Are the Relevant Risk Factors Being Adequately Captured in Empirical Studies of Smoking Initiation? A Machine Learning Analysis Based on the Population Assessment of Tobacco and Health Study

Abstract

Introduction

Cigarette smoking continues to pose a threat to public health. Identifying individual risk factors for smoking initiation is essential to further mitigate this epidemic. To the best of our knowledge, no study today has used machine learning (ML) techniques to automatically uncover informative predictors of smoking onset among adults using the Population Assessment of Tobacco and Health (PATH) study.

Aims and Methods

In this work, we employed random forest paired with Recursive Feature Elimination to identify relevant PATH variables that predict smoking initiation among adults who have never smoked at baseline between two consecutive PATH waves. We included all potentially informative baseline variables in wave 1 (wave 4) to predict past 30-day smoking status in wave 2 (wave 5). Using the first and most recent pairs of PATH waves was found sufficient to identify the key risk factors of smoking initiation and test their robustness over time. The eXtreme Gradient Boosting method was employed to test the quality of these selected variables.

Results

As a result, classification models suggested about 60 informative PATH variables among many candidate variables in each baseline wave. With these selected predictors, the resulting models have a high discriminatory power with the area under the specificity-sensitivity curves of around 80%. We examined the chosen variables and discovered important features. Across the considered waves, two factors, (1) BMI, and (2) dental and oral health status, robustly appeared as important predictors of smoking initiation, besides other well-established predictors.

Conclusions

Our work demonstrates that ML methods are useful to predict smoking initiation with high accuracy, identifying novel smoking initiation predictors, and to enhance our understanding of tobacco use behaviors.

Implications

Understanding individual risk factors for smoking initiation is essential to prevent smoking initiation. With this methodology, a set of the most informative predictors of smoking onset in the PATH data were identified. Besides reconfirming well-known risk factors, the findings suggested additional predictors of smoking initiation that have been overlooked in previous work. More studies that focus on the newly discovered factors (BMI and dental and oral health status,) are needed to confirm their predictive power against the onset of smoking as well as determine the underlying mechanisms.

Introduction

During the past decade, the tobacco product landscape has evolved rapidly with a remarkable decrease in cigarette smoking prevalence,¹ a dramatic increase in the popularity of electronic nicotine delivery systems (ENDS), and the emergence of other novel tobacco products such as heated tobacco products.^2,3 However, cigarette smoking is still the main cause of tobacco-related morbidity and mortality, being responsible for hundreds of thousands of premature deaths and costing the U.S. economy hundreds of billions of dollars annually.^4,5 Despite the significant decline in smoking prevalence, reducing initiation rates, and increasing cessation rates remain the focus of tobacco control to further reduce the burden of tobacco use and prevent a potential reversal in the reductions of population-level smoking rates from the past decades. To better support tobacco regulations, more studies predicting smoking behaviors as well as identifying risk factors of such behaviors are needed.

Using survey data and traditional regression methods, previous research has uncovered a list of individual-specific characteristics associated with smoking initiation. Because of the standard assumptions of these models and the correlation between potential explanatory variables, these analyses can generally only incorporate a relatively small number of expert-selected predictors.^6–9 Consequently, they generally do not exploit the vast amount of information in survey data such as the many variables collected by the U.S. Population Assessment of Tobacco and Health (PATH) study.¹⁰ In addition, the findings from traditional regression methods are likely to depend on the independent variables included in the model. This dependence may lead to different conclusions if different sets of independent variables are considered⁶ and to the omission of some unanticipated predictors. Furthermore, previous studies have shown that these traditional methods may not perform optimally as prediction models, especially when dealing with class-imbalanced data (having an imbalance between class samples).^11,12 Therefore, in predicting individuals’ smoking behaviors, alternative methods are needed.

