This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted noncommercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Tobacco use among older adults remains a pressing public health concern in South Korea, particularly among men. Despite the decline in tobacco use rates with age, a range of sociodemographic, psychological, and behavioral factors persist to influence tobacco use behaviors. However, prior studies have primarily relied on linear modeling approaches, which may overlook indirect associations. This study addressed this gap using classification and regression tree (CART) analysis to identify hierarchical and interactive patterns among factors associated with tobacco use.
Methods
We analyzed data from the 2023 Korea Community Health Survey, which included 34,924 Korean men aged ≥65 years. Using CART analysis, we identified distinct subgroups and patterns across multiple predictors of tobacco use behavior, including demographic factors, lifestyle characteristics, and chronic health conditions.
Results
CART analysis identified age as the strongest predictor of tobacco use. Men aged ≤72.5 years had higher tobacco use rates, especially those with low social engagement (24.8%). Younger men with higher social activity and aged ≤67.5 years had the lowest rate (12.9%). Among men ≥72.5 years, alcohol use was the key predictor. Non-drinkers had the lowest tobacco use rate (9.3%), whereas drinkers aged ≤78.5 years showed elevated rates (22.0%), suggesting persistent risk even in later life.
Conclusion
These findings underscore the importance of promoting social engagement and reducing alcohol use to decrease tobacco use among older men, particularly those under 73 years of age.
Tobacco use remains a significant health concern for older adults in South Korea, particularly for men [1,2]. Although tobacco use rates tend to decline with age, various sociodemographic, psychological, and behavioral factors contribute to continued tobacco use later in life. The Korea National Health and Nutrition Examination Survey reported a current tobacco use rate of 19.6% among men aged 65 years and older [3]. In addition, lower educational attainment, unemployment, and being single are associated with higher tobacco use prevalence among older adults in South Korea [1,4].
Psychological distress, including elevated stress and depressive symptoms, increases the risk of tobacco use [5,6]. Similarly, social isolation and loneliness are distinct psychosocial factors associated with a higher likelihood of tobacco use, particularly among socially isolated older men [7,8]. In contrast, engagement in physical activity is associated with reduced tobacco use [9].
Health literacy and self-rated health also play complex roles in tobacco use behavior. Although some studies have not directly measured these constructs, they have highlighted related factors, such as lower educational attainment and perceived importance of smoking cessation, which may indirectly contribute to continued tobacco use and cessation difficulties [10,11]. More directly, poor health literacy and negative health perceptions have been associated with sustained tobacco use and challenges in cessation efforts [12,13]. However, most prior research has relied on linear modeling approaches, limiting the understanding of the interactive effects among key predictors.
This study addresses this gap by applying classification and regression tree (CART) analysis to a nationally representative sample of older Korean men from the 2023 Korea Community Health Survey [14,15]. The central aim of this study was to explore how demographic, behavioral, and psychosocial factors, individually and combined, shape patterns of current tobacco use in later life. By leveraging CART’s strength in modeling nonlinear and hierarchical interactions, this study aimed to answer two key research questions. First, what combinations of factors place older Korean men at an elevated risk of continued tobacco use? Second, how can these interaction-based risk profiles inform more precise and effective public health interventions? Unlike traditional models that test additive main effects, CART enables the identification of distinct data-driven subgroups, such as younger unmarried men with low social engagement, who may benefit from specifically tailored cessation strategies. Accordingly, the study makes both methodological and practical contributions to the literature by demonstrating the value of using machine learning approaches to uncover hidden risk structures and supporting the development of interventions that are better matched to the lived realities of vulnerable populations.
Literature review
Tobacco use remains a leading cause of death worldwide and adversely affects nearly every organ of the body [16]. Globally, there are 1.25 billion adult tobacco users [17]. Tobacco use is responsible for approximately 8 million deaths [18], contributing to a substantial global health burden. The economic impact is equally significant, with global tobacco use estimated to cost US dollar (USD) 1.85 trillion annually [19]. Similarly, tobacco use remains a significant health concern in South Korea. In 2022, tobacco use-related deaths totaled 72,689 (63,452 male, 9,237 female), with the associated socioeconomic burden estimated at approximately USD $10 billion [20]. Moreover, the Korea Disease Control and Prevention Agency (KDCA) reported a steady increase in tobacco-use-related mortality over 3 years, from 2020 to 2022, with men disproportionately affected.
Although prior studies conducted in the South Korean context have rarely specified an explicit theoretical framework, many have examined tobacco-use-related factors across multiple levels: individual (e.g., socioeconomic status, stress, depressive symptoms, and nicotine dependence), community (e.g., social networks and peer norms), and national (e.g., tobacco control policies and health inequality).
This multilevel approach aligns with the socio-ecological model, which posits that individuals influence and are influenced by multiple nested layers in their social and physical environments [21]. Guided by this perspective, this study adopts the socio-ecological model as a conceptual framework to inform variable selection and interpret tobacco use behavior in a multilevel context.
Sociodemographic factors and tobacco use
Previous research in the South Korean context examining tobacco use cessation patterns has revealed significant challenges among men aged ≥65 years, with 66.1% showing persistent tobacco use [2] and only 26.3% expressing cessation intention [1], reflecting the broader national pattern described in the 2023 national survey [3]. However, individuals aged ≥70 years are significantly more likely to quit tobacco use, potentially because of their medical situation (e.g., presence or family history of chronic disease) and increased health awareness through cancer screening or physical examinations [2,4]. Factors such as lower educational attainment, unemployment, and being single are associated with higher tobacco use rates [4,22].
