Study design and setting

This cross-sectional study was conducted from August to October 2025 in the provinces of rural northeastern Thailand. The study included 1020 participants aged ≥16 years recruited through multi-stage cluster random sampling.

Sample size calculation

Sample size calculation assumed e-cigarette prevalence of 70.0% in exposed versus 55.0% in unexposed groups, with 80.0% power and α=0.05, yielding a required minimum of 894 participants plus a 15.0% non-response buffer¹³, resulting in a final sample of 1020 participants.

Sampling procedure

Multi-stage cluster sampling was conducted as follows: Stage 1 – Three provinces in northeastern Thailand were purposively selected based on high e-cigarette prevalence rates; Stage 2 – Within each province, three districts were randomly selected using probability proportional to size; Stage 3 – Two sub-districts were randomly selected from each district; and Stage 4 – Within each sub-district, households were systematically sampled using a random start point and fixed interval until the target sample size was reached. One eligible participant per household was selected using the Kish method^14,15.

Data collection instruments

Data were collected using validated, self-administered questionnaires comprising: 1) Sociodemographic characteristics (age, gender, education level, income); 2) E-cigarette use patterns (current use, frequency, duration); 3) Alcohol consumption (frequency and quantity); and 4) Knowledge about e-cigarettes, with scores categorized as low (<60%), moderate (60.0–79.0%), and high (≥80.0%). The scale demonstrated good internal consistency (Cronbach’s α=0.82) and content validity index of 0.89 as assessed by five experts; 5) Danger perception – assessed using an 8-item Likert scale (α=0.79), content validity index 0.86; 6) Family support and social support – measured using adapted Multi-dimensional Scale of Perceived Social Support with good reliability (α=0.85 and 0.83 respectively); and 7) Perceived law enforcement – single item assessing perception of enforcement strictness.

All instruments were originally developed in English and translated into Thai using forward-backward translation by two independent bilingual experts. The Thai versions underwent content validation by a panel of five experts in tobacco control and public health. Pilot testing was conducted with 30 participants to assess clarity and comprehension. Permission was obtained from the original scale developers for adaptation and use. The response rate was 89.2% (1020 completed questionnaires from 1143 eligible participants approached).

Statistical analysis

Variable definition

The dependent variable was current e-cigarette use (yes, no). Independent variables included: gender (male, female, LGBTQA+), age (16–39, 40–59, 60–69, ≥70 years), alcohol consumption (yes, no), family support (low, moderate, high), social support (low, moderate, high), knowledge about e-cigarettes (low, moderate, high), danger perception (poor, moderate, good), and perceived law enforcement (strict, slack). Potential confounding variables were identified a priori based on literature review and included all demographic and psychosocial factors measured.

Analytical methods

Statistical analyses used both traditional multivariable logistic regression and propensity score (PS) methods. Crude and adjusted odds ratios (AORs) with 95% confidence intervals (CIs) were calculated for associations between exposures and e-cigarette use. For key exposures, PS statistics were estimated using logistic regression including all measured confounders¹⁶. Four complementary PS methods were applied: 1) one-to-one nearest-neighbor matching with a 0.2 standard deviation caliper¹⁷; 2) inverse probability of treatment weighting with stabilized weights¹⁸; 3) PS stratification into quintiles¹⁹; and 4) PS covariate adjustment²⁰. Doubly robust estimation combining PS weighting and outcome regression provided additional robustness. Covariate balance was assessed using standardized mean differences (SMD), with SMD <0.10 indicating adequate balance²¹, and visualized through distribution plots. Participants outside the common support region (PS: 0.15–0.92) were trimmed to ensure positivity assumption.

Sensitivity analysis consisted of E-value calculation to quantify the minimum strength of unmeasured confounding required to nullify observed associations²², which was computed as E-value = OR + √[OR × (OR - 1)] for both point estimates and confidence limits. Subgroup analyses, stratified by gender, age, and knowledge level, explored effect heterogeneity, with interaction terms tested at p<0.10. Multiple imputation by chained equations (20 datasets) addressed potential missing data bias. Clinical utility measures including number needed to harm and risk differences supplemented the ORs for meaningful effect interpretation. Model performance was evaluated using C-statistics, Hosmer-Lemeshow tests, and Brier scores. All analyses were conducted using R version 4.3.0. Statistical significance was tested at two-tailed p<0.05.

