The Relationship Between Electronic Cigarette Use and Conventional Cigarette Smoking Is Largely Attributable to Shared Risk Factors

Abstract

Introduction

The growing popularity of electronic cigarettes (e-cigarettes) among youth raises concerns about possible causal effects on conventional cigarette smoking. However, past research remains inconclusive due to heavy confounding between cigarette and e-cigarette use. This study uses propensity score methods to robustly adjust for shared risk in estimating the relationship between e-cigarette use and conventional smoking.

Methods

Cross-sectional data from 8th and 10th graders were drawn from the 2015–2016 waves of Monitoring the Future (n = 12 421). The effects of (1) lifetime and (2) current e-cigarette use on (A) lifetime and (B) current conventional cigarette smoking were examined using logistic regression analyses with inverse propensity weighting based on 14 associated risk factors.

Results

After accounting for the propensity for using e-cigarettes based on 14 risk factors, both lifetime and current e-cigarette use significantly increased the risk of ever smoking a conventional cigarette (OR = 2.49, 95% CI = 1.77 to 3.51; OR = 2.32, 95% CI = 1.66 to 3.25, respectively). However, lifetime (OR = 2.17, 95% CI = 0.62 to 7.63) and current e-cigarette use (OR = 0.95, 95% CI = 0.55 to 1.63) did not significantly increase the risk of current conventional cigarette smoking.

Conclusions

E-cigarette use does not appear to be associated with current, continued smoking. Instead, the apparent relationship between e-cigarette use and current conventional smoking is fully explained by shared risk factors, thus failing to support claims that e-cigarettes have a causal effect on concurrent conventional smoking among youth. E-cigarette use has a remaining association with lifetime cigarette smoking after propensity score adjustment; however, future research is needed to determine whether this is a causal relationship or merely reflects unmeasured confounding.

Implications

This study examines the relationship between e-cigarette use and conventional smoking using inverse propensity score weighting, an innovative statistical method that produces less-biased results in the presence of heavy confounding. Our findings show that the apparent relationship between e-cigarette use and current cigarette smoking is entirely attributable to shared risk factors for tobacco use. However, e-cigarette use is associated with lifetime cigarette smoking, though further research is needed to determine whether this is a causal relationship or merely reflects unaccounted-for confounding. Propensity score weighting produced significantly weaker effect estimations compared to conventional regression control.

Introduction

The use of electronic cigarettes (e-cigarettes) has risen drastically between 2011 and 2018 among adolescents, with a corresponding decline in conventional cigarette smoking.¹ In fact, e-cigarette use is more prevalent today among youth than adults.² In 2018, 20.8% of high school students and 7.2% of middle school students were current users, showing that e-cigarette use is now even more common and faster growing than conventional smoking.^1,3 However, at least part of the apparent increases in e-cigarette use may be due to questionnaire design changes,⁴ raising uncertainty about the exact prevalence and implications of e-cigarettes.

The health-related implications of these recent trends remain unclear. Aside from possible toxic effects of e-cigarettes themselves,⁵ the growing popularity of e-cigarette in youth on one hand has raised other significant concerns. A specific concern is that e-cigarettes might act as a “gateway” or causal factor leading to conventional cigarette smoking. In support of this, about half of adolescents who used a tobacco product reported using multiple tobacco products.⁶ On the other hand, some are optimistic: e-cigarettes may aid in reducing or quitting conventional cigarette smoking,^7–9 therefore, resulting in substantial harm reduction.^10–12

A recent report by the National Academies that presented a systematic review concluded that there is substantial evidence for e-cigarettes having a causal effect on increasing conventional smoking in youth and young adults.¹³ Smoking in adolescence often continues into adulthood,¹⁴ which leads to increased health risks including various cancers and lung diseases.¹⁵ Therefore, establishing the relationship between e-cigarette use and subsequent conventional cigarette smoking will have a critical impact on public health by directing future tobacco-related interventions and policies. However, nearly all of the studies included used conventional approaches to confounding (namely, controlling for risk factors in a regression),¹³ which are known to be biased in the presence of heavily confounded data.^16,17 This is the case with e-cigarette use and conventional cigarette smoking: e-cigarette users are more similar to conventional cigarette smokers than they are to nontobacco users in terms of demographics, smoking-related mediators, and behavioral characteristics.^18,19 These shared characteristics may serve as a common liability for any tobacco use: e-cigarette use seemingly causing conventional cigarette smoking may, in fact, merely be due to the user’s predisposition to tobacco use. In accounting for this possibility, existing research that has attempted to estimate the effect of e-cigarette use on conventional smoking most likely contains residual bias due to confounding.^16,17,20 A result of this methodological bias is that it remains unclear whether e-cigarettes attract new youth who would not have otherwise used tobacco products,²¹ or whether tobacco users are more often initiating smoking behaviors with e-cigarettes instead of conventional cigarettes.²²