Psychological and mental health factors and tobacco use
Perceived stress is associated with tobacco use, with individuals who experience high stress levels being more likely to use tobacco [5,6]. Individuals who intend to quit tobacco use report higher levels of stress, suicidal ideation, and recent depressive episodes [5]. Furthermore, current tobacco use is associated with increased suicide mortality, even after adjusting for depression [23]. These findings are consistent with those of a previous study conducted in a different cultural context that identified tobacco use as a coping mechanism for stress and depressive episodes or mood [24].
Social and lifestyle factors and tobacco use
Previous studies have demonstrated significant associations between various social and lifestyle factors and tobacco use behaviors in older adult populations. Social isolation has emerged as a particularly salient risk factor associated with tobacco and alcohol use, with socially isolated individuals exhibiting significantly higher odds of continued tobacco use and diminished cessation success rates [7,8]. Older adults experiencing loneliness or inadequate social support networks demonstrate heightened vulnerability to both tobacco and alcohol use behaviors [25]. Conversely, social engagement through participation in sports clubs and community activities is associated with more positive health behaviors and outcomes [26,27]. Accordingly, tobacco users tend to be more physically inactive than non-tobacco users [9]. However, social gatherings in predominantly male social environments frequently demonstrate positive correlations with tobacco use. In rural Korean populations, higher levels of social support and participation are positively associated with tobacco use behavior among older men, indicating that social-contextual factors may reinforce continued tobacco use [28].
Furthermore, alcohol and tobacco use frequently co-occur as health-risk behaviors with common social determinants. Research has demonstrated that certain living arrangements, particularly living alone, are associated with increased rates of tobacco use and heavy alcohol consumption [22,29].
Health perceptions and literacy and tobacco use
Prior research across diverse cultural settings has shown that limited health literacy among older adults is associated with a higher prevalence of tobacco use [10], increased difficulty in cessation efforts, and higher relapse rates [12]. Similarly, other studies have found that current tobacco users frequently report poor self-rated health [11,13]. For some individuals, this perception motivates them to quit tobacco use, whereas others may continue tobacco use despite recognizing their poor health, possibly because of delayed concern or a low perceived need to quit [13].
Based on the literature, we created a guiding conceptual framework and conducted analyses accordingly. The framework is illustrated in Figure 1.
Methods
Data
The Korea Community Health Survey has been conducted in person by the KDCA since 2008 to support the construction and development of local public health planning. The 2023 survey was conducted across South Korea from May 16 to July 31, 2023. The target population was individuals aged 19 years or older. A total sample of 231,752 individuals (45.6% male and 54.4% female) was drawn using a stratified probability proportional to size systematic sampling method. The 2023 survey was administered through face-to-face interviews conducted by trained field interviewers. For this study, we selected a subset of 34,924 men aged 65 years and older from the original dataset. This restriction was applied to focus on older male adults, a population particularly vulnerable to health risks and behavioral factors relevant to the study objectives. The original dataset is available upon request from the KDCA (https://chs.kdca.go.kr/chs/index.do).
This study used publicly available, de-identified secondary data; therefore, it was not considered human subjects research and did not require Institutional Review Board review or informed consent.
Dependent variable
The dependent variable in this study was a binary measure of current tobacco use (“0=no,” “1=yes”) constructed using three items: current use of conventional cigarettes, current use of heated tobacco products, and current use of electronic cigarettes with nicotine. The variables of current use of conventional cigarettes and current use of heated tobacco products comprised six response categories: “1=smoke every day,” “2=smoke occasionally,” “3=used to smoke but not currently,” “7=refused to answer,” “8=not applicable,” and “9=don’t know.” For analysis, we recoded responses 1 and 2 as 1 (current tobacco user) and responses 3 and 8 as 0 (non-user). The remaining response categories (7 and 9) were treated as system missing. Current use of electronic cigarettes with nicotine was originally coded as a continuous variable, with participants indicating how many days they had used it within the past month. Responses ranged from 0 to 31, with categories of 88 (non-users) and 99 (system missing). We recoded responses 1–31 as “1=yes,” and 0 and 88 as “0=no.” After summing the three recoded items, if at least one was coded as 1=yes, the final variable was recoded as “1=yes, currently using tobacco.” Otherwise, it was recoded as “0=no, not using tobacco.”
Independent variables
We examined the associations between the dependent variable and the following factors: age, walking days per week, perceived stress level, depression, social isolation, social activity, self-rated health, health literacy, educational attainment, economic activity status, monthly household income level, marital status, alcohol use, and chronic disease.
The perceived stress level item originally comprised six response categories: “1=extremely stressed,” “2=very stressed,” “3=somewhat stressed,” “4=barely stressed,” “7=refused to answer,” and “9=don’t know.” For the analysis, the response scale was reverse-coded, with higher values representing greater levels of perceived stress. Specifically, 1 was recoded as 4, 2 as 3, 3 as 2, and 4 as 1. Response categories 7 (“refused to answer”) and 9 (“don’t know”) were treated as system missing. For descriptive purposes, this variable was considered continuous.
The survey used the Patient Health Questionnaire-9 (PHQ-9) to assess depression. The PHQ-9 comprises nine items that assess depressive symptoms over the past 2 weeks. Responses range from “0=not at all” to “3=nearly every day,” with higher scores indicating more severe depressive symptoms [30]. In the survey, the original response categories for each item of the PHQ-9 were “1=not at all,” “2=several days,” “3=more than a week,” and “4=nearly every day.” However, for scoring purposes, the response categories were recoded according to the original PHQ-9 as follows: “0=not at all,” “1=several days,” “2=more than a week,” and “3=nearly every day.” All items were then summed to construct the depression item. This scale demonstrated acceptable internal consistency in our sample (Cronbach’s α=0.789), although its reliability may vary across populations.