Quality assurance

Quality assurance procedures included double data entry with discrepancy resolution, logical consistency checks, random re-interviewing of 10.0% of participants for verification, and regular supervisory monitoring during data collection²³. Trained research assistants conducted household visits ensuring participant privacy during questionnaire completion.

Ethical considerations

This study was conducted in accordance with the Declaration of Helsinki and ethical principles for human research. All participants provided written informed consent prior to participation in the study, with procedures clearly explained in the local language. The study received institutional ethical approval from the Institutional Review Board of Human Research Ethics Subcommittee, Rajamangala University of Technology Isan Sakon Nakhon Campus, Thailand (No. HEC-04-68-020). Participant confidentiality and anonymity were maintained throughout the research process; participants were informed of their right to withdraw from the study at any time without penalty.

Baseline characteristics of participants

In total, 1020 participants were included in this study, with females comprising the majority (49.8%), followed by males (35.2%) and LGBTQA+ individuals (15.0%). The mean age was 54.8 years (SD=13.12), with participants predominantly in the age group 40–59 years (49.0%).

More than one-half of the participants (54.1%) reported high levels of social support, while 42.6% had moderate support. Knowledge about e-cigarettes was variable, with nearly one-half (49.7%) having moderate knowledge and 13.0% having low knowledge. Most participants (51.9%) demonstrated high perception toward the dangers of e-cigarettes, while 42.8% had moderate perception and only 5.3% had low perception.

Analysis of different types of vaping behavior revealed that 54.9% of the participants used both cigarettes and e-cigarettes, with an equal proportion vaping after eating. Nearly one-half (48.2%) reported vaping with alcohol, while 35.0% vaped when under stress. Bars were the most common venue for vaping (92.3%), followed by vaping everywhere when the opportunity arose (20.2%). The primary reason for vaping was peer invitation (59.7%), followed by stress from school or work (39.3%) and the perception of vaping as cool or stylish (36.4%) (Table 1).

Analysis of factors associated with e-cigarette vaping behavior

Gender was the strongest predictor of e-cigarette use. Males demonstrated 5.29-fold (AOR=5.29; 95% CI: 3.56–7.89) and females 4.59-fold (AOR=4.59; 95% CI: 3.02–6.96) higher likelihood of e-cigarette vaping compared to the LGBTQA+ group (p<0.001). Alcohol consumption exhibited a significant association, with drinkers showing 2.05-fold higher likelihood of e-cigarette vaping compared to non-drinkers (AOR=2.05; 95% CI: 1.53–2.74, p<0.001). Perception of lax law enforcement increased vaping likelihood by 1.52-fold (AOR=1.52; 95% CI: 1.10–2.10, p=0.015).

Individuals with low danger perception had 2.45-fold higher likelihood of vaping than those with high perception (AOR=2.45; 95% CI: 1.02–5.89, p=0.043). Knowledge about e-cigarettes displayed a U-shaped pattern, with individuals with low knowledge having 1.95-fold higher likelihood of vaping compared to those with high knowledge (AOR=1.95; 95% CI: 1.10–3.45, p=0.012). There was no significant difference between those with moderate knowledge and the high knowledge group (AOR=0.87; 95% CI: 0.63–1.20, p>0.05).

Family support showed a dose-response relationship, with individuals receiving low support demonstrating 1.98-fold higher likelihood of vaping compared to those with high support (AOR=1.98; 95% CI: 1.22–3.22, p=0.045). Vaping prevalence increased progressively across support levels (high: 72.2%, moderate: 76.8%, low: 84.5%). Individuals with high social support had 2.05-fold higher likelihood of vaping than those with low support (AOR=2.05; 95% CI: 1.99–4.23, p<0.001), and those with moderate support had 2.95-fold higher likelihood (AOR=2.95; 95% CI: 1.49–4.06, p<0.001) (Table 2).