Propensity score methods offer an advantage over traditional methods in minimizing the likelihood of confounding, even in observational data.¹⁶ For example, inverse propensity weighting (IPW) weights each individual according to their “propensity” for receiving the “treatment” (using e-cigarettes) according to a set of confounding variables. Individuals that resemble the other “treatment” group are weighted heavily, as these provide a more appropriate comparison. In this way, propensity score analyses approximate a randomized design, provided all relevant confounders are measured.¹⁶

This study examines the relationship between e-cigarette use and conventional cigarette smoking using IPW to rigorously adjust for shared risk factors. We estimate the causal effects of (1) lifetime and (2) current e-cigarette use on concurrent measures of (A) lifetime and (B) conventional cigarette smoking, after IPW based on 14 risk factors. Data were extracted from Monitoring the Future (MTF) surveys²³ conducted in 2015 and 2016 on 8th and 10th graders.

Methods

Sample

Data were drawn from the MTF study, years 2015–2016. MTF was selected for this study because of the wide range of confounding variables that are available: due to the importance of including as many confounders as possible in IPW, this consideration outweighed the drawback of the cross-sectional nature of the data for this study. MTF is an ongoing, cross-sectional, nationally representative study about drug use, behavior, attitudes, and values of young people. Owing to availability of questions about e-cigarettes, only data from 2015 to 2016 from 8th to 10th graders, from form 2 of the questionnaire, were included in this study. The response rate was 88%–90% in 2015 and 87%–89% in 2016. The sample included a total of 21 343 respondents. Omitting missing values for ever and current use of electronic and conventional cigarettes reduced the sample size to n = 19 535. Omitting missing data on all core questions and covariates (described later) resulted in a final analytic sample size of n = 12 421.

Measures

E-cigarette use was self-reported with five possible response options to a question about lifetime using: never, 1–2 times, occasionally, regular in the past, or regular now. This was recoded into two separate dichotomous variables: (1) lifetime e-cigarette use (never vs. at least once in lifetime), and (2) current e-cigarette use (regular now vs. anything less).

Conventional cigarette use was also measured in the same way using five-option self-reports about lifetime smoking. Responses were also recoded into two dichotomous variables: (1) lifetime conventional cigarette smoking (never vs. at least once in lifetime), and (2) current conventional cigarette smoking (regular now vs. anything less).

For demographics, grade (8th vs. 10th grade), race, and sex (male or female), were self-reported. The race was collected as follows: black or African American, Mexican American, Cuban American, Puerto Rican, Other Hispanic or Latino, Asian American, white (Caucasian), American Indian or Alaska Native, Native Hawaiian, or Other Pacific Islander. For this study, these responses were recoded into four levels: white, black, Hispanic, and other.

Alcohol and marijuana use were treated as separate covariates due to their relatively common usage among respondents. These variables were assessed as the number of substance use occasions in last 30 days (“On how many occasions have you had alcoholic beverages to drink (more than just a few sips)/marijuana or hashish during the last 30 days?’’). Binary variables for any versus no current use of alcohol and marijuana were made based on each response.

The use of other illicit substances (Lysergic acid diethylamide, hallucinogen, methamphetamine, tranquilizer, inhalant, crack, 3,4-methylenedioxy-methamphetamine, steroid) were combined into a single variable due to the very low frequency of each. Each substance was originally assessed separately as the number of substance use occasions in the lifetime. For this study, these were combined into a single dichotomous variable indicating the use of any versus none of these substances in the lifetime.