Social isolation was assessed based on reported frequency of contact with relatives, neighbors, and friends. Each item had the following response categories: “1=less than once a month,” “2=once a month,” “3=2–3 times a month,” “4=once a week,” “5=2–3 times a week,” “6=4 or more times a week,” “7=refused to answer,” and “8=don’t know.” For the analysis, responses coded as 7 (“refused to answer”) and 8 (“don’t know”) were treated as system missing. The remaining categories (1–6) were recoded as 0–5. Finally, the scores for the three recoded items were summed to construct the social isolation variable.
The social activity variable was created using three items: religious activity, friendship/social gathering, and leisure/recreational activity. Each item included the following response categories: “1=yes,” “2=no,” “7=refused to answer,” and “9=don’t know.” For the analysis, responses 7 and 9 were treated as system missing, whereas the remaining categories were recoded as “0=no” and “1=yes.” The three recoded items were then summed to generate the final social activity variable.
The self-rated health variable comprised seven response categories: “1=excellent,” “2=good,” “3=fair,” “4=bad,” “5=very bad,” “7=refused to answer,” and “9=don’t know.” Categories 7 and 9 were classified as missing systems. For ease of interpretation, the response categories were reverse-coded, with higher values representing better self-rated health. Although originally a categorical variable, it was treated as continuous for statistical analysis and interpretation.
The health literacy variable was constructed by combining items related to the recognition of early symptoms of cerebral infarction (e.g., weakness in the arms and legs, slurred speech, one-sided vision loss, difficulty maintaining balance, and severe headache) and myocardial infarction (e.g., pain in the jaw, neck, or back, dizziness, chest pain, pain in the arms and shoulders, and shortness of breath). The internal consistency demonstrated good reliability for the cerebral infarction (Cronbach’s α=0.833) and myocardial infarction (Cronbach’s α=0.828) symptom recognition items. The original response categories for these items were as follows: “1=yes,” “2=no,” “7=refused to answer,” and “9=don’t know.” For analysis, responses coded as 7 (“refused to answer”) and 9 (“don’t know”) were set as system missing, whereas responses of 1 and 2 were recoded as a binary variable (“0=no,” “1=yes”). Finally, all 10 recoded items were summed to construct the health literacy variable.
The original educational attainment variable included the following: “1=no education,” “2=traditional Korean village school,” “3=elementary school,” “4=middle school,” “5=high school,” “6=associate degree,” “7=bachelor’s degree,” “8=graduate degree,” “77=refused to answer,” and “99=don’t know.” For analysis, categories 1–4 were recoded as “1=less than high school,” categories 5 and 6 as “2=high school,” categories 7 and 8 as “3=college and more,” and 77 and 99 were treated as system missing.
The economic activity status item was also recoded. The item comprised response categories of “1=yes,” “2=no,” “7=refused to answer,” and “9=don’t know.” Response categories 7 and 9 were treated as system missing, whereas categories 1 and 2 were recoded as “0=no” and “1=yes.”
The monthly income variable was used to measure economic level (n=29,647) because the annual income variable had a high nonresponse rate (n=4,866; 13.9%). Monthly income was originally recorded as a continuous variable. However, the response distribution was right-skewed; therefore, it was divided into quartiles and used as a categorical variable (“1=low,” “2=lower-middle,” “3=upper-middle,” and “4=high”), because quartile-based methods are robust to skewed data distributions and less sensitive to extreme values that can cause bias [31].
Marital status was categorized as follows: “1=married, living together,” “2=married, but not living together,” “3=widowed,” “4=divorced,” “5=never married,” “7=refused to answer,” and “9=don’t know.” In this study, 7 and 9 were classified as system missing. Response categories 1 and 2 were recoded as “1=married” and “2=separated.” Categories 3–5 were recoded as “3=single.”
Alcohol use was calculated based on annual frequency. The relevant item included nine response categories: “1=did not drink at all during the past year,” “2=less than once a month,” “3=about once a month,” “4=2–4 times a month,” “5=2–3 times a week,” “6=4 or more times a week,” “7=refused to answer,” “8=not applicable,” and “9=don’t know.” Considering this study’s focus on factors related to tobacco use among older Korean men, the item was recoded to measure current alcohol use: categories 7–9 were classified as system missing, categories 2–6 were recoded as “1=yes,” and category 1 was recoded as “0=no.”
We constructed the chronic disease variable by combining the history of diabetes and hypertension diagnosis items. Each original item had four response categories: “1=yes,” “2=no,” “7=refused to answer,” and “9=don’t know.” For analysis, responses 7 and 9 were treated as missing values. Each condition was first recoded as binary (“1=yes” if the original response was 1, and “0=no” if the original response was 2). Respondents were classified as having a chronic disease (coded as “1=yes”) if they were diagnosed with diabetes, hypertension, or both. Respondents were classified as not having a chronic disease (coded as “0=no”) only if they answered “no” for both conditions.
Data analysis strategies
Data analysis was conducted using univariate, bivariate, and multivariate modeling. Univariate analyses were conducted to examine the overall distribution and central tendencies of each variable. Means and standard deviations were computed for continuous variables, such as age, walking days per week, perceived stress, depression (PHQ-9), social isolation, social activity, self-rated health, and health literacy. Frequencies and percentages were reported for categorical variables, including educational attainment, economic activity, monthly household income level, marital status, alcohol use, chronic disease status, and tobacco use. These initial summaries helped identify variable distributions, detect potential outliers or missing data patterns, and describe sample characteristics.