Propensity score model performance and covariate balance

The propensity score model for alcohol consumption exposure demonstrated acceptable discrimination (C-statistic=0.672; 95% CI: 0.640–0.704) and good calibration (Hosmer-Lemeshow χ²=8.94, p=0.347). Gender emerged as the strongest predictor of alcohol consumption, with males (OR=2.27; 95% CI: 1.56–3.30, p<0.001) and females (OR=1.57; 95% CI: 1.04–2.37, p=0.032) having significantly higher likelihood than LGBTQA+ individuals. Family support and social support also predicted alcohol consumption patterns.

Before propensity score matching, substantial baseline imbalances existed between drinkers and non-drinkers, particularly for gender (SMD=0.328), social support (SMD=0.251), and family support (SMD=0.217). After 1:1 nearest-neighbor matching with a caliper width of 0.2 standard deviations, excellent covariate balance was achieved across all measured confounders (all SMD <0.10, all p>0.05), creating comparable groups. The matched sample comprised 796 participants (398 pairs), representing 78.0% of the original cohort. Common support assessment revealed that 979 participants (96.0%) fell within the propensity score overlap region (0.15–0.92), with only 41 participants (4.0%) trimmed due to extreme propensity scores, ensuring the positivity assumption was satisfied (Table 3).

Effect of alcohol consumption on e-cigarette use across multiple methods

Consistency was observed regarding the association between alcohol consumption and e-cigarette use across all analytical approaches. In the original unadjusted analysis, alcohol drinkers showed 82.6% e-cigarette use prevalence compared to 69.2% among non-drinkers (OR=2.11; 95% CI: 1.59–2.80, p<0.001). Traditional multivariable logistic regression yielded an AOR of 2.05 (95% CI: 1.53–2.74, p<0.001) after controlling for all measured confounders. Propensity score methods produced estimates as follows: PS matching, OR=1.77 (95% CI: 1.30–2.41); inverse probability of treatment weighting, OR=1.87 (95% CI: 1.38–2.54); PS stratification, OR=1.81 (95% CI: 1.34–2.45); PS covariate adjustment, OR=1.92 (95% CI: 1.43–2.58); and doubly robust estimation, OR=1.85 (95% CI: 1.37–2.50). All methods achieved statistical significance (p<0.001), with effect estimates ranging from 1.77 to 2.05 (Table 4).

Sensitivity analysis and robustness to unmeasured confounding

E-value analysis quantified the robustness of the observed associations to potential unmeasured confounding. Gender effects demonstrated E-values of 9.99 for males versus LGBTQA+ (AOR=5.29) and 8.63 for females versus LGBTQA+ (AOR=4.59). Alcohol consumption showed an E-value of 3.53 (AOR=2.05). Family support demonstrated an E-value of 3.41, and social support showed an E-value of 3.53. Danger perception showed an E-value of 4.32, and law enforcement perception demonstrated an E-value of 2.48. Knowledge about e-cigarettes showed an E-value of 3.35 (Table 5).

Effect modification by gender and clinical utility measures

Subgroup analysis using the propensity score matched sample revealed significant effect modification by gender (interaction p=0.042). Among males, alcohol consumption demonstrated an association with e-cigarette use (OR=2.89; 95% CI: 1.51–5.53, p=0.001), with prevalence of 91.2% among male drinkers compared to 78.3% among male non-drinkers. The effect was attenuated in females (OR=1.62; 95% CI: 0.98–2.68, p=0.061) and LGBTQA+ individuals (OR=1.87; 95% CI: 0.99–3.53, p=0.053). Clinical utility measures derived from the inverse probability of treatment weighting revealed a risk difference of 11.2% (95% CI: 6.8–15.6) and number needed to harm of 9 (95% CI: 6–15) (Supplementary file Table 1).

Consistency across multiple exposures using propensity score methods

Comparative analysis across all key exposures demonstrated consistency between traditional regression and propensity score methods. For males versus LGBTQA+, estimates ranged from OR=4.98 (PS matching) to OR=5.29 (regression); for females versus LGBTQA+, from OR=4.35 to OR=4.59. Family support showed effects ranging from OR=1.85 to OR=1.98 for low versus high support. Social support findings ranged from OR=1.92 to OR=2.05. Knowledge effects ranged from OR=1.82 to OR=1.95 for low versus high, and danger perception effects ranged from OR=2.21 to OR=2.45 for poor versus good. Law enforcement perception demonstrated effects ranging from OR=1.46 to OR=1.52 for slack versus strict (Supplementary file Table 2).