Perceived peer cigarette use was assessed by asking participants what proportion of friends they think smoke. Responses were given on a 5-point scale: 0 (none), 1 (a few), 2 (some), 3 (most), and 4 (all).

Exposure to health warnings about cigarettes was assessed with a dichotomous question about whether the respondent has ever noticed the warning sign on a cigarette pack.

Discipline problems were represented as the frequency of respondents getting disciplined (getting sent to the office or having to stay after school) for their misbehavior, and responses were collected on a 5-point scale: 0 (never), 1 (seldom), 2 (sometimes), 3 (often), and 4 (always). This variable was further categorized into three levels to achieve balance between two groups: 0 (never), 1 (seldom or sometimes), 2 (often or always).

Risk-taking behavior was examined with how much the respondent agrees with the statement “I like new and exciting experiences, even if I have to break the rules.” Responses were collected on a 5-point scale from 1 (disagree) to 5 (agree).

Positive mood was assessed by how much respondents feel happy these days, with answers reported on a 3-point scale: 1 (not happy), 2 (pretty happy), and 3 (very happy).

Dissatisfaction of being near smokers was evaluated with how much respondents agree with the statement “I strongly dislike being near people who are smoking.” Responses were collected on a 5-point scale ranging from 1 (disagree) to 5 (agree).

Disapproval of smoking behavior assessed as how much respondents disapprove of people smoking one or more packs of cigarettes per day. Responses were collected on a 3-point scale: 1 (don’t disapprove), 2 (disapprove), and 3 (strongly disapprove).

Paternal education level was collected as the highest level of schooling the respondent’s father had finished. Responses were given on a 6-level scale: 1 (completed grade school), 2 (some high school), 3 (high school), 4 (some college), 5 (college), and 6 (graduate or professional school after college).

Analyses

Several analyses were run examining the cross-sectional association between e-cigarette use and conventional cigarette use. First, survey-weighted, unadjusted logistic regression models were used to examine outcomes of (1) lifetime and (2) current conventional cigarette use each as a function of (A) lifetime and (B) current e-cigarette use (four separate analyses). Survey weighting was conducted using the “survey” package²⁴ in R software.²⁵

Possible confounders from prior literature were matched with questions asked in the MTF survey. MTF covered demographics, smoking perception, exposure to health warning, substance use, discipline problems, risk-taking behaviors, and mental health. For risk categories which had multiple questions available (smoking perception, discipline problems, risk-taking behavior, mental health, and parental education), preliminary weighted regression was performed. The variable that has highest coefficient on regression model was considered to be most representative of the group and, therefore, selected for final analyses.

In order to reduce extreme dissimilarity between treatment and control groups prior to IPW analysis, covariate balance between the treatment and control groups was examined using “cobalt” package.²⁶ Preliminary analyses of balance used the original forms of the variables, and in cases of a substantial imbalance (standardized mean difference >0.2), the corresponding variable was categorized or dichotomized until balance was achieved. The final set of 14 covariates were the following: (1) grade, (2) race, (3) sex, (4) current alcohol use, (5) current marijuana use, (6) lifetime use of any other illicit substances, (7) perceived peer smoking, (8) exposure to health warnings on cigarette packaging, (9) discipline problems, (10) risk-taking behavior, (11) positive mood, (12) disapproval of smoking, (13) dissatisfaction being near people smoking, and (14) highest level of father’s education.

Next, survey-weighted, confounder-adjusted logistic regressions were repeated for the effect of (1) lifetime and (2) current e-cigarette use on outcomes of (A) lifetime and (B) current conventional cigarette smoking. This was done to compare results of traditional confounding adjustment with those of IPW.

Finally, IPW based on propensity scores was used to estimate the significance and the strength of the associations. The propensity score was calculated from a logistic regression analysis of e-cigarette use on the 14 covariates described earlier. The inverse propensity score for the treatment group was 1, and the inverse propensity score for the control group was computed as $(Propensityscore)/[1-(Propensityscore)]$. Treatment and control groups showed large differences in inverse propensity weights, ranging from 0.017 to 79.68 for lifetime e-cigarette use (Supplementary Figure 1) and from 0.0006 to 3.56 for current e-cigarette use (Supplementary Figure 2), which shows strong confounding in these data and highlights the need for propensity score methods.

