Why do apprentices smoke much more than high school students? Understanding educational disparities in smoking with a Oaxaca-blinder decomposition analysis

The ESCAPAD survey (Enquête sur la Santé et les Consommations lors de l’Appel de Préparation A la Défense) is a repeated cross-sectional, standardized, nationally representative survey that provides estimates of prevalence of drug use among French adolescents aged 17 [26].

The ESCAPAD survey is conducted by the French Monitoring Centre for Drugs and Drug Addiction and the Department of National Civil Service and Youth during the National Defence and Citizenship Day (JDC): a one-day session of civic and military information that is compulsory for later enrolment in public exams (i.e. driver’s license, university exams, etc.). All French nationals are summoned to participate shortly after having turned 17. About 4% of youths never attend the JDC.

The ESCAPAD survey takes place during 2 weeks in March in all centres across the French territory where the JDC is organized. All of the adolescents present are invited to participate in the survey. The pen and paper questionnaires are self-administered, anonymous and follow the recommendations of the European Monitoring Centre for Drugs and Drug Addiction [27]. Participants are guaranteed complete confidentiality and anonymity and can refuse to complete the form. The survey was approved by the National Council for Statistical Information (CNIS), as well as the ethics commission of the French National Data Protection Authority (CNIL).

Our analysis focuses on the 2008 and the 2017 waves of the ESCAPAD survey. Response rates were consistently above 90% (number of filled questionnaires out of number of adolescents who attended the JDC the day of the survey). Respondents aged 17 consistently represented more than 90% of the samples, the others were aged 18. The data are calibrated to guarantee the weight of 17 years-olds in each “département” (official geographic and administrative unit) and their sex ratio inside each “département”.

We compared two educational tracks: high school (starting after middle school, around age 15) and apprenticeship. In France, apprenticeship contracts are work contracts opened to people aged 16 to 29. They combine 2 weeks of in-company training alternated with 1 week of courses [28]. Apprentices are at the same time students and employees, often in manual employment, and thus have regular contacts with an older and professional environment and receive a salary. Most apprentices enter the workforce after their diplomas, while most high school students enter tertiary education.

Our outcome variable was daily smoking at the time of the survey. This was measured through the question “during the last 30 days, have you smoked cigarettes?” with the following choices: never; less than once a week; less than once a day; 1–5 per day; 6–10 per day; 11–20 per day; more than 20 per day. People who declared smoking one or more cigarettes per day in the past 30 days were considered daily smokers.

The independent variables included the following demographic and familial characteristics: gender (girl/boy), age (as a continuous variable), living with no relatives (yes/no), parents living together (yes/no), father and mother smoking status (occasional or current smoker yes/no) and parental occupational status. Parental occupational status was defined as the highest occupational category of the parents, as reported by adolescents, based on the National Institute for Statistics and Economic Studies’ official typology (more details can be found elsewhere [29]). This yielded a final 5-category scale: low (both parents are unemployed or inactive, or both situations are missing), disadvantaged (the highest parental occupational category is manual worker or clerk), intermediate (highest parental occupational category is intermediate: technician and associate professional), advantaged (one parent is a manager, or has a liberal or intellectual occupation), high (parents are both managers, or have a liberal or intellectual occupation). In addition, suicidal attempts that lead to hospitalization in the lifetime (yes/no) were included as a proxy measure of mental health. We approached school performance by the question “Have you ever repeated a class?” (No; yes). We accounted for the use of cannabis (yes/no) and use of alcohol (yes/no) in the past month and the amount of money received in the past month (as a continuous variable). In 2008, this amount was obtained through three subquestions distinguishing pocket money, salary from a job (baby-sitting etc.) or from an internship or apprenticeship, and money for a special occasion. In 2017, salary from a job (baby-sitting etc.) was additionally distinguished from salary received from an internship or apprenticeship. The amount received superior to zero was capped at the 99th percentile for each subquestion to avoid obvious irrelevant reports. In 2017, parental smoking was available only for a subsample (thereafter called 2017-A) and the amount of money received only for another subsample (thereafter called 2017-B). Age at smoking initiation (“What was your age the first time you smoked a cigarette”?) and age at transition to daily smoking (“If you are smoking daily, at what age did you start smoking every day?”) were also measured but not included in the Oaxaca-Blinder models due to the correlation with our outcome.

In 2008, the survey included 39,542 adolescents (33,253 high school students, 4564 apprentices, 1725 school dropouts and 0 with missing education). In 2017, the subsample 2017-A included 12,471 adolescents (11,093 high school students, 877 apprentices, 459 school dropouts and 42 with missing education) and the subsample 2017-B 13,314 adolescents (11,769 high school students, 949 apprentices, 546 school dropouts and 50 with missing education). Those who dropped out of school and those who did not fill their educational background were excluded from the analyses.

First, we compared the distribution of all independent variables between apprentices and high school students as well as the daily smoking rate for each characteristic separately for apprentices and high school students. To do so, we calculated the absolute standardized mean differences (SMD) since they are not influenced by sample size [30].

Then, we applied a nonlinear Oaxaca-Blinder decomposition, designed for nonlinear outcomes. This approach decomposes the observed difference in daily smoking between apprentices and high school students into an explained and an unexplained part [31]. High school students were used as the reference group. The explained part is the part of the daily smoking gap associated with differences in measured characteristics between two groups. It shows the expected change in daily smoking gap that would be observed if apprentices were given the covariates’ distribution of high school students. The contribution of each variable to the daily smoking gap was reported as log odds and percentage contribution (percentage of the total difference). The explained part expressed in % is the proportion of the total disparity that would be reduced following a hypothetical intervention equalizing the average value of each covariate in both groups. The explained part thus quantifies the change in daily smoking disparities that would be observed if the averaged values for each covariate of interest were set to be equal between apprenticeships and high school students.

