This population-based cross-sectional study used data from the 2017 Canadian Tobacco, Alcohol, and Drugs Survey (CTADS) that included 16,349 Canadian residents aged 15+ (excluding institutional residents) from ten provinces (excluding Yukon, Northwest Territories and Nunavut) [26]. The CTADS is a well-validated survey with a range of studies being published using the 2017 data [27–30]. We used the question, “Have you ever tried an electronic cigarette, also known as e-cigarette?” to identify all of the 3,059 respondents with an experience of vaping.

A binary outcome variable was created to represent vaper type (current vs. former vaper). Respondents were defined as a current vaper if they answered, “every day” or “occasionally” to the question, “At the present time, do you use an electronic cigarette, also known as an e-cigarette every day, occasionally or not at all?” People responding “not at all” to the same question were former vapers.

A wide range of person-level characteristics were explored as potential correlates of vaper type. These 46 variables were mostly categorical, except for age and years of smoking that were continuous. These variables described demographics, socioeconomic factors, household information, health, vaping behaviours, substance use and perceived risk of vaping and smoking (see below).

We summarized the characteristics of current vs. former vapers and used two-sample tests (Fisher’s exact test, t-test or the Mann-Whitney U-test) to compare their distributions.

Variable selection has been shown to be a necessary procedure prior to classification tree analysis to reduce the risk of model overfitting and spurious associations [20]. Hence, we used each candidate correlate to predict the odds of being a current vaper in a logistic mixed-effects model with a random intercept indicating provincially based random effects. This model estimates an unadjusted fixed effect for each correlate on all individuals and allows this effect to vary across provinces. Correlates associated with a significant fixed effect (using a 2-sided P-value < 0.05) were selected into the machine learning analysis. This procedure was performed on R using the “lme4” package [31].

We followed the Classification and Regression Tree (CART) algorithm to develop a tree model to classify current and former vapers [32] using the R package “rpart” [33]. CART is a non-parametric method that develops a classification tree by recursive partitioning. In this tree, a node is where a splitting variable, X, and one of its level of value, c, divides the dataset into two regions, X ≤ c and X > c (or X = 0 and X = 1 for binary X), that correspond to the predicted classes of current and former vapers. An optimal splitting variable X and value c minimize the Gini Index at a node, which is an impurity criterion that measures how well a split correctly separates true current vapers from true former vapers (i.e. how “pure” the separation is). The splitting procedure terminates when a node contains < 5 data. “Pruning” of a tree is often necessary as a full tree may be excessively large and complicated that leads to model overfitting. Hence, a cross-validation method is used to identify an optimal number of splits that minimizes a cost complexity function of total misclassified cases with a penalized term for larger tree size. This procedure yields a tree with manageable size and interpretable structure while maintaining its performance.

We used a ratio of 9:1 to randomly split the dataset into a training set (n = 2,753) and a validation set (n = 306). The training set was used to develop and to internally validate the model, while the validation set was used to establish model performance externally as it comprised independent data not used in model construction. Using the data from the training set, we first performed oversampling as the number of former vapers significantly exceeded that of current vapers (with a ratio of 4:1), causing concerns on outcome imbalance that may deter model performance. Hence, a random oversampling with replacement was conducted on current vapers on the training set so that their size was increased to be that of former vapers. We then developed a full tree using data from the oversampled training set to classify current and former vapers with the set of correlates selected from the logistic mixed-effects regression analysis. After that, the pruning procedure was performed using a ten-fold cross-validation method. Classification accuracy, specificity and sensitivity of the pruned tree were calculated during the cross-validation process to establish performance of the model on the training set. Finally, the pruned tree was applied to data from the validation set and accuracy, specificity and sensitivity were computed to demonstrate the external performance of the model.

We assessed the performance of two parsimonious trees, including the one that only had the first split of the full tree and another one with the first two splits. This procedure was adopted from a recent machine learning paper that also used a classification tree to predict smoking cessation status in efforts of quantifying the significance of the top predictors in this model [20]. For both parsimonious tree models, we calculated their accuracy, sensitivity and specificity using data from the oversampled training set and from the validation set separately.

Multiple imputation by chained equation [34] was used to address the very small portion of data missing from the dataset (totaled 1.0%). After visually confirming the assumption of missing at random, five imputed data copies were generated independently, and all analytical procedures were repeated on each of these data copies to compare results with our primary findings (Supplementary Material 1). Analyses were performed on R (version 3.5.1).