Weighted logistic regressions were conducted to determine the causal effect of e-cigarette use on conventional smoking. In addition to the inverse propensity weights, the sampling weights were also accounted for in the analyses, using R’s “survey” package²⁴; the two sets of weights were multiplied together in order to account for both the IPW and the survey weighting.

Finally, the relative contributions of e-cigarettes versus the set of 14 covariates to the final models was examined by comparing their respective proportions of variance explained, using the Nagelkerke’s pseudo R-square as calculated by the “BaylorEdPsych” package²⁷ in R.

Results

Among the 12 421 respondents, 25.33% (n = 3147) reported that they had ever used e-cigarettes. Of these lifetime users, 9.02% (n = 284) were current e-cigarette users. Those who ever smoked conventional cigarettes comprised 14.45% (n = 1795) of participants. Of these lifetime smokers, 6.57% (n = 118), still reported currently smoking conventional cigarettes.

Comparisons of descriptive statistics based on lifetime e-cigarette use are shown in Table 1. Of those who had ever tried e-cigarettes, 44.5% of them also had smoked conventional cigarettes before (n = 1401) whereas only 4.25% of those who are e-cigarette naïve had smoked conventional cigarettes at least once in their lifetime (n = 394); this difference is statistically significant (p < .001). Similarly, although 3.53% of lifetime e-cigarette users were current smokers (n = 111), only 0.08% (n = 7) of e-cigarette naïve participants were current smokers (p < .001).

Table 1.

Descriptive Statistics of Sample by Lifetime E-cigarette Use

	Ever used e-cigarettes (n = 3147)	Never used e-cigarettes (n = 9274)	Total	p
Ever smoked conventional cigarettes	1401 (44.52%)	394 (4.25%)	1795 (14.45%)	<.001
Currently smoke conventional cigarettes	111 (3.53%)	7 (0.08%)	118 (0.95%)	<.001
Grade
8th	1015 (32.25%)	4629 (49.91%)	5644 (45.44%)	(Reference)
10th	2132 (67.75%)	4645 (50.09%)	6777 (54.56%)	<.001
Race
White	1881 (59.77%)	5335 (57.53%)	7216 (58.10%)	(Reference)
Black	196 (6.23%)	860 (9.27%)	1056 (8.50%)	<.001
Hispanic	535 (17.00%)	1444 (15.57%)	1979 (15.93%)	.387
Others	535 (17.00%)	1635 (17.63%)	2170 (17.47%)	.187
Sex
Male	1605 (51%)	4325 (46.63%)	5928 (47.73%)	(Reference)
Female	1542 (49%)	4951 (53.37%)	6493 (52.27%)	<.001
Currently drink alcoholic beverages	1225 (38.93%)	688 (7.42%)	1913 (15.4%)	<.001
Currently smoke marijuana	925 (29.39%)	316 (3.41%)	1241 (9.99%)	<.001
Lifetime illicit substance use	1193 (37.91%)	932 (10.05%)	2125 (17.11%)	<.001
Perceived proportion of friends who smokea	1 (0–2)	0 (0–1)	0 (0–1)	<.001
Exposure to health warnings on cigarette packaging	1530 (48.62%)	2912 (31.4%)	4442 (35.76%)	<.001
Discipline problems
Any misbehaviors	1112 (35.34%)	1358 (14.64%)	2470 (19.89%)	<.001
Frequency of getting disciplined from misbehaviorsb	1 (1–1)	1 (1–1)	1 (1–1)
Risk-taking behaviorc	4 (3–5)	3 (2–4)	3 (2–4)	<.001
Positive moodd	2 (2–2)	2 (2–2)	2 (2–2)	<.001
Level of dissatisfaction being near people smokingc	5 (3–5)	5 (3–5)	4 (3–5)	<.001
Level of disapproval of adults smoking more than one pack per daye	2 (2–3)	3 (2–3)	3 (2–3)	<.001
Highest level of education father finishedf	4 (3–5)	5 (3–5)	4 (3–5)	<0.001

Categorical variables are summarized as n (valid percentage), and quantitative variables are summarized as median (interquartile range). p values are based on unweighted, unadjusted logistic regression. E-cigarette = electronic cigarette.

^aThe scale represents 0 (none), 1 (a few), 2 (some), 3 (most), and 4 (all).

