship on the total drug usage score. Escore(extraversion) and Ascore(Agreeableness) shows weak decreasing nonlinear relationships. On the other hand, SS(sensation seeking), Oscore(open to experience) shows strong increasing nonlinear relationships. In order to specifically see the difference in nonlinear effects across education levels for each personality, we have extracted the smoothed term for each personality and created a linear model in the form of:

total drug usage = B0+ B1(s_personality) + B2(Education) + B3 s_personality:Education

S_personality represents the smoothed term of personality score (estimated contribution of the personality to determining total drug usage from the original GAM model) and s_personality:Education represents the interaction variable.

Then, we collect the new total drug usage data from the new equation, and plot the slope between new total drug usage and personality score for each education level. This enables us to see if there are certain subgroups that are more reactive to the smoothed terms than others.

Fig. 7: Drug useage by smoothed personality scores, split by education level.

For SS(sensation seeking), Oscore(openness to experience), and Impulsiveness plots, all education groups display similar steepness of their slopes. Higher scores for sensation seeking, openness to experience, and impulsiveness increases the drug score. Groups that are in college and does not have a diploma and those who left school before 16 show steeper slopes, showing more reactiveness of the personality scores.

For Cscore (Conscientiousness) and Nscore(Neuroticism) plots, all groups show positive linear trends but with different levels of steepness in their slopes. For Cscore, Bachelor, Masters, Left Before 16, and some colleges have similar slopes while those who left at 16 and 18 have steeper levels. On the other hand, for Nscore, those who left school at 18,17, before 16 show steeper slopes while other groups have significantly lower gradients. In both of these groups, those whole do not have a diploma but went to college have the highest drug score while those who have bachelors, masters, or PhDs have the lowest levels.

For Ascore(agreeableness) and Escore(extraversion) plots, those who left school before 16 have negative slopes while others have positive slopes (Masters group have negative slope for Escore). For Ascore, those who left school at 17 and 18 show steeper slopes for both plots.

In summary, the effect of sensation seeking, openness to experience, and impulsiveness do not differ significantly across education groups. However, those who left school before 18 show a greater reactiveness in other personality scores than those who completed higher education.

Research Question 3: Is there a significant difference between genders for the usage of hard drugs and its association with personality?

Now that we have explored whether or not education plays a role in the usage of hard drugs and its association with personality, we wish to explore whether there are any differences between how two genders (male and female) use hard drugs based on their personalities. To do so, we will first look at the distribution of hard drug use for each gender using a stacked bar chart. This will show us at a high level if there are any differences in hard drug use between the genders. Then, we will create boxplots, one for each personality variable, and facet based on gender to look at the distribution for hard drug use versus non hard drug use. This will allow us to compare drug use by personality variable for each gender and identify potential differences between males and females.

Fig. 8: Stacked bar plot of hard drug use by gender.

From this figure, we can see that many more respondents in the dataset admit to using hard drugs than those who do not, which is an interesting observation and suggests that the dataset may not represent the general population or may be biased in some other way due to how the data was collected. In the stacked bar chart, red represents females and blue represents males. For the participants who used hard drugs, we see a pretty even split between males and females, with slightly more males than females claiming to have used hard drugs at least once in their lifetimes. For those who did not use hard drugs, there are more females than males; there appear to be at least twice as many females who did not use hard drugs as males who did not. To test whether or not drug use systematically differs by sex, we conducted a two-sided test for proportions to determine whether or not there was a statistically significant difference in the proportion of hard drug users among males and females.

## 
##  2-sample test for equality of proportions with continuity correction
## 
## data:  c(female.drug.use, male.drug.use) out of c(n.female, n.male)
## X-squared = 68.759, df = 1, p-value < 2.2e-16
## alternative hypothesis: two.sided
## 95 percent confidence interval:
##  -0.1750481 -0.1077669
## sample estimates:
##    prop 1    prop 2 
## 0.7692308 0.9106383

Fig. 9: Results from the two-sided proportions test between female and male drug use rates.

Having chosen our significance threshold at 0.05, we found a statistically significant difference in the proportion of female hard drug users compared to the proportion for males, at a p-value of less than 2.2e-16. Because this p-value is much smaller than our significance threshold, we conclude that there is a statistically significant difference between the proportion of females who use drugs (~77%) and the proportion of males who use drugs (91%), and that these data are too unlikely under the null hypothesis (that the proportions are the same) to occur by chance.