Principal findings and international context

This study identified gender, alcohol consumption, family support, and psychosocial factors as significant determinants of e-cigarette use among rural Thai people. The findings were corroborated through multiple propensity score methods. The gender disparity observed, with males and females demonstrating substantially higher likelihood compared to LGBTQA+ individuals, aligned with emerging evidence of differential vulnerability across populations documented internationally²⁴. The lower prevalence among LGBTQA+ individuals warrants further investigation into protective factors operating within this community²⁵.

The association between alcohol consumption and e-cigarette use documented in the current study exemplifies the clustering of risk behavior phenomena extensively reported in the literature on substance use. This finding demonstrated consistency across analytical methods, from traditional regression through the four propensity score approaches, with E-value analysis indicating moderate robustness to unmeasured confounding. This co-occurrence pattern reflects broader evidence of behavioral clustering documented internationally²⁶. The significant gender interaction identified in subgroup analysis suggests that targeted interventions should be gender-stratified.

The relationship between social support and e-cigarette use, wherein high social support showed increased associations, represents a finding requiring interpretation²⁷. This finding persisted across all propensity score methods. The persistence after achieving excellent covariate balance indicates that social networks may simultaneously provide support while transmitting health-risk behaviors¹⁹.

Family support demonstrated a protective dose-response relationship. This gradient strengthens confidence in a potential association and aligns with theoretical models emphasizing family-level protective factors²⁸. The distinction between family and social support effects, suggested that family-level interventions may operate through different mechanisms than peer-based approaches²⁹.

Risk perception and knowledge patterns

The U-shaped relationship between knowledge and e-cigarette use suggests that basic health literacy may provide substantial protection¹⁷. The finding that individuals with poor danger perception had higher likelihood of vaping underscores the risk perception-behavior gap documented across substance use research.

Policy implications and law enforcement perceptions

The finding that perceived lax law enforcement was associated with increased e-cigarette use supported deterrence theory but revealed implementation challenges in Thailand’s regulatory framework^5,6,10. Despite their legal prohibition since 2014^3,10, e-cigarettes remain widely accessible through online platforms and cross-border sales. The E-value analysis for law enforcement perception suggested moderate vulnerability to unmeasured confounding.

Methodological contributions

The primary methodological contribution of the current study lies in applying comprehensive propensity score methods to tobacco research in a low- and middle-income country context, consistent with recent advances in observational epidemiology. The consistency of effect estimates across traditional regression and the four propensity score approaches, strengthens confidence in the current findings^28,30. The explicit demonstration of covariate balance achievement created pseudo-randomization approximating experimental conditions within observational constraints. The E-value analyses quantified robustness to unmeasured confounding.

Strengths and limitations

The strengths of this study include the use of multiple propensity score methods to strengthen the robustness of findings, comprehensive sensitivity analyses through E-value calculations, and explicit assessment of covariate balance. The propensity score methods offered substantial advantages over conventional adjustment techniques by creating pseudo-randomization through explicit covariate balancing^17,19, mimicking experimental conditions within observational constraints, and enabling transparent assessment of assumptions^17,19,20. This methodological advancement was essential for rural Thailand, where randomized controlled trials are impractical due to ethical constraints regarding substance use exposure assignment.

However, important limitations warrant acknowledgment. The cross-sectional design precluded definitive temporal inference despite advanced analytical methods. Propensity score methods only balance measured covariates; unmeasured confounders such as nicotine dependence severity, mental health disorders, and genetic susceptibility may bias estimates^28,30. Self-report bias may lead to underreporting of socially undesirable behaviors^31,32. Additionally, self-reported data may introduce information bias and misclassification^32,33. Residual confounding from unmeasured variables cannot be entirely ruled out.

Future research should apply longitudinal designs with comprehensive confounding measurement, validated biomarkers of tobacco exposure, and mixed-method approaches integrating quantitative epidemiology with qualitative investigation of social contexts and individual motivations^33,34.