Bivariate analyses were used to explore the relationships between each independent variable and the dependent variable, current tobacco use. Owing to unequal variances and group sizes, Welch’s t-tests were used for continuous variables, whereas chi-square tests were used for categorical variables. Considering the large sample size, effect sizes (e.g., Hedges’ g and Cramer’s V) were calculated to assess the magnitude of the associations beyond statistical significance. These analyses provided insights into the potential predictors of tobacco use and informed variable selection for subsequent modeling.
CART analysis was the primary multivariate modeling strategy, offering a more flexible and explanatory approach than traditional regression techniques. Unlike binomial logistic regression, which assumes linearity in the log-odds and additive effects of predictors, CART is a non-parametric method that recursively partitions data based on predictor variables to produce homogeneous subgroups. This is beneficial for identifying complex interactions and nonlinear relationships among predictors without requiring strict statistical assumptions. Furthermore, CART results are presented as decision trees, which enhance interpretability and have practical applications in policy and intervention settings. This underscores the relevance of our findings in real-world scenarios.
All analyses were performed using IBM SPSS ver. 29.0 (IBM Corp.) for descriptive and bivariate analyses. The CART module in Modeler in Minitab Statistical Software ver. 27.0 (Minitab LLC), another industry-standard tool, was used to construct and validate the classification tree. Using these familiar and reliable tools ensured the credibility and robustness of our analysis.
Results
Descriptive statistics
Our study examined 34,924 men aged 65 and older, with an average age of 73.9 years (standard deviation [SD]=6.6). Participants reported walking an average of 4.2 days per week (SD=2.9). Regarding mental health, the mean perceived stress level was 1.8 (SD=0.7) and the mean depression score (PHQ-9) was 2.0 (SD=3.0).
Social engagement measures revealed mean social isolation and social activity scores of 8.4 (SD=4.0) and 1.0 (SD=0.9), respectively. Participants typically rated their health as moderate, with mean self-rated health and health literacy scores of 3.0 (SD=1.0) and 7.7 (SD=3.0), respectively.
Educational attainment varied, with most participants (54.5%) reporting less than a high school education, whereas 31.5% completed high school, and 14.0% finished college or higher. Approximately half of the participants were not economically active. Among those who reported their monthly income, income levels were distributed across the following quartiles: low (17.8%), lower-middle (24.9%), upper-middle (17.0%), and high (25.2%). Most were married (81.7%), with a few reporting being single (14.5%) or separated (3.8%). Slightly more than half (54.0%) of the participants consumed alcohol, whereas 33.9% were non-drinkers. The majority (60.9%) reported having at least one chronic disease, and 80.2% were non-tobacco users. Among tobacco users specifically, 19.6% were current conventional cigarette users, 0.5% were heated tobacco product users, and 0.4% were electronic cigarette users, after applying sampling weights for age and sex standardization (crude prevalence is shown in Table 1).
Bivariate analysis
Most participants in our sample were non-tobacco users (n=27,993; 80.2%), whereas 6,917 (19.8%) were tobacco users. Welch’s t-tests were conducted because of differences in group sizes. Tobacco users were typically younger (mean±SD: 71.61±5.90 years) than non-users (74.48±6.68 years), and this difference was statistically significant (t(11,705.82)=35.33, P<0.001, effect size Hedges’g=0.44), suggesting a moderate association between age and tobacco use.
Regarding physical activity, non-tobacco users reported walking slightly more frequently (4.27±2.85 days per week) than tobacco users (4.02±3.08 days per week; t(10032.83)=5.99, P<0.001, g=0.08). This association was statistically significant; however, the effect size was negligible.
Tobacco users reported modestly higher perceived stress levels (1.85±0.77) than non-tobacco users (1.74±0.71; t(10020.30)= −10.57, P<0.001, g=−0.15), indicating a positive association between tobacco use and stress, although the practical significance was small.
Depressive symptoms followed a similar pattern, with tobacco users showing slightly higher PHQ-9 scores (2.17±3.15) than nonusers (1.95±2.96; t(10099.08)=−5.44, P<0.001, g=−0.08). However, the effect size was very small, indicating minimal practical significance (Table 2).
Regarding social isolation, tobacco users reported slightly fewer social contacts (8.17±4.06) than non-tobacco users (8.46±3.97; t(10390.10)=5.28, P<0.001, g=0.07). Despite the statistical significance, the effect size was negligible, likely owing to the large sample size.
Similarly, tobacco users engaged in fewer social activities (0.84±0.83) than non-tobacco users (1.04±0.91; t(11361.31)=17.39, P<0.001, g=0.22). Considering the effect size, the difference between the two groups was of limited real-world significance. In contrast, virtually no differences were observed in self-rated health between tobacco users (2.97±0.93) and non-tobacco users (2.97±0.95; t(10830.80)=−0.31, P=0.76, g=−0.004). Unlike self-rated health, health literacy differed slightly between tobacco users (7.45±3.09) and non-tobacco users (7.74±2.95; t(9279.57)=6.77, P<0.001, g=0.10). However, the effect size was small, suggesting minimal practical significance despite statistical significance.
Tobacco use and educational attainment were significantly associated (χ2(2)=56.48, P<0.001). A higher proportion of tobacco users had less than a high school education (55.1%) than non-tobacco users (54.3%). High school graduates accounted for 33.5% of tobacco users and 31.0% of non-tobacco users. The most notable difference appeared in the highest education category, with only 11.4% of tobacco users having a college degree or higher compared to 14.7% of non-tobacco users. Despite the statistical significance, the effect size was minimal (Cramer’s V=0.04), suggesting that the practical significance of this relationship was quite small.