Machine learning (ML) methods such as random forest (RF) are powerful tools that can address some of these limitations. Many ML algorithms can automatically identify hidden patterns in complex data that researchers would otherwise struggle to uncover to predict future outcomes. Applications of ML techniques have just started infiltrating tobacco control research, despite the popularity of its applications in other research fields.¹³ In tobacco control, a growing number of research articles have focused on predicting smoking and vaping behaviors and identifying the most important predictors of such behaviors through ML algorithms. For instance, researchers used ML methods such as RF classifiers and penalized logistic regression models to predict ENDS use status and identify the most important predictors of this behavior based on various surveys.^13–15 Coughlin et al. employed decision trees to classify smoking cessation outcomes in group cognitive-behavioral therapy at a 6-month follow-up and identify the most relevant risk factors of treatment outcomes.¹⁶ Kim et al. applied regression tree models to identify complex interactions among various factors that differentiate global adolescents’ levels of tobacco susceptibility and use.¹⁷

Most studies used data from clinical trials or local representative surveys.^13,15,16,18 Therefore, the findings may not be generalizable to the whole U.S. population. The PATH study has been a primary data source of research work on population-level tobacco use behaviors.¹⁰ PATH is a large, longitudinal, and nationally representative survey.¹⁰ It contains detailed individual-specific information, including demographics, socioeconomic and health status, perceptions, substance use, and tobacco use. To the best of our knowledge, no study to date has applied ML models to predict smoking initiation and identify relevant predictors of this behavior in the adult PATH data. Here we employed an RF classifier paired with recursive feature elimination (RF-RFE), a popular ML method, to identify critical baseline factors that predict smoking initiation among adults who have never smoked at baseline between two consecutive PATH waves. Our analysis incorporates all PATH variables relevant to the smoking onset. This study is intended to enrich the literature on applications of ML methods in tobacco control, advance our understanding of smoking initiation, and aid tobacco regulators in designing targeted tobacco interventions.

Methods

Data Preparation

The PATH survey is a national longitudinal survey of tobacco use and its effects on the U.S. population. PATH conducted its first wave in 2013 and has publicly released five waves of adult data (wave 1: September 2013–December 2014, wave 2: October 2014–October 2015, wave 3: October 2015–October 2016, wave 4: December 2016—January 2018, and wave 5: December 2018–November 2019).¹⁰ Using the PATH survey, we aimed to predict smoking initiation between two consecutive waves among adults who have never smoked at baseline and identify relevant predictors of this behavior. Here, individuals who never tried smoking even one or two puffs of a cigarette are referred to as individuals who have never smoked. In PATH, a nationally representative sample of adults was surveyed in wave 1, completing follow-up interviews at each follow-up wave. In wave 4, a new replenishment sample was added to the study to address attrition over time and reset the sample size, making it again nationally representative.

In this study, we analyzed the earliest (waves 1 and 2) and more recent (waves 4 and 5) pairs of waves. Considering these two pairs of waves, we employed the RF-RFE method to identify the most important predictors of smoking initiation in the PATH survey and investigated whether these predictors changed over time. These important predictors were selected by the RF-RFE method as most informative for the prediction of this smoking behavior, see section “Statistical Analysis” for details. We used the wave 1 (wave 4) variables to predict the past 30-day (P30D) smoking status in wave 2 (wave 5). The P30D smoking status of an individual is defined as “Yes” if he/she smoked a cigarette in the past 30 days and “No” otherwise.

For waves 1 and 2, we first extracted a subpopulation of individuals who have never smoked in wave 1 and combined it with the outcome variable that indicates the P30D smoking status in wave 2 using individuals’ identity numbers. We then removed non-relevant variables such as personal identity numbers, random questions, sample weights, imputed variables, adult survey type, and specimen collection status. Furthermore, variables with more than 5% missing values of the total sample size were dropped to maintain the highest possible number of predictors. We also excluded individuals with missing smoking status in wave 2. To use these datasets for model development later, all the remaining missing values were imputed using a supervised imputation method.¹⁹ Finally, we removed highly correlated variables. As a result, we obtained a final analytic dataset which includes data from 5776 adults with 145 predictors, see Table 1.

Table 1.

A Description of the Clean and Complete Datasets Extracted From the PATH Data After Data Processing

Baseline smoking status	Outcome smoking status	Value of outcome smoking status		Number of predictors	Total
		Yes	No
Individuals who have never smoked in wave 1	P30D smoking status in wave 2	197 (3.4%)	5579 (96.6%)	145	5776 (100%)
Individuals who have never smoked in wave 4	P30D smoking status in wave 5	208 (2.6%)	7687 (97.4%)	182	7895 (100%)

PATH = Population Assessment of Tobacco and Health.