^bThe scale represents 0 (never), 1 (seldom), 2 (sometimes), 3 (often), and 4 (always).

^cThe scale represents 1 (disagree), 2 (mostly disagree), 3 (neither), 4 (mostly agree), and 5 (agree).

^dThe scale represents 1 (not happy), 2 (pretty happy), and 3 (very happy).

^eThe scale represents 1 (don’t disapprove), 2 (disapprove), and 3 (strongly disapprove).

^fThe scale represents 1 (completed grade school), 2 (some high school), 3 (high school), 4 (some college), 5 (college), and 6 (graduate or professional school after college).

Comparisons of descriptive statistics broken down by current e-cigarette use are presented in Table 2. Of those 284 current e-cigarette users, 74.3% (n = 211) reported ever smoking conventional cigarettes, whereas only 13% (n = 1584) of those not currently using e-cigarettes had ever smoked conventional cigarettes (p < .001). Likewise, although 10.2% (n = 29) of current e-cigarette users reported being current cigarette smokers, only 0.7% (n = 89) of those not currently using e-cigarettes reported currently smoking conventional cigarettes (p < .001).

Table 2.

Descriptive Statistics of Sample by Current E-cigarette Use

	Current user of e-cigarette (n = 284)	Not a current user of e-cigarette (n = 12 137)	Total	p
Ever smoked conventional cigarettes	211 (74.3%)	1584 (13.05%)	1795 (14.45%)	<.001
Currently smoke conventional cigarettes	29 (10.21%)	89 (0.73%)	118 (0.95%)	<.001
Grade
8th	94 (33.1%)	5550 (45.73%)	5644 (45.44%)	(Reference)
10th	190 (66.9%)	6587 (54.27%)	6777 (54.56%)	<.001
Race
White	196 (69.01%)	7020 (57.84%)	7216 (58.10%)	(Reference)
Black	13 (4.58%)	1043 (8.59%)	1056 (8.50%)	.0052
Hispanic	30 (10.56%)	1949(16.06%)	1979 (15.93%)	.0026
Others	45 (15.85%)	2125 (17.51%)	2170 (17.47%)	.0981
Sex
Male	185 (65.14%)	5743 (47.32%)	5928 (47.73%)	(Reference)
Female	99 (34.86%)	6394 (52.68%)	6493 (52.27%)	<.001
Currently drink alcoholic beverages	179 (63.03%)	1734 (14.29%)	1913 (15.4%)	<.001
Currently smoke marijuana	153 (53.87%)	1088 (8.96%)	1241 (9.99%)	<.001
Lifetime illicit substance use	172 (60.56%)	1,953 (16.09%)	2125 (17.11%)	<.001
Perceived proportion of friends who smokea	2 (1–2)	0 (0–1)	0 (0–1)	<.001
Exposure to health warnings on cigarette packaging	175 (61.62%)	4267 (35.16%)	4442 (35.76%)	<.001
Discipline problems
Any misbehaviors	140 (48.61%)	2330 (19.20%)	2470 (19.89%)	<.001
Frequency of getting disciplined from misbehaviorsb	1 (1–1)	1 (11)	1 (1–1)
Risk-taking behaviorc	5 (3–5)	3 (2–4)	3 (2–4)	<.001
Positive moodd	2 (2–2)	2 (2–2)	2 (2–2)	<.001
Level of dissatisfaction being near people smokingc	3 (1–3)	4 (3–5)	4 (3–5)	<.001
Level of disapproval of adults smoking more than one pack per daye	2 (1–3)	3 (2–3)	3 (2–3)	<.001
Highest level of education father finishedf	4 (3–5)	4 (3–5)	4 (3–5)	<.001

Categorical variables are summarized as n (valid percentage), and quantitative variables are summarized as median (interquartile range). p values are based on unweighted, unadjusted logistic regression. E-cigarette = electronic cigarette.

^aThe scale represents 0 (none), 1 (a few), 2 (some), 3 (most), and 4 (all).

^bThe scale represents 0 (never), 1 (seldom), 2 (sometimes), 3 (often), and 4 (always).

^cThe scale represents 1 (disagree), 2 (mostly disagree), 3 (neither), 4 (mostly agree), and 5 (agree).