Of the 3,059 Canadians aged 15+ who reported to have tried vaping, their average age was 24.4 [standard deviation (SD) 11.0] years, with 41.9% of them being female, 8.5% of them being aboriginal and 74.3% residing in urban areas (Table 1). A total of 2,503 (81.8%) were former vapers and 556 (18.2%) reported to be current vapers.

The two groups of vapers differed significantly in their characteristics. Notably, current vapers were younger by 2-year on average (mean age = 22.5 vs. 24.9 years) and had lower education (high school or above: 60.4% vs. 79.9%). They were less likely to be female (37.2% vs. 42.9%), currently working (61.2% vs. 68.6%), having been or currently married (13.7% vs. 19.0%) or living with a child (27.7% vs. 23.1%). The most striking difference between the two groups was their perceived risk of vaping. Current vapers were less likely to consider vaping occasionally (22.5% vs. 33.8%), regularly (55.0% vs. 70.2%) or during pregnancy (78.1% vs. 84.7%) to have moderate/great risks rather than no/slight risks.

Twenty-nine correlates were identified by the logistic mixed-effects regression analysis to be potentially important (Table 2; full results see Supplementary Material 2).

Using these correlates, a classification tree (Figure 1) was developed and pruned with a final form comprising just three predictors-attraction to vaping flavor (yes/no), age (with 18-years being chosen as the optimal threshold; age < 18 or age ≥ 18) and believing vaping was less harmful than smoking to oneself (yes/no). Using cross-validation, the accuracy, sensitivity and specificity of this tree model on the training set was 0.71 (95% CI 0.65–0.77), 0.70 (95% CI 0.61–0.76) and 0.71 (95% CI 0.66–0.73), respectively. Applying this model to data from the validation set yielded accuracy, sensitivity and specificity of 0.72, 0.65 and 0.73.

The tree model used attraction to flavor as the first splitting variable, followed by age as the second splitting variable and believing vaping to be less harmful than smoking to oneself as the third splitting variable. People who reported to vape due to flavors were predicted to have the highest probability (0.64) of being a current vaper. Among those who vaped for other reasons, minors with ages < 18 had the second highest probability (0.63) of vaping currently. For adults who did not vape for flavors, their probability of current vaping could reach a high of 0.60 if they believed vaping to be less harmful than cigarette smoking to users and otherwise was a low of 0.26 if they did not have such health belief.

In order to understand the importance of the top two correlates (attraction to flavor and age), we compared the performance of the full tree to two parsimonious trees that comprised only the first split (attraction to flavor) or the first two splits (attraction to flavor and age; Table 3). Using data from the oversampled training set, classification accuracy and specificity were the highest in the full tree, but sensitivity was the highest in the 2-split tree that used only attraction to flavor and age for prediction (sensitivity = 0.74 vs. 0.70). Similar results were observed on the validation set where the 2-split tree exceeded the full tree in terms of sensitivity (sensitivity = 0.67 vs. 0.65). However, in general, improvement of model performance from a 1-split tree to a 2-split tree to the full tree was minor (Table 3).

Five imputed data copies were generated using the multiple imputation by chained equation method after visual inspection confirmed the assumption of missing at random (Supplementary Material 1). All analytical procedures were repeated on the five imputed datasets and the same classification tree involving attraction to flavor (first-split), age (second-split) and believing vaping to be less harmful than smoking to oneself (third-split) was reached at each iteration. Hence, we conclude the tree model is generally insensitive to missing data.

Due to the cross-sectional nature of the survey data, we were unable to identify true predictors of current vaping, but rather important correlates. Future study with access to longitudinal data could outcome this limitation. Next, our analysis depends entirely on self-reported measures, which may introduce recall bias. However, we believe that the CTADS survey mechanism has been carefully constructed to ensure sufficient answer time for interviewees to adequately recall long-term memory. Third, there are other factors that we do not have access to, such as household income and neighborhood characteristics, that may influence the pattern of vaping. Future researchers with a more comprehensive tracking of people with vaping experiences, preferably through the use of linked administrative dataset, may provide additional insights. Finally, the data used for this study was collected in 2017, which was before the enactment of the Tobacco and Vaping Products Act (in May 2018) [43] and the legalization of recreational cannabis (in October 2018) [44] in Canada. Future researchers may leverage more recent data to explore the impact of these new regulations on vaping behaviours.

In conclusion, by using a classification tree, we identified attraction to flavors to be one of the most important correlates of current vaping. This finding is relevant to future development of regulations on e-cigarettes as most flavored vaping products are still legal in Canada. Furthermore, interventions that target youths are needed to prevent their e-cigarette uptake and help those who have already initiated vaping to quit.