However, as shown in Table 1, most wave 1 individuals who have never smoked (more than 95%) remained their never smoking status in wave 2. Therefore, the distribution of the outcome P30D smoking status in wave 2 is highly imbalanced, which often has a negative impact on predicting the minority class (ie, individuals who have smoked in the past 30 days). We divided each dataset into training and test data (80%:20%). The training data were used for training and validating the chosen model classifier, and for feature selection, while the test data, which remained untouched, was used to reevaluate the model performance on the RF-RFE selected variables.

Analogously, we repeated this whole process and obtained the final clean dataset for waves 4 and 5 (7895 adults with 182 predictors) shown in Table 1. A difference in the number of predictors between these two pairs of waves is because of the changes in the PATH survey questionnaires between wave 1 and wave 4. These data preparation processes are illustrated in Figure 1.

Figure 1.

Diagram of the data preparation process, where N is the number of participants and P is the number of variables including the outcome variable.

Statistical Analysis

After data preprocessing, we obtained two datasets corresponding to two pairs of waves in Table 1. Our datasets with these many variables are likely to contain many redundant or irrelevant predictors. Searching for an optimal subset of predictors is usually needed to produce more effective data mining algorithms, improve the predictive accuracy of the considered model, and increase the comprehensibility of the underlying process that generated the data.²⁰ As such, for each pair of waves, we used an RF classifier combined with RFE to obtain a subset of predictors on which the RF classifier performs best.²¹ RF is an ensemble method that consists of many decision trees trained in parallel on different subsets bootstrapped from the original data. The final RF decision is generated by aggregating the decisions of all individual decision trees. For RF-RFE, the RF classifier was, first, trained on the entire training set, providing a rank order of all the predictors based on the mean decrease of accuracy. The mean decrease in accuracy was computed by randomly permuting the values of each predictor. Variables whose permutation causes a significant loss in accuracy correspond with high-importance scores. Variable(s) with the lowest important ranking was (were) then removed from the dataset. The model was re-trained on the reduced predictors, and its performance was recorded. This process was repeated until all the predictors were explored.²¹ We trained RF-RFE on the training data with 5-fold cross-validation repeated 3 times to score different subsets of predictors, select the best subset of predictors, and avoid selection bias.²² To overcome this class imbalance issue and improve the classification power of the model, we performed random oversampling examples (ROSE – an R package)²³ of the training set within each cross-validation iteration (oversampling the training folds, but not the validation one). To guarantee the stability of our selected predictors, we repeated this RF-RFE training process until the intersection of the best subsets of the RF-RFE selected predictors of all simulations remains the same after many iterations (a total of 100 simulations were enough for this). Finally, the final most important predictor subset, which may contain all predictors if they are all informative, was a set of the RF-RFE selected predictors which were chosen in at least 95% of all simulations. For more details on how RF-RFE works, we refer the reader to Kuhn’s work.²² Here, we chose RF as a core method for RFE due to its appealing features including high prediction accuracy, ability to deal with high dimensional data, and internal measure of variable importance.^24,25

To evaluate how well the model performance would generalize to new data drawn from the same distribution, but not included in the training data, we trained the eXtreme Gradient Boosting method (XGBoost) using the training data with the RF-RFE selected variables and tested the model’s ability to classify the P30D smoking status using the test data. XGBoost, an efficient and effective open-source implementation of Gradient Boosting, which was developed by Chen and Guestrin in 2016,²⁶ has been shown to outperform many other classification techniques.