Tobacco use was significantly associated with economic activity status (χ2(1)=56.05, P<0.001). Specifically, tobacco use was more prevalent among those engaged in economic activity (53.7%) than among those who were not (48.6%). Although statistically significant, the effect size was minimal (Cramer’s V=0.04), indicating that this association may have been partially driven by the large sample size.
Monthly household income was significantly associated with tobacco use (χ2(3)=38.07, P<0.001). The proportion of current tobacco users was higher in the lowest- and highest-income groups than in the lower- and upper-middle-income groups. Despite the statistical significance, the effect size was negligible (Cramer’s V=0.04), indicating that the statistical significance might have been inflated because of the large sample size.
Another significant association was observed between marital status and tobacco use (χ2(2)=369.94, P<0.001). Compared to non-tobacco users, tobacco users were less likely to be married and more likely to be single or separated. This suggests that individuals who are not married or living with a partner are more prone to tobacco use. However, the effect size was small (Cramer’s V=0.10), indicating that the association had limited strength.
Alcohol and tobacco use showed a significant association (χ2(1)=382.69, P<0.001). Notably, alcohol use was more prevalent among tobacco users (72.3%) than among non-tobacco users (58.7%), indicating that alcohol and tobacco use tend to co-occur. Despite the statistical significance, the effect size was small (Cramer’s V=0.11), demonstrating a weak practical association.
Chronic disease and tobacco use were significantly associated (χ2(1)=79.16, P<0.001). However, the effect size was small (Cramer’s V=0.05), indicating a weak association. Notably, tobacco users were less likely to report having a chronic disease (56.3%) than non-tobacco users (62.1%).
CART analysis
We conducted a CART analysis to better understand the complex patterns of tobacco use among older Korean men (Figure 2). The analysis was based on a nationally representative sample of 34,910 men aged 65 years and older. The data were divided into two groups: 70% for training the model and 30% for testing its performance. The final tree, selected using the Gini impurity measure and pruned for optimal simplicity and accuracy, retained 13 significant predictors and resulted in eight distinct end groups, or “terminal nodes.”
The model performed modestly well. It correctly identified 56.4% of tobacco users and 67.2% of non-users in the test data, with an overall accuracy of 58.6% (Tables 3, 4). It also showed fair discriminative ability, with an area under the receiver operating characteristic curve of 0.6449 and a lift of 1.57, indicating a better-than-random classification.
The tree began with age, which is the strongest predictor of tobacco use. Men aged 72.5 years or younger were more likely to smoke than those older than 72.5 years. Within this younger group, social activity was the next most influential factor. Men with low social engagement (≤1.5) were routed into higher-risk subgroups. For example, one branch (terminal node 3) had the highest tobacco use prevalence among younger men (24.8%). In contrast, younger men who engaged in more social activity were further split by age. Those ≤67.5 years (terminal node 2) had the lowest tobacco use rate among the younger group (12.9%), whereas those aged >67.5 and ≤72.5 years (terminal node 4) had a moderate prevalence of 17.7%. These findings suggest that social isolation increases the risk of tobacco use in younger men, while higher engagement and younger age were protective factors.
For men older than 72.5 years, tobacco use was generally less common; however, the risk varied depending on alcohol use. Non-drinkers (terminal node 8) had the lowest prevalence of tobacco use in the entire tree (9.3%), underscoring the protective role of abstaining from alcohol. Drinkers were further divided according to age. Those ≤78.5 years (terminal node 6) had a tobacco use prevalence of 22.0%, whereas those >78.5 years (terminal node 7) had a lower rate of 15.3%. This pattern indicates that drinking elevates risk even in later life, although the effect is somewhat reduced with advancing age.
In addition to the decision tree structure, we examined the relative variable importance (RVI) scores to identify the predictors that contributed the most to the model. RVI scores quantify the importance of each variable in classifying the outcome, scaled to a maximum value of 100 for the most influential predictor. As Table 5 shows, age had the highest RVI (100%), followed by social activity (51.1%) and alcohol use (24.4%). Other variables such as economic activity (7.1%), marital status (6.9%), perceived stress (6.5%), health literacy (6.4%), and daily exercise (5.9%) demonstrated a moderate influence on classification, whereas factors such as educational attainment, depression, and chronic disease status showed lower relative importance.
Discussion
This study used a CART model to examine multiple predictors of current tobacco use among older Korean men. The findings address our research aims by identifying discrete subgroups with elevated or reduced tobacco use risk based on interactive combinations of age, social activity, and alcohol use. The CART model validated the statistical association between these factors and tobacco use, and demonstrated how they interact to stratify risk in ways that conventional linear techniques cannot detect. Consistent with previous studies, tobacco use prevalence declined with advancing age, likely reflecting increasing health awareness, smoking cessation in response to illness, or survivor effects among older cohorts [4]. Similarly, the relationship between social engagement and health behaviors has been well-documented in the literature, although the strength and direction of the association have tended to vary across different cultural contexts. In our study, socially engaged younger men (≤72.5 years) were less likely to smoke than their socially isolated counterparts, suggesting that social participation may serve as a protective factor by fostering norms that discourage tobacco use. This finding aligns with the evidence that social connectedness supports healthier behaviors, although in other cultural settings, social gatherings may sometimes reinforce tobacco use norms [25].