The unexplained part depicts, in addition to the effect of unmeasured characteristics, the heterogeneity across educational groups in the effect of characteristics on the smoking outcome, as well as the effect of the apprenticeship environment on educational disparities in smoking.

All analyses were performed with Stata 15. To take into account the “identification problem” that arises when including categorical variables in the model [32], we used the -oaxaca- command in Stata with the options ‘logit’, ‘weight(0)’ and ‘normalize’ [33]. We compared with results obtained using the command -mvdcmp- with the option ‘normal’ [34] and did not find any difference.

The decomposition analysis was conducted for the 2008 survey, the 2017-A and the 2017-B subsamples separately. Because of concurrent use of cannabis and alcohol use with daily smoking in the last 30 days, we performed sensitivity analyses by removing these two variables. In addition, for comparison purposes, we performed two decomposition analyses using the 2008 survey, each including similar variables as the one in the subsamples 2017-A and 2017-B (See Tables 1S and 2S, Additional File 1). Finally, we tested the robustness of our models by decomposing educational disparities in daily smoking for each independent characteristic separately (See Table 3S, Additional File 1).

Table 1 displays descriptive statistics of apprentices and high-school students aged 17. Regardless of the sample, compared to high school students, apprentices have received much more money, are more likely to be boys, to have repeated a class, to have used cannabis and alcohol in the past month, as well as to have parents who smoke, have a low occupational status, and do not live together. In the subsample 2017-B, apprentices were more likely to live alone than high school students. Suicide attempts were more frequent among apprentices in 2008 and in the subsample 2017-B. Daily smoking rates were about two times higher among apprentices than among high school students: 49.9% compared to 24.5% in 2008, 44.8% compared to 22.1% in the subsample 2017-A and 46.8% compared to 21.9% in the subsample 2017-B. Age at smoking initiation and especially age at transition to daily smoking were earlier for apprentices of all samples.

Table 2 presents daily smoking rates at 17 for each characteristic separately for apprentices and high school students and thus showcases the differences in the covariates of interest between apprenticeships and high school students. The same smoking profile was observed for high school students and apprentices for most variables: higher daily smoking rates were reported among adolescents whose parents smoked or did not live together as well as among adolescents who lived alone, reported at least one suicidal attempt in their lifetime, and used cannabis or alcohol in the past month. Daily smoking rates also gradually increased with the amount of money received in both groups. Repeating a class was associated with higher daily smoking rates among high school students but not among apprentices. Both among high school students and apprentices, parental occupational status did not clearly affect daily smoking rates. In the subsample 2017-A, female apprentices smoked less than male apprentices (38.1% vs 47.3%).

Table 3 describes the results from the non-linear Oaxaca-Blinder decomposition analyses. The explained part varied from 28.6 to 51.2% depending on the sample and the model. This represents the share of the observed daily smoking gap between high school students and apprentices that could be attributed to differences in measured characteristics. The variables that accounted the most for the smoking gap between high school students and apprentices were, in decreasing order of importance, cannabis use in the past month (12.9 to 16.9%), sum of money received (8.7 to 12.8%), alcohol use in the past month (4.3 to 8.0%), mother’s smoking (4.4 to 5.2%) and father’s smoking (3.4 to 4.9%). In 2008, gender and suicide attempt also explained a small part of the smoking gap between high school students and apprentices. Parental occupational status, grade repetition and living alone consistently showed no contribution to educational disparities in smoking.

In the sensitivity analyses excluding alcohol and cannabis use (Table 4), the explained part decreased in each model. The share of the explained proportion increased for money received, gender, parents living together and parental smoking.

The additional analyses for the 2008 survey using the same variables in 2008 as in the 2017 subsamples yielded similar findings (Tables 1S and 2S, Additional File 1). The explained share and the percentage contribution of the variables were quite similar in 2008 and in 2017, except for differences in 2008 already identified in the main analyses. Results from the additional analysis with each characteristic tested separately yielded consistent results (Table 3S, Additional File 1).

Our analysis is based on a unique dataset representative of all French adolescents aged 17, while most surveys interviewing adolescents tend to be school-based, thus missing adolescents out of the general educational system. With an original approach using a decomposition method, we aim to fill the gap in the literature about educational disparities in smoking among adolescents. That being said, some limitations have to be acknowledged. First, in addition to accounting for the effect of the apprenticeship setting, the unexplained part could also include the effect of some unmeasured characteristics (such as perceived peer smoking, perceived accessibility of cigarettes or specific youth socialization). Second, since our data are cross-sectional our analyses cannot conclude on causal relationships between the measured covariates and the gap in daily smoking. Third, although some variables may be determinant of the educational track (e.g. parental smoking, parental occupational class), others may be determined by the educational track (e.g. money received, other drug use). Fourth, the discrepancies in predictors for the 2017 a and b samples do not allow for comparing the effects of parental smoking and amount of money received in 2017. Asking different questions to different samples of adolescents allows for multiplying the topics covered in the survey without extending the length of the questionnaire but it does come with disadvantages. Finally, French high schools are divided in general and vocational high schools. Among French high school students, smoking differences also exist depending on their type of high schools [12]. This information was not available in ESCAPAD, and more research is needed to investigate smoking disparities within high school students.