Performance Measures of the RF Classifier

A receiver operating characteristic (ROC) curve is a graphical representation of a classifier’s performance as a tradeoff between specificity and sensitivity. It is typically obtained by plotting the true positive rate (sensitivity) against the true negative rate (specificity) for a single classifier at various probability cutoff values. For each outcome wave,

$${\displaystyle \begin{array}{ll}{\displaystyle \mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}}& =\frac{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}\\ & =\frac{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}30\mathrm{D}\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{s}}{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}30\mathrm{D}\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{s}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{s}}\end{array}}$$

$${\displaystyle \begin{array}{ll}{\displaystyle \mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}}& =\frac{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}\\ & =\frac{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{s}}{\mathrm{t}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{s}+\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{P}30\mathrm{D}\mathrm{s}\mathrm{m}\mathrm{o}\mathrm{k}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{s}},\end{array}}$$

where individuals with true P30D smoking status (false never smoking status) are their actual P30D smoking status in the outcome wave who are correctly (incorrectly) predicted by the model, and individuals with true never smoking status (false P30D smoking status) are their actual never smoking status in the outcome wave who are correctly (incorrectly) predicted by the model. Sensitivity and specificity show the ability of a classifier to correctly identify individuals’ P30D smoking status in wave 2 (wave 5) among wave 1 (wave 4) those who have never smoked. The area under the ROC curve (AUC) is usually a preferred measure of the model performance in binary classification problems. AUC is a more discriminating measure of performance than accuracy in classification.²⁷ The closer AUC gets to 100%, the better the classifier performance. A perfect classifier corresponds to an AUC of 100%, while a classifier without any predictive power corresponds to an AUC of 50%. Finally, to provide a reliable estimate of AUC, we also computed the 95% confidence interval (CI) of AUC with 2000 stratified bootstrap replicates.

The study was exempted from the review by the Institutional Review Board at the authors’ institution because of the public availability of the data.

Results

Most Relevant Predictors of Smoking Initiation

RF- RFE selected a list of only 64 and 56 relevant variables of 145 and 182 possible variables in waves 1 and 4, respectively, for predicting smoking initiation among adults who have never smoked at baseline as shown in detail in Supplementary Tables A1, A2 in the appendix. With these recommended variables, the classification model’s performance metrics (Accuracy and Kappa²⁸) reach their best levels. Based on the description of each variable, we grouped them into different categories as detailed in Supplementary Tables A1, A2 in the appendix. Variables describing the same characteristics are put into the same category. For instance, the “Social influence” category includes the variables in waves 1 and 4 that describe the number of hours in past 7 days that an individual was in close contact with others when they were smoking. We refer the readers to Supplementary Tables A1, A2 in the appendix for the description of each category in Table 2.

Table 2.

Individuals’ Characteristics Related to the Transition from Individuals Who Have Never Smoked to Those Who Have Smoked in the Past 30 Days Between Two Consecutive Waves

Wave 1	Wave 4
BMI
Census region
Dental and Oral care
Education
Employment status
Experimentation with other tobacco products
Exposure to radio and TV
Financial status
Gender, Race
Mental and Physical health status
Social influence
Tobacco risk perceptions and awareness
Alcohol use status
Exposure to tobacco products and tobacco advertisements	Exposure to anti-tobacco advertisements
Household rules about tobacco use	Marital status
Insurance status	Smoking intention
Relatives’ asthma status
Taking anti-inflammatory or pain medication
Presence of youth in the household
Internet use

As a result, we obtained a set of individual characteristics associated with the onset of cigarette smoking, see Table 2 and Supplementary Tables A1, A2 in the appendix for more details. Table 2 indicates that across the considered baseline waves, census region, BMI, dental and oral care, education, employment status, experimentation with other tobacco products, exposure to radio and TV, financial status, mental and physical health status, gender, race, social influence, and tobacco risk perceptions and awareness are important factors associated with smoking initiation by the following wave. In addition to those variables, alcohol use status, exposure to tobacco products and tobacco advertisements, household rules about tobacco use, insurance status, relatives’ asthma status, taking anti-inflammatory or pain medication, presence of youth in the household, and internet use are related to smoking initiation between waves 1 and 2. However, exposure to anti-tobacco advertisements, marital status, and smoking intention are predictive factors identified only in the wave 4 to 5 analysis.

Supplementary Figures A1, A2 in the appendix comprehensively present the rankings of all the RF-RFE selected predictors and their impact directions on P30D smoking status.