A unique contribution of our CART model is that it highlights the dual roles of social activity and alcohol use across age groups. Among the younger men in our sample (≤72.5 years), low social engagement emerged as the most salient risk factor, with socially isolated individuals showing the highest prevalence of tobacco use. In contrast, alcohol use was the most influential lifestyle factor among older men (>72.5 years). Older non-drinkers represented the lowest-risk group overall, whereas older drinkers, particularly those in their early-to-mid-70s, had elevated rates of tobacco use. Although tobacco use prevalence declined again among the oldest drinkers (>78.5 years), the risk remained notably higher than that of their non-drinking peers.
Collectively, these results underscore that tobacco use in later life is shaped not by single variables but by combinations of demographic and lifestyle factors. The CART analysis revealed that risk pathways differ across age groups: (1) younger men (≤72.5 years) with limited social activity and (2) older men who drink alcohol represent distinct high-risk subgroups. These subgroup-specific patterns offer critical insights for tailoring public health strategies. Interventions that enhance social connectedness may be particularly effective for younger older men, whereas efforts to reduce tobacco use among older adults may benefit from integrating alcohol reduction strategies.
From a practice standpoint, these findings reinforce the need for targeted tobacco use cessation and prevention programs that go beyond broad age segmentation. Interventions should prioritize younger older men with limited social ties and incorporate social support–building components. Among older men, programs should address the intersection between alcohol and tobacco use, highlighting the need for integrated substance use interventions. The protective roles of social activity and health literacy also suggest that community-based health education and engagement in senior centers could provide secondary benefits for tobacco control. At the policy level, incorporating behavioral health screening into routine geriatric care may help identify older adults at elevated risk owing to combined behavioral and social vulnerabilities.
Although this study presents essential insights based on a large nationally representative sample, some limitations should be noted. First, the cross-sectional design limited our ability to draw causal conclusions. Second, the use of self-reported data for our selected variables, such as tobacco use and social activity, introduces the possibility of response bias. Third, although the CART model improves interpretability and captures interaction effects, it does not provide measures of statistical significance in the same way as traditional regression. Finally, the generalizability of the findings may be limited to older Korean men and may not extend to women or other populations without further research.
In conclusion, this study identified age, social engagement, and alcohol use as key predictors of tobacco use among older Korean men. Men aged ≤72.5 years with low social activity and those who consumed alcohol were more likely to use tobacco, whereas higher social engagement and abstaining from alcohol were associated with lower tobacco use. These findings highlight the protective roles of social connectedness and alcohol avoidance in reducing tobacco use in later life. Future studies should explore longitudinal patterns of tobacco use behavior and examine the potential mediating effects of mental health, social support, and access to cessation services. Furthermore, mixed-method studies could add depth by capturing how older adults perceive the effects of social roles, drinking habits, and health knowledge on their tobacco use behaviors.
Article Information
Conflict of interest
No potential conflict of interest relevant to this article was reported.
Funding
None.
Data availability
This study utilized secondary data of the 2023 Korea Community Health Survey. The dataset is accessible upon request via Korea Disease Control and Prevention Agency (https://chs.kdca.go.kr/chs/rawDta/rawDtaProvdMain.do).
Author contribution
Conceptualization: SSM. Data curation: JL. Formal analysis: SSM, JL. Methodology: SSM, HCH. Software: JL. Supervision: HCH. Validation: SSM, HCH. Visualization: JL. Writing–original draft: SSM, JL. Writing–review & editing: SSM, HCH, JL. Final approval of the manuscript: all authors.
Figure 1
Conceptual framework of multilevel factors associated with tobacco use.
Figure 2
Classification and regression tree (CART) analysis of risk factors associated with tobacco use among Korean older adults. Alcohol use: 0=no, 1=yes.
Table 1
Sociodemographic variables of the participants (n=34,924)
Characteristic
Value
Age (y)
73.9±6.6
Walking days per week
4.2±2.9
Perceived stress level
1.8±0.7
Depression (PHQ-9)
2.0±3.0
Social isolation
8.4±4.0
Social activity
1.0±0.9
Self-rated health
3.0±1.0
Health literacy
7.7±3.0
Education level
Less than high school
19,013 (54.5)
High school
10,985 (31.5)
College and higher
4,902 (14.0)
Economic activity status
No
17,595 (50.4)
Yes
17,327 (49.6)
Monthly household income level
Low
6,209 (17.8)
Lower-middle
8,708 (24.9)
Upper-middle
5,936 (17.0)
High
8,794 (25.2)
Marital status
Married
28,525 (81.7)
Separated
1,317 (3.8)
Single
5,077 (14.5)
Alcohol use
No
11,850 (33.9)
Yes
18,853 (54.0)
Missing
4,221 (12.1)
Chronic disease
No
13,638 (39.1)
Yes
21,282 (60.9)
Current CC use
No
28,080 (80.4)
Yes
6,842 (19.6)
Current HTP use
No
34,730 (99.4)
Yes
191 (0.5)
Current EC use
No
34,781 (99.6)
Yes
132 (0.4)
Current tobacco use (CC, HTP, or EC)
No
27,993 (80.2)
Yes
6,917 (19.8)
Values are presented as mean±standard deviation or number (%).
Values are presented as number (%) or mean±standard deviation unless otherwise stated. Percentages in the total row are based on the full sample; all other percentages are column-based.
df, degrees of freedom; PHQ-9, Patient Health Questionnaire-9.
***P<0.001 (Statistically significant).
a)By chi-square,
b)By Hedges’ g/Cramer’s V.