^dThe scale represents 1 (not happy), 2 (pretty happy), and 3 (very happy).

^eThe scale represents 1 (don’t disapprove), 2 (disapprove), and 3 (strongly disapprove).

^fThe scale represents 1 (completed grade school), 2 (some high school), 3 (high school), 4 (some college), 5 (college), and 6 (graduate or professional school after college).

In terms of demographic characteristics, a higher proportion of 10th graders were lifetime and current e-cigarettes users compared to 8th graders (both p < .001). The odds for e-cigarette use were significantly lower for blacks compared to whites (p < .001, p = .0052, respectively), and for females compared to male (both p < .001). Hispanics were more likely than whites to currently use e-cigarettes (p = .0052) but did not significantly differ with respect to lifetime use (p > .05). All of the covariates accounted for in the final analyses showed significant differences across “treatment” (e-cigarette use) and control (no e-cigarette use) groups (all ps < .05).

Unadjusted, survey-weighted regressions of (A) lifetime and (B) current cigarette smoking on (1) lifetime and (2) current e-cigarette use are presented in Table 3 (first column). Prior to adjusting for shared risk factors, those who had ever used an e-cigarette had approximately 17 times the odds of ever having smoked a conventional cigarette, 36 times the odds of currently smoking conventional cigarettes. Similarly, current users of e-cigarettes had 22 times the odds of having ever smoked conventional cigarettes in their life, and 16 times the odds of currently smoking conventional cigarettes, relative to those who did not currently use e-cigarettes.

Table 3.

Comparison of Results From Survey-Weighted Unadjusted Regression, Survey-Weighted Adjusted Regression, and Survey- and Inverse Propensity-Weighted Regression

	Unadjusted regression OR (95% CI), p	Adjusted regression OR (95% CI), p	IPW regression OR (95% CI), p
Ever e-cigarette use on ever smoking	17.23 (14.86 to 19.97), p < .001	5.63 (4.71 to 6.73), p < .001	2.49 (1.77 to 3.51), p < .001
Ever e-cigarette use on current smoking	35.86 (15.85 to 81.11), p< .001	4.45 (1.73 to 11.40), p =.002	2.17 (0.62 to 7.63), p = .228
Current e-cigarette use on ever smoking	22.08 (16.19 to 30.13), p < .001	3.79 (2.53 to 5.68), p < .001	2.32 (1.66 to 3.25), p < .001
Current e-cigarette use on current smoking	16.46 (10.22 to 26.50), p < .001	1.37 (0.66 to 2.84), p = .403	0.95 (0.55 to 1.60), p = .849

Results are presented as odds ratio (OR), p value, and 95% confidence interval (CI). Inverse propensity weighting (IPW) accounted for covariates (grade, sex, race, current alcohol use, current marijuana use, lifetime use of other illicit substances, perceived peer smoking, perceived exposure to health warnings on cigarette packaging, frequency of getting disciplined for misbehavior, willingness to break rules for new and exciting experiences, positive mood, disapproval of smoking, dissatisfaction being near people smoking, highest level of education of father), bolded if p < .05. E-cigarette = electronic cigarette.

Next, traditional risk adjustment was performed by including all 14 confounding variables as covariates in a survey-weighted, logistic regression model (Table 3, second column). As expected, accounting for confounders significantly lowered the strength of the association between e-cigarette use and conventional cigarette smoking. Specifically, those who ever used e-cigarettes had approximately 5.6 times the odds of having ever smoked a conventional cigarette, and 4.5 times the odds of currently smoking conventional cigarettes, compared to those who never used e-cigarettes. In addition, those who currently used e-cigarettes had a 3.8-fold increase in the odds of having ever smoked a conventional cigarette. Notably, however, current e-cigarette use was no longer associated with current cigarette smoking.

Finally, IPW- and survey-weighted regressions are presented in Table 3 (third column). After IPW, neither lifetime nor current e-cigarette use had a remaining significant association with current conventional cigarette smoking. E-cigarette use remained associated with having ever smoked a conventional cigarette, though these relationships were weaker using IPW than using traditional regression control (as indicated by the odds ratio under IPW being outside of the confidence interval under a traditional regression). Specifically, lifetime e-cigarette use was associated with approximately a 2.5-fold odds of having ever smoked a conventional cigarette, and 2.3 times the odds of currently smoking conventional cigarettes.

