Sample
Respondents were from the 7th Wave of Understanding Society, a panel survey based on annual interviews conducted within UK households [17]. In fieldwork spanning 2015-2017, 4534 youth aged 10-15 were eligible for inclusion because they lived in a household with a member of the (adult) study sample. Youth were not interviewed directly but confidentially filled in a self-completion questionnaire after their parent or carer had given permission for them to take part. A total of 3635 youth (80.2% of those eligible) in 2759 households completed questionnaires. Valid survey weights designed to account for household attrition, non-response and over-sampling were available for 3291 youth (90.5% of those responding) and used throughout to render the sample representative of the UK [28]. Multiple imputation of missing values (25 datasets, using an unconstrained model in which all analysis variables predicted all others) enabled inclusion of all observed data from respondents with valid weights [29] (proportions missing for most variables were between 0 and 5.3%, though 18.6% for ethnicity; see Table 1).
Measures
Youth and parents self-reported vaping in response to the question: “Do you ever use electronic cigarettes (e-cigarettes)?” (Yes/No). This was the first survey wave in which respondents had been asked about vaping. However, smoking was self-reported by youth in this and up to five earlier waves of the survey, depending on when they had reached age 10. Youth were first asked “Do you ever smoke cigarettes at all?” (yes/no), and if ‘yes’ were asked to tick a statement that best described them (only smoked once or twice; used to smoke but don’t now; sometimes smoke but not every week; usually smoke between one and six cigarettes a week; usually smoke more than 6 cigarettes a week). Youth responding ‘no’ in wave 7 and all previous waves for which data were available were coded as ‘never smoked’, while a ‘yes’ response in any wave was coded as having ‘ever smoked’. Current smoking was coded as smoking sometimes or more frequently at wave 7. Initiation of smoking was coded as current smoking with no indication of smoking in earlier survey waves. Parents were asked about current smoking (yes/no) in wave 7 and questions on smoking history from previous survey waves were used to distinguish never from ex-smokers. Parental smoking (never, ex, current) and vaping (yes/no) were coded according to the highest level of use from either parent, i.e. parental current smoking/vaping indicates that at least one parent smoked or vaped, while parental ex-smoking indicates no parents smoked currently but at least one was an ex-smoker.
Socioeconomic position (SEP) was measured with three variables at the household level (taking the more advantaged responses from couple parents): highest educational level (degree or higher, A-Level or equivalent, GCSE or equivalent, or no qualifications); occupational status using NS-SEC codes (managerial or professional, intermediate, routine, or not employed); and household income, equivalised for household composition and split into quartiles. For ease of presentation, SEP measures were treated as ordinal when assessing confounder balance, with higher values indicating greater socioeconomic disadvantage. Indicators of gender (male vs female), ethnicity (White UK vs ethnic minority), family structure (couple vs single parents), UK country (England, Wales, Scotland, and Northern Ireland) and interview date (to account for temporal trends in smoking and vaping during fieldwork) were also included, with youth age (in years) as a continuous variable.
Statistical analyses
We estimated ATEs and ATTs using a propensity weighting procedure, which is designed to balance measured confounders across the main exposure groups, i.e. youth who did and did not vape, and youth with parents who did and did not vape [27, 30]. This involves first running logistic regression models to predict each exposure, based on measured confounders (identified a priori). Gender, age, ethnicity, family structure, household SEP, UK country and interview date were treated as potential confounders throughout, as was parental smoking (except when stratifying on this variable). For estimating effects of youth vaping, parental vaping was included as an additional confounder.
The predicted probability of each individual’s observed exposure status was used to calculate weights for estimating ATEs and ATTs. Table 2 details these calculations. ATE weights re-weight exposed and unexposed respondents to resemble the total sample (with regards to measured confounders), while ATT weights re-weight the unexposed respondents to resemble the exposed group. Prior to using these weights to estimate effects, validity of the weights was assessed by examining mean differences in confounders associated with the relevant exposure [30]. Weights were deemed valid if confounder differences, expressed in standard deviation units, were reduced close to zero (with differences <0.2 considered close to 0). Models predicting exposure probability initially used main effects of confounders only, but where imbalance remained, the model was revised by adding interactions terms and then re-assessed. Improvements in confounder balance from model revisions were balanced against sufficient overlap of propensity distributions between exposed and unexposed groups by confirming that mean ATE weights were close to 1 [30, 31] (the same is not expected of mean ATT weights). Deviations from this would suggest that some individuals were being assigned extreme weights, indicating risk of making inferences not strongly supported by the available data.
ATEs and ATTs were then estimated in weighted logistic regressions of each outcome on the exposure of interest. For comparison, we also present associations weighted for sample selection only (labelled “sample weighted associations”). Standard errors were adjusted for clustering of youth within households. Z-tests were used to compare differences in effect estimates between strata of parental smoking and between ATEs and ATTs [32].
Analyses of smoking initiation excluded 216 youth who had reported ever smoking in previous survey waves. These prior smokers were older, more likely to be vaping and to have single parents. Since this could introduce selection bias, these differences were reduced by additional weighting back to the total sample for analyses of initiation.
Since these estimates may still be biased by unmeasured confounding, we calculate e-values for each point estimate and for the lower limit of the confidence interval [33]. E-values represent the minimum strength of association (OR in our analysis) that a set of unmeasured confounders would need to have with both the outcome and exposure of interest (independent of measured confounders), in order to respectively explain away the association, or cause its lower confidence interval to include the null (if it already includes the null the e-value for the lower limit will be 1). We include e-values for the sample-weighted associations, to indicate how much these were reduced by weighting for measured confounders.