Table 3
Confusion matrix
Actual class
Count
Predicted class (training)
Correct (%)
Predicted class (test)
Correct (%)
1.00
0.00
1.00
0.00
1.00 (event)
4,843
3,271
1,572
67.5
1,394
680
67.2
0.00
19,601
8,531
11,070
56.5
3,658
4,734
56.4
All
24,444
11,802
12,642
58.7
5,052
5,414
58.6
Table 4
Classification performance table
Statistic
Training (%)
Test (%)
True positive rate (sensitivity or power)
67.5
67.2
False positive rate (type I error)
43.5
43.6
False negative rate (type II error)
32.5
32.8
True negative rate (specificity)
56.5
56.4
Table 5
Relevant importance variables
Variable
Relative importance (%)
Age
100.0
Social activity
51.1
Alcohol use
24.4
Economic activity
7.1
Marital status
6.9
Perceived stress level
6.5
Health literacy
6.4
Walking days per week
5.9
Education level
4.2
Depression
2.0
References
1. Lee I, Park YS, Kim JH, Han SH. Factors associated with the intention to quit smoking in elderly Korean men: the Korea National Health and Nutrition Examination Survey 2010–2015. Korean J Fam Med 2020;41:237-42.
2. Lim HK, Park YS, Kim JH, Kim JH. Factors that predict smoking cessation among older Koreans: based on the Korea National Health and Nutrition Examination Survey 2016–2018. Korean J Fam Med 2021;42:363-8.
3. Korea Disease Control and Prevention Agency. Korea national health and nutrition examination survey: the 2023 report [Internet] Korea Disease Control and Prevention Agency; 2024 [cited 2025 Jun 13]. Available from: https://knhanes.kdca.go.kr/knhanes/
4. Kim Y, Lee JS, Cho WK. Factors associated with successful smoking cessation according to age group: findings of an 11-year Korea National Survey. Int J Environ Res Public Health 2021;18:1576.
5. Kim JY, Kang HS, Jung JW, Jung SY, Park HJ, Park JS, et al. Nicotine dependence and stress susceptibility in e-cigarette smokers: the Korea National Health and Nutrition Examination Survey 2013–2017. Tuberc Respir Dis (Seoul) 2021;84:159-66.
6. Lee H, Lee HH, Kang A, Cha Y, Operario D. Psychological stress, smoking, and hazardous drinking behaviors among South Korean adults: findings from the Korean National Health and Nutrition Examination Survey. J Subst Use 2021;26:13-20.
7. Adebisi YA, Lin X, Kouwenhoven MB, Lucero-Prisno DE III. Loneliness is associated with smoking and alcohol consumption in older adults: findings from the Chinese Longitudinal Healthy Longevity Survey. J Med Surg Public Health 2024;3:100075.
9. You MA, Choi J, Son YJ. Associations of dual use of tobacco cigarettes and e-cigarettes, sleep duration, physical activity and depressive symptoms among middle-aged and older Korean adults. Nurs Open 2023;10:4071-82.
12. Stewart DW, Cano MA, Correa-Fernandez V, Spears CA, Li Y, Waters AJ, et al. Lower health literacy predicts smoking relapse among racially/ethnically diverse smokers with low socioeconomic status. BMC Public Health 2014;14:716.
13. Qiu D, Chen T, Liu T, Song F. Smoking cessation and related factors in middle-aged and older Chinese adults: evidence from a longitudinal study. PLoS One 2020;15:e0240806.
14. Breiman L, Friedman J, Olshen RA, Stone CJ. Classification and regression trees. Chapman and Hall/CRC; 1984.
15. Lemon SC, Roy J, Clark MA, Friedmann PD, Rakowski W. Classification and regression tree analysis in public health: methodological review and comparison with logistic regression. Ann Behav Med 2003;26:172-81.
21. Bronfenbrenner U. The ecology of human development: experiments by nature and design. Harvard University Press; 1979.
22. Lee SW, Han B, Cho SJ, Jung SJ, Huh Y, Kim J, et al. Associations between living alone and smoking and alcohol consumption in Korean adults. Korean J Fam Med 2020;41:306-11.
23. Park HJ, Han B, Kim B, Han K, Kim S, Kim H, et al. Relationship between smoking experience and risk of suicide mortality in South Korean adults: a nationwide population-based retrospective cohort study. J Affect Disord 2024;367:67-74.
25. Luo M, Ding D, Bauman A, Negin J, Phongsavan P. Social engagement pattern, health behaviors and subjective well-being of older adults: an international perspective using WHO-SAGE survey data. BMC Public Health 2020;20:99.
26. Feng Z, Cramm JM, Nieboer AP. Social participation is an important health behaviour for health and quality of life among chronically ill older Chinese people. BMC Geriatr 2020;20:299.
27. Kwon J, Jang J. The relationship between sports club participation, physical activity, and health behaviors among older Korean adults. Healthcare (Basel) 2024;12:2411.
28. Yun EH, Kang YH, Lim MK, Oh JK, Son JM. The role of social support and social networks in smoking behavior among middle and older aged people in rural areas of South Korea: a cross-sectional study. BMC Public Health 2010;10:78.
29. Noh JW, Kim KB, Cheon J, Lee Y, Yoo KB. Factors associated with single-use and co-use of tobacco and alcohol: a multinomial modeling approach. Int J Environ Res Public Health 2019;16:3506.
Tobacco use among older Korean men: a classification and regression tree analysis
Figure 1
Conceptual framework of multilevel factors associated with tobacco use.
Figure 2
Classification and regression tree (CART) analysis of risk factors associated with tobacco use among Korean older adults. Alcohol use: 0=no, 1=yes.