Performance of the RF Classifier

Figure 2 shows the predictive power of the trained RF classifiers on the test data using the selected variables (around 60 variables). Critical metrics for evaluating the classifier’s performance including AUCs, Specificity, and Sensitivity with their 95% CIs are computed. Column 3 indicates that the RF classifier performs very well on all considered datasets with AUCs of around 80% (82% [CI: 75.5% to 87.5%] for waves 1 and 2 and 78% [CI: 70% to 85.5%] for waves 4 and 5). Column 3 displays the suggested probability cutoffs which are the closest points to the top-left parts of the plots with perfect sensitivity and specificity. These suggested thresholds are computed using the closest-topleft method.²⁹ Individuals with predicted probability greater than these thresholds are classified as individuals who have never smoked and individuals who have smoked in the past 30 days otherwise. Columns 4 and 5 are the values of sensitivity and specificity corresponding with the thresholds in column 3. Increasing sensitivity, more individuals who have smoked in the past 30 days identified correctly, leads to a decrease in specificity (more individuals who have never smoked would be incorrectly identified) and vice versa. Therefore, these thresholds are just suggestive and should be adjusted to achieve one’s desirable balance between sensitivity and specificity. The ROC curves of the XGBoost classifiers are plotted in Figure 2. The list of optimized XGBoost hyperparameters can be found in Supplementary Table A3 in the appendix.

Figure 2.

The receiver operating characteristic curves of all the XGBoost classifiers. The diagonal line in each plot represents random guessing. Each plot shows the area under the ROC curve (AUC) curve (95% confidence interval [CI]) together with a suggested threshold (specificity, sensitivity).

Discussion

This is the first study to leverage all possible original public adult PATH variables to extract a subset of the most informative variables for predicting the transition from having never smoked to having smoked in the past 30 days between two consecutive waves using RF-RFE. We uncovered about 60 variables associated with the smoking onset between two considered PATH waves among adults who have never smoked at baseline. Despite the severely skewed class distribution shown in Table 1, with these RF-RFE selected variables, the XGBoost model performs well in classifying smoking status in the follow-up wave among individuals who have never smoked at baseline with AUC = 80% approximately (see Figure 2). The summary in Table 2 shows that most of the discovered risk factors are robust across waves and their association with smoking initiation is well reported in the literature.

Before starting our discussion about the association between the resulting risk factors in Table 2 and the smoking initiation behavior, we would like to emphasize that the direction of the impact (ie, negative or positive) of each risk factor on the onset of smoking, as discussed below, is based on the exhaustive information provided in Supplementary Tables A1, A2 and Figures A1, A2 in the appendix.

For the joint predictive variables between waves 1 and 4, most of them (ie, census region,³⁰ education,^31,32 employment status,³³ experimentation with other tobacco products,^32,34–38 exposure to radio and TV,³⁹ financial status,^32,40 mental and physical health status,⁴¹ gender,³⁰ race,^31,32,42 social influence,^43–45 and tobacco risk perceptions and awareness^31,45) are familiar risk factors of smoking initiation. Meanwhile, other factors including BMI and dental and oral health status, have not been consistently identified as predictors for smoking initiation in the literature. Across the considered waves, our model consistently indicates BMI as one of the main predictors of the onset of cigarette smoking in the following waves. Our findings indicate that individuals with lower BMI are likely to be at a higher risk of smoking initiation. Several studies have found a strong connection between BMI and smoking initiation while analyzing different datasets.^45–48 This association could be related to weight concerns since many believe that smoking helps control their weight or via other sociodemographic factors such as lower socioeconomic status, income, and educational attainment, which are highly correlated to BMI.^45–48 Oral and dental care robustly appears among the predictors of smoking initiation in both baseline waves. The association between oral and dental care and smoking onset has not been widely reported in the literature, except in earlier studies that analyzed data from a randomized clinical trial in California to predict tobacco initiation over a 2-year follow-up interval among adolescent never-tobacco users.^48,49 People who take good care of their oral health seem to be less likely to start smoking since they are probably aware of the harmful impacts of smoking on their teeth and gum health. More work is required to investigate and understand the impacts of these factors on smoking initiation.