Now, let us explore if there are similar or different trends between male and female hard drug use based on personality.

Fig. 10. Box plots of personality measurements by hard drug use, facetted by sex.

In these boxplots, red represents never having used hard drugs and blue represents having used hard drugs. Overall, there seem to be similar distributions for hard drug use and non hard drug use across the different personality variables for males and females. For impulsivity, in general, hard drug users show higher median scores compared to non hard drug users, and this difference is slightly more pronounced in males than in females. Neuroticism shows higher scores for hard drug users compared to non hard drug users, and the variance in scores appears to be larger for females compared to males. For openness, hard drug users tend to have higher scores than non hard drug users, which is relatively consistent across both genders. Conscientiousness shows slightly higher scores for non hard drug users than hard drug users; this difference is more evident in females than in males. Neither extraversion nor agreeableness show any significant differences between hard drug users and non hard drug users for either gender, and the medians and distributions overlap. Altogether, for females, the differences between non hard drug users and hard drug users seem to be more evident in neuroticism and conscientiousness. On the other hand, for males, these differences appear to be more pronounced in impulsivity and openness. In summary, both genders show differences in personality traits between hard drug users and non hard drug users, but the magnitude of these differences varies by gender.

Research Question 4: Are there significant differences in how hard drug use is associated with different personality traits across different ethnic groups?

Now, we will move on to understand whether ethnicity plays a role in hard drug use and its association with personality. To understand the role of ethnicity, we will first make a stacked bar graph which highlights the distribution of hard drug use by ethnicity. To further understand whether different ethnicities’ use of hard drugs is based on their personalities, we will create boxplots. There are different plots for each personality type, each of which showcases the distribution of hard drug users belonging to different ethnicities.

Fig. 11: Number of hard drug users by ethnicity in the dataset.

The graph above examines the distribution of hard drug use by ethnicity, showing that white people make up the majority of the proportion of hard drug users in the dataset (and that, in general, the majority of respondents were white). Additionally, this graph also shows that white individuals also outweigh the other ethnic groups in the dataset, but as compared to the other ethnicities white individuals have a significantly higher rate of hard drug use. From this graph we can conclude that there is a potential link between ethnicity and the likelihood of using hard drugs, with white individuals being the most likely. To understand, the relationship involving personality traits we made the following box plots.

Fig. 12: Box plots of personality traits by ethnicity and drug use.

These are boxplot graphs that help us understand the effect of ethnicity on the relationship between hard drug usage and personality. In the above boxplots, we see that there are side by side boxplots for each ethnicity, wherein the red box plot represents individuals who don’t do hard drugs and the blue box plots represent hard drug users. From the first plot for neuroticism we can see that the spread for White individuals is broader as compared to other ethnicities, which indicates a higher variability within the group. Additionally, we can see that for White and Mixed-White/Black individuals drug users experience higher neuroticism as compared to non drug users. In contrast to this, Black and Asian individuals show relatively no difference between drug users and non users. The next plot explores Extraversion by ethnicity and drug use. We can see from this plot that for most ethnicities (except Mixed-White/Black), non users have a higher extraversion score as compared to hard drug users, which implies that non-drug users tend to be more social and outgoing. Additionally, We examined the openness to experience score which showed us that hard drug users had a higher score for openness to experience as compared to non-users. This shows us a potential association between drug use and the willingness to participate in new experiences. In terms of different ethnicities, we can see that there is variability in openness scores among different ethnicities. For example Mixed-White/Black hard drug users, have lower openness scores as compared to other ethnicities. The fourth graph shows agreeableness which varies by ethnicity and drug use. We can see from the graph that White individuals and the other group, drug-users display lower agreeableness than non-users which can help us conclude that drug use lowers people’s ability to cooperate. For other ethnicities such as black individuals we can see that drug use does not have as big of an impact on agreeableness.

Conscientiousness is lower among drug users in different ethnic groups, specifically among white individuals. Mixed-Black Asian drug users have the lowest conscientiousness which indicates a strong association between drug use and conscientiousness for this ethnicity. Finally, Impulsivity is shown in the last graph. From this graph we can see that for almost all ethnicities drug users tend to exhibit higher impulsivity scores. This leads to the possible association that drug use leads to a tendency to act without thinking. White individuals display a particularly high value for impulsivity, whereas Mixed-White/Black drug users have a narrower range of impulsivity scores as compared to non users, which implies less variability in this group. From the above analysis, we can see that there are different trends in personality for hard drug users and non users that belong to different ethnicities, but the magnitude of the effect varies between ethnicities. Thus, we can conclude that different ethnicities do in fact affect how hard drug use is linked to different personality traits.