Graphical abstract
Figure 1
Figure 2
Graphical abstract
Tobacco use among older Korean men: a classification and regression tree analysis
Sociodemographic variables of the participants (n=34,924)
Characteristic
Value
Age (y)
73.9±6.6
Walking days per week
4.2±2.9
Perceived stress level
1.8±0.7
Depression (PHQ-9)
2.0±3.0
Social isolation
8.4±4.0
Social activity
1.0±0.9
Self-rated health
3.0±1.0
Health literacy
7.7±3.0
Education level
Less than high school
19,013 (54.5)
High school
10,985 (31.5)
College and higher
4,902 (14.0)
Economic activity status
No
17,595 (50.4)
Yes
17,327 (49.6)
Monthly household income level
Low
6,209 (17.8)
Lower-middle
8,708 (24.9)
Upper-middle
5,936 (17.0)
High
8,794 (25.2)
Marital status
Married
28,525 (81.7)
Separated
1,317 (3.8)
Single
5,077 (14.5)
Alcohol use
No
11,850 (33.9)
Yes
18,853 (54.0)
Missing
4,221 (12.1)
Chronic disease
No
13,638 (39.1)
Yes
21,282 (60.9)
Current CC use
No
28,080 (80.4)
Yes
6,842 (19.6)
Current HTP use
No
34,730 (99.4)
Yes
191 (0.5)
Current EC use
No
34,781 (99.6)
Yes
132 (0.4)
Current tobacco use (CC, HTP, or EC)
No
27,993 (80.2)
Yes
6,917 (19.8)
Values are presented as mean±standard deviation or number (%).
Current tobacco use by multilevel factors (n=34,924)
Variable
Tobacco use
t-value (Welch’s t-test)a)
df
Effect sizeb)
Yes
No
Total
6,917 (19.8)
27,993 (80.2)
Age (y)
71.61±5.90
74.48±6.68
35.33***
11,705.82
0.44
Walking days per week
4.02±3.08
4.27±2.85
5.99***
10,032.83
0.08
Perceived stress level
1.85±0.77
1.74±0.71
−10.57***
10,020.30
−0.15
Depression (PHQ-9)
2.17±3.15
1.95±2.96
−5.44***
10,099.08
−0.08
Social isolation
8.17±4.06
8.46±3.97
5.28***
10,390.10
0.07
Social activity
0.84±0.83
1.04±0.91
17.39***
11,361.31
0.22
Self-rated health
2.97±0.93
2.97±0.95
−0.31
10,830.80
−0.00
Health literacy
7.45±3.09
7.74±2.95
6.77***
9,279.57
0.10
Education level
56.48***
2.00
0.04
Less than high school
3,809 (55.1)
15,193 (54.3)
High school
2,317 (33.5)
8,666 (31.0)
College and higher
785 (11.4)
4,116 (14.7)
Economic activity status
56.05***
1.00
0.04
Yes
3,711 (53.7)
13,613 (48.6)
No
3,205 (46.3)
14,379 (51.4)
Monthly household income
38.07***
3.00
0.04
Low
1,387 (23.8)
4,819 (20.2)
Lower-middle
1,609 (27.6)
7,094 (29.8)
Upper-middle
1,157 (19.8)
4,777 (20.1)
High
1,676 (28.8)
7,117 (29.9)
Marital status
369.94***
2.00
0.10
Married
5,096 (73.7)
23,419 (83.7)
Separated
361 (5.2)
954 (3.4)
Single
1,458 (21.1)
3,617 (12.9)
Alcohol use
382.69***
1.00
0.11
Yes
4,443 (72.3)
14,406 (58.7)
No
1,704 (27.7)
10,137 (41.3)
Chronic disease
79.16***
1.00
0.05
Yes
3,892 (56.3)
17,380 (62.1)
No
3,025 (43.7)
10,609 (37.9)
Values are presented as number (%) or mean±standard deviation unless otherwise stated. Percentages in the total row are based on the full sample; all other percentages are column-based.
df, degrees of freedom; PHQ-9, Patient Health Questionnaire-9.
***P<0.001 (Statistically significant).
a)By chi-square,
b)By Hedges’ g/Cramer’s V.
Confusion matrix
Actual class
Count
Predicted class (training)
Correct (%)
Predicted class (test)
Correct (%)
1.00
0.00
1.00
0.00
1.00 (event)
4,843
3,271
1,572
67.5
1,394
680
67.2
0.00
19,601
8,531
11,070
56.5
3,658
4,734
56.4
All
24,444
11,802
12,642
58.7
5,052
5,414
58.6
Classification performance table
Statistic
Training (%)
Test (%)
True positive rate (sensitivity or power)
67.5
67.2
False positive rate (type I error)
43.5
43.6
False negative rate (type II error)
32.5
32.8
True negative rate (specificity)
56.5
56.4
Relevant importance variables
Variable
Relative importance (%)
Age
100.0
Social activity
51.1
Alcohol use
24.4
Economic activity
7.1
Marital status
6.9
Perceived stress level
6.5
Health literacy
6.4
Walking days per week
5.9
Education level
4.2
Depression
2.0
Table 1
Sociodemographic variables of the participants (n=34,924)
Values are presented as mean±standard deviation or number (%).
Table 2
Current tobacco use by multilevel factors (n=34,924)
Values are presented as number (%) or mean±standard deviation unless otherwise stated. Percentages in the total row are based on the full sample; all other percentages are column-based.
df, degrees of freedom; PHQ-9, Patient Health Questionnaire-9.