Besides the shared predictors of smoking initiation across the considered waves, each baseline wave has its own set of relevant predictors, in part because of the changes in the PATH questionnaire. Again, most wave-specific variables are well reported in previous studies except taking anti-inflammatory or pain medication, insurance status, relatives’ asthma status, and internet use in wave 1. We found that taking anti-inflammatory or pain medication is also associated with less risk of smoking initiation. Individuals taking these medications are likely to be already in bad health status, and thus do not want to start smoking. In addition, our results show that individuals with no insurance are related to a higher likelihood of smoking initiation.³⁰ This link is probably because of the confounding effects of other factors such as income and employment status. Individuals who have any of their biological relatives (father, mother, brothers or sisters, sons or daughters, and both living and deceased) with asthma are associated with lower smoking probability. This can be because of an increase in tobacco risk awareness since asthma is a well-known tobacco-related disease. Finally, our findings show that the more time an individual spends on the internet, the more likely that person is to initiate smoking. Internet use is likely to be an indicator of exposure to smoking-stimulating factors, including onscreen tobacco use, tobacco advertisements, and digital social influence. A recent study¹⁴ reported a similar observation while using a penalized logistic regression model on the PATH data to predict the use of ENDS among never users of any tobacco products. These authors found an association between digital media use with future ENDS use when using different waves in U.S. adolescents. There could be other possible explanations for the connection between these uncommon factors and smoking initiation. As such, more studies on these matters are needed.

Like other similar studies,^14,18 our work has limitations. First, since the PATH survey is self-reported, the responses are potentially subject to recall bias. Second, our analysis was conducted based on the unweighted PATH data, therefore our results may not generalize to the whole population. However, since the PATH survey is nationally representative, for example, the survey sample captures the characteristics of most U.S. individuals, our conclusions are likely to hold in general. External validation of these findings is needed. Third, factors not included in the PATH study were not considered. Fourth, even though we simulated RF-RFE a hundred times to obtain the final converged set of most informative predictors of smoking onset, other ML methods may provide a slightly different set with which the XGBoost classifier’s discriminative power remains similar to our presented results. As such, a future study that is devoted to comparing the best sets of the most informative predictors of this smoking behavior that result from different feature selection techniques is encouraged. Finally, it should be noted that our results cannot establish the causal direction of the association of the model-selected predictors with smoking initiation because of the nature of these ML algorithms.

While most smoking experimentation happens during adolescence, more established smoking behavior is usually observed during young adulthood.⁴ In addition, recent studies have reported a shift in the age of smoking initiation towards older ages.⁵⁰ As such, with our outcome of interest, the current P30D smoking status, it is relevant to analyze the young adult PATH data to understand smoking initiation. In PATH, according to our definition of the baseline population and the outcome of interest, we found that the unweighted proportions of individuals who have never smoked at baseline and have smoked in the past 30 days in the following wave were similar in both the adult and youth data. We chose to start with the adult data first to demonstrate the feasibility of this approach. The analysis based on the youth data, our ongoing work, is more complex because of the aged-out individuals whose behaviors are different from the rest. As such, analyses of predictors of adolescent initiation will be the focus of a follow-up paper.

In summary, this work demonstrates the potential applications of ML methods in developing highly accurate classification models to predict smoking initiation using nationally representative survey data and systematically identifying all potential risk factors of smoking behaviors in the PATH data. Besides reconfirming well-known risk factors in the literature, our findings discovered additional predictors of smoking initiation that have not been paid much attention to in previous studies. More studies that focus on the newly discovered factors are needed to confirm their predictive power against changes in smoking behaviors as well as determine the underlying mechanisms.

Supplementary Material

A Contributorship Form detailing each author’s specific involvement with this content, as well as any supplementary data, are available online at https://academic.oup.com/ntr.

Funding

Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health (NIH) and FDA Center for Tobacco Products (CTP) under Award Number U54CA229974. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or the Food and Drug Administration.

Declaration of Interests

The authors do not report any conflicts of interest.

Data Availability

All data produced are available online at https://www.icpsr.umich.edu/web/NAHDAP/studies/36231