Conclusions and Limitations

While previous efforts to understand the drug use-individual environment interaction has extensively been studied from the lens of socioeconomic status, there have been fewer attempts to study this relationship from the viewpoint of personality traits and other demographic variables. Here, we found associations between several personality traits, most notably openness and impulsivity, and drug use; we also found that sex was predictive of hard drug use, but not education. Finally, we noted several patterns (such as increased variability) when looking at the data by ethnicity, hard drug use, and personality, but there were no clear trends by ethnicity, and the existing observations should be validated with more data.

This study consists of multiple limitations that constrains our findings and conclusions. The primary concern of this study is the reliability of our data. Although the data does try to filter out response bias by adding a fake drug, we still do not know if those who have been filtered are answering the questions truthfully. This is specifically apparent in drug ratings and personality scores. Thus, application of these statistical tests and visualization to other similar datasets are necessary. Furthermore, due to the nature of real-world data and other unaccounted variables, our inferences are limited in their scope and may be inaccurate. For instance, GAM model has the problem of overfitting due to uncertainty and its sensitivity to smoothing parameters. We have treated SS and Impulsive scores as continuous variables despite displaying discrete variable traits. Consequently, other combinations of smoothing parameters for GAM models should’ve been made for comparison and determined through AIC or GVC.

Another concern of this study is the oversimplification and generalization of some of the variables. Specifically, the data collects the drug usage level by “last used” (used over decade ago, year ago, and etc.). This is a bit inaccurate as there may be those who only tried it once and never tried again, rather than being a regular user. Thus, using another method of measurement such as frequency of usage would have been more accurate. In order to overcome this, we have totaled the drug usage score to at least attain the idea of drug usage score (where a high score would include those who use multiple drugs often or recently). However, this method also comes with a limitation where some will not have higher scores than our intentions. For instance, an older person who has used multiple drugs decades ago will have a lower score than a younger person who has used multiple drugs recently. Overall, as this dataset cannot clearly differentiate between regular users and occasional users, the measure of “drug use” represents variability of drug use (how many drugs a person has tried) rather than frequency of drug usage.

Similarly, personality scores limit our interpretation. The scores simplify complex human traits which is a fundamental problem with these types of variables. This can lead to overgeneralization of individuals and inaccurate representations. For instance, an individual’s score for openness to experience may be different from the time when one took the test and when one tried a drug. Thus, personality scores for those who never took a drug would tend to be more accurate than those who took drugs. However, as it is difficult to measure the personality of the individual during the moment of taking a drug, there are alternative ways. Rather than recording the times of the tests, we would collect data for diagnosed mental illness and the time when they were diagnosed. Although this measure yields a bit different information, it could also be informative on tracking drug usage patterns by mental illness.

In addition to this, there are other confounding variables not accounted for in this study. The tendency of drug usage also depends on income, job, location, family(have a family or not). Furthermore,there are several aspects within this data set that this study fails to cover. First, there isn’t a specific analysis on the relationship for legal drug usage to illegal drug usage or personality scores. By only focusing on illegal and hard drugs, the study is limited in application and ignores a potential variable that may be another predictor variable between personality and hard drug usage.

References

  1. Fehrman E, Mirkes EM, Muhammad AK, Egan V, and Gorban AN (2017). The Five Factor Model of personality and evaluation of drug consumption risk. Preprint on arXiv. URL https://arxiv.org/pdf/1506.06297
  2. Patrick ME, Wightman P, Schoeni RF, and Schulenberg JE (2012). Socioeconomic status and substance use among young adults: A comparison across constructs and drugs. Journal of Studies on Alcohol and Drugs 73(5).
  3. National Center for Drug Abuse Statistics (2024). Drug Related Crime Statistics. Accessed December 4, 2024. URL https://drugabusestatistics.org/drug-related-crime-statistics/
  4. Daley DC (2013). Family and social aspects of substance use disorders and treatment. Journal of Food and Drug Analysis 21(4).