Corroborating these main findings that e-cigarette use has a weak or absent association with conventional smoking after risk adjustment with IPW, e-cigarettes explained little additional variance in conventional smoking status over and above the variance explained by shared risk factors. In particular, Nagelkerke’s pseudo R-squared showed that the 14 covariates alone were able to explain almost half of the variance for cigarette smoking in the final model (45% for lifetime smoking and 42% for current smoking). Adding e-cigarette use to the model explained only 1%–7% additional variance.

Discussion

This study is among the first studies attempting to estimate the relationship between e-cigarette use and conventional cigarette smoking by using propensity score methods for causal inference, which produce less-biased effect estimates compared to conventional regression adjustment. Raw, unadjusted analyses showed a very strong and significant relationship between e-cigarette use and conventional cigarette use. Conventional regression control yielded significantly weaker relationships, and IPW reduced them further. In particular, lifetime and current e-cigarette use was associated with slightly more than doubled odds of ever smoking a conventional cigarette; but e-cigarette use was not significantly associated with currently smoking conventional cigarettes, according to IPW adjustment based on 14 shared risk factors.

Corroborating these main findings, e-cigarette use only marginally increased explanatory power of conventional smoking status, above that of the set of shared risk factors. These findings suggest that shared risk factors for tobacco use can fully explain this relationship. In the case of current conventional cigarette smoking, we find no evidence of a residual association with either lifetime or current e-cigarette use.

The current findings are consistent with current knowledge about risk factors for tobacco use. Recent studies^18,28 showed that the determinants of e-cigarette use showed little to no difference compared to those of conventional cigarette smoking. Specifically, major risk factors for both conventional smoking and e-cigarette use include parental education^21,29–32 and smoking^30–33; peer smoking^{19,28,29,32–35}; sensation-seeking behavior^{18,19,30,31,33,34}; impulsivity³²; delinquent behavior^18,19,32,36; internalizing symptoms (depression, anxiety, etc.)^18,33; alcohol,^{18,28–30,37} marijuana,^18,29 or other illicit substance use^18,32,33; exposure to health warning labels on cigarette packs^30,31; and cigarette advertising receptivity,^30,33 The fact that electronic and conventional cigarettes share all of these major risk factors suggests that adolescents have a near-identical preexisting propensity for using e-cigarettes as they do for smoking conventional cigarettes. This strong overlap of risk factors underscores the need for propensity score methods to reduce bias in effect estimation. Moreover, taken together with the current findings, this suggests that conventional cigarette smoking may be entirely attributable to adolescents’ preexisting propensity to smoke, rather than their use of e-cigarettes.

This study contradicts some previous work showing that e-cigarette use is associated with a higher likelihood of smoking conventional cigarettes,^13,30,31 as this study found no evidence for a role of e-cigarette use in promoting regular and/or current cigarette smoking in youth. One reason for the discrepancy could be methodological: conventional regression control is known to be biased^16,17 in cases of strong confounding, and some previous literature may reflect unmeasured or unaccounted for confounding. Doran et al.³⁸ made a novel attempt to apply propensity score methods into this research question; however, the unavailability of known confounders in the dataset (eg, sensation-seeking behavior, other substance use) again raises the possibility of biased effect estimates.

The current results do show a possible causal role of e-cigarette use on lifetime use or initiation of conventional cigarettes, consistent with other previous studies.^21,32,34,35 Several mechanisms have been hypothesized to explain this transition, such that 99% of e-cigarette products contain nicotine³⁹ and, therefore, e-cigarette use promotes the development of nicotine dependence^21,29,30,34; however, evidence of this mechanism actually occurring is mixed.¹⁰ Other possible mechanisms include lowering the perceived risk of smoking,⁴⁰ becoming curious about cigarettes,³¹ or mimicking the behavioral, sensory act of smoking.^34,41

On the other hand, given that IPW drastically reduced the strength of all associations, even compared to traditional regression control, the residual association between e-cigarette use and conventional smoking may reflect imperfect risk adjustment in this study. Namely, some variables were categorized to reduce extreme dissimilarity between treatment and control groups. Further, a crucial assumption of propensity score methods is that there are no unmeasured confounders; this assumption is likely violated, despite the many confounding variables available in the MTF study. Thus, if more refined risk adjustment were possible, this may explain the remaining, now-weaker relationship between e-cigarette use and lifetime conventional smoking. Further research is needed using a more complete set of risk variables, a larger sample size, and longitudinal data to evaluate whether there is any remaining relationship between e-cigarette and initiation of conventional cigarettes.

