Recruitment of sample
—We sent letters to 30 randomly selected middle schools in New Hampshire and Vermont with at least 150 students (fig ). Half the schools agreed to participate. The socioeconomic profiles of participating and non-participating schools did not differ. About half (52%) of the schools were in rural communities of less than 10
000 residents. In September 1999 proctors administered the confidential survey during class time (parents were informed by mail beforehand). The average participation by school was 92.5%; 128 (2.1%) parents or students refused participation, and 380 (6.3%) students were absent. We excluded 571 surveys for missing (n=565) or inconsistent (n=15) responses. Excluded students were likely to be younger (for example, fifth grade), to report poorer school performance, and to have seen fewer films than those with usable surveys, but smoking behaviour did not differ between included and excluded students.
Exposure to smoking in films—Figure illustrates our procedure for determining exposure to smoking in films. We counted occurrences of smoking in each of 601 popular contemporary films. We estimated exposure to these films by asking respondents whether they had seen 50 films randomly selected from the larger pool. On the basis of the films that adolescents reported seeing, we calculated the number of occurrences of smoking seen by each survey respondent.
Assessment of exposure to tobacco use in movies
Primary outcome—We determined whether students had ever tried smoking by asking the question “How many cigarettes have you smoked in your life?” We categorised a response of none as “never smoked” and all other responses (just a few puffs, 1-19 cigarettes, 20-100 cigarettes, >100 cigarettes) as “tried smoking.”
—We measured the following categories of factors that might be associated with trying smoking: sociodemographic characteristics (for example, school, age, sex, parents' education), social influences (parent smoking, sibling smoking, friend smoking, receptivity to tobacco promotions18–19
), and other characteristics of the child and family (self reported school performance, propensity to sensation seeking,20–21
two measures of authoritative parenting,24
and students' perception of parental disapproval of smoking). We measured reliability by using Cronbach's α.25
Table lists the questions used in the survey to assess these variables, with their reliability.
Measures for characteristics of child and parenting
Validity of responses to film questions—To evaluate the validity of adolescents' recollection of films they had seen, we re-contacted 49 adolescents who had participated in a longitudinal study in which they reported each month the films they had seen in the past week. Adolescents had excellent recognition of the films they reported seeing during the previous year, identifying films correctly 88% of the time. In addition, the adolescents rarely reported seeing false film titles with false actors (3.0%) or false film titles with real actors (2.7%).
Statistical analysis—We used the χ2 test or analysis of variance to evaluate the association between trying smoking and each of the confounding variables. We used logistic regression to determine the crude odds ratios, adjusted odds ratios, and 95% confidence intervals. Firstly, we used a crude model in which exposure to smoking in films was entered as four categories that corresponded to fourths of exposure in the student population. Next, we added controls for sociodemographic characteristics only. Then we added social influence variables, and finally we added other characteristics of the child and family. Age and indexed variables (sensation seeking, rebelliousness, self esteem, and the authoritative parenting measures) were entered as continuous variables. We did not include the number of R rated (restricted) films seen as a covariate because of its high correlation with occurrences of tobacco use (r=0.89). All tests were considered significant at the 0.05 level.
—We conducted a sensitivity analysis to determine whether an unmeasured confounder could explain our results.26
We considered the effect of adding a missing confounder (independent of other covariates) on the relation between seeing tobacco use in films and smoking in adolescents. The results of this analysis indicate how strongly an unmeasured confounder would have to be associated with exposure and outcome in order to lead to false reporting of an association.