Taken together, the current findings fail to support the gateway theory, which predicts that e-cigarettes will promote the initiation as well as continued^21,31,42 and heavier⁴³ smoking of conventional cigarettes. However, this study also does not support e-cigarettes as a significant agent in cessation^44,45 or primary prevention⁴⁶ of e-cigarettes, as our results did not show e-cigarettes to have an inverse effect on conventional smoking. Our results instead suggest that tobacco users are a stable population determined by a host of preexisting risk factors. Thus, e-cigarette use primarily represents a diversification of tobacco products but does not seem to affect the size of the user population—either an increase (through hypothesized “gateway” effects) or a decrease (through desired cessation effects). This is consistent with national-level data in the United States showing that, despite the increase in e-cigarettes and decrease in conventional cigarettes, the overall prevalence of the use of any tobacco product has remained stable across recent years.^47,48

This study has some limitations that should be taken into consideration. First of all, due to the cross-sectional nature of data used in this study, temporality cannot be examined in this causal inference analysis, which is one of the major assumptions for establishing causality. However, the large set of available confounders in MTF overrode the drawbacks of MTF’s cross-sectional nature, given that the purpose of this study was to demonstrate the reduced bias of IPW methods relative to conventional regression control. Related to the cross-sectional sample, the order of substance initiation is not accounted for; thus, the temporality assumption that these adolescents initiated with e-cigarettes prior to initiating with conventional cigarettes does not necessarily hold for the whole sample. Thus, the current findings can neither support nor dispute the causal effect of e-cigarette use on subsequent conventional cigarette smoking. However, this is a plausible assumption for a majority of the sample given figures pointing to e-cigarettes as the most common introductory tobacco product.⁴⁹ Moreover, the question of temporality becomes less relevant if the association is due to common risk factors.³² Second, this study used self-report data, which are vulnerable to potential response bias and may affect the results in unknown ways. Finally and most importantly, the possibility of unmeasured confounding variables cannot be excluded. For example, important risk factors such as parental smoking and advertising receptivity were not included in the final data analyses due to not being available in the MTF survey. However, the MTF study contains more confounding variables compared to many other studies, leading to comparatively less bias. Moreover, additional risk factors would likely serve to reduce the causal estimate of e-cigarettes’ effect on conventional smoking even further. Nevertheless, this study offers an important first step in evaluating the causal relationships of e-cigarette use on conventional smoking, through a demonstration of the reduced bias offered by propensity score methods. The current findings related to causality of e-cigarette use on conventional smoking should be further confirmed by future studies using longitudinal data.

In spite of these limitations, this study provides a novel contribution to the debate on the effects of e-cigarettes through its innovative methodological approach. Propensity score methods can produce less-biased estimates of the effect of e-cigarette use on conventional cigarette smoking through their rigorous approach to risk adjustment, even in observational data.¹⁶ In addition, by using a nationally representative, recent, and large MTF dataset along with survey-weighted analyses, the results of this study are designed to be generalizable to the larger population of US adolescents.

Conclusions

After adjusting for tobacco use risk factors, e-cigarette use remained weakly associated with conventional smoking initiation. It remains unclear currently whether this reflects a true causal effect, or whether it reflects residual confounding due to imperfect measurement of risk factors. However, e-cigarette use was not found to be associated with current, continued conventional cigarette smoking after adjusting for these risk factors. These findings support neither the concerns that e-cigarettes act as a “gateway” to conventional cigarette smoking, nor hopes that e-cigarettes are significantly diverting from conventional cigarettes.

Funding

This work was supported by an Early Career Grant Award to AS from the University of North Dakota, the National Institute on Drug Abuse (grant number L40 DA042431), and the National Institute for General Medical Sciences (grant number 1P20GM121341-01).

Declaration of Interests

None declared.