Study selection, inclusion, and exclusion.
Odds ratios (ORs) and 95% confidence intervals (CIs) for studies conducted in the 1980s. The total represents the average OR derived from a mixed-effects model analysis. Lowercase letters after a reference number indicate different effects from the same study.
Odds ratios (ORs) and 95% confidence intervals (CIs) for studies conducted in the 1990s. The total represents the average OR derived from a mixed-effects model analysis. Lowercase letters after a reference number indicate different effects from the same study.
Odds ratios (ORs) and 95% confidence intervals (CIs) for studies conducted in the 2000s. The total represents the average OR derived from a mixed-effects model analysis. Lowercase letters after a reference number indicate different effects from the same study.
Wellman RJ, Sugarman DB, DiFranza JR, Winickoff JP. The Extent to Which Tobacco Marketing and Tobacco Use in Films Contribute to Children's Use of TobaccoA Meta-analysis. Arch Pediatr Adolesc Med. 2006;160(12):1285-1296. doi:10.1001/archpedi.160.12.1285
Copyright 2006 American Medical Association. All Rights Reserved. Applicable FARS/DFARS Restrictions Apply to Government Use.2006
To quantify the effect of exposure on initiation of tobacco use among adolescents.
A systematic literature search of MEDLINE, PsychINFO, ABI/INFORM, and Business Source Premier through October/November 2005 was conducted. Unpublished studies were solicited from researchers.
Of 401 citations initially identified, 51 (n = 141 949 participants) met the inclusion criteria: reporting on exposure and tobacco use outcomes and participants younger than 18 years. Included studies reported 146 effects; 89 were conceptually independent effects. Data were extracted independently by 3 of us using a standardized tool. Weighted averages were calculated using a linear mixed-effects model. Heterogeneity and publication bias were assessed.
Exposures (tobacco advertising, promotions, and samples and pro-tobacco depictions in films, television, and videos) were categorized as low or high engagement based on the degree of psychological involvement required.
Main Outcome Measures
Outcomes were categorized as cognitive (attitudes or intentions) or behavioral (initiation, tobacco use status, or progression of use).
Exposure to pro-tobacco marketing and media increases the odds of youth holding positive attitudes toward tobacco use (odds ratio, 1.51; 95% confidence interval, 1.08-2.13) and more than doubles the odds of initiating tobacco use (odds ratio, 2.23; 95% confidence interval, 1.79-2.77). Highly engaging marketing and media are more effective at promoting use (odds ratio, 2.67; 95% confidence interval, 2.19-3.25). These effects are observed across time, in different countries, with different study designs and measures of exposure and outcome.
Pro-tobacco marketing and media stimulate tobacco use among youth. A ban on all tobacco promotions is warranted to protect children.
Approximately 1.4 million children younger than 18 years in the United States begin smoking cigarettes each year.1 Addiction to tobacco begins quickly, and the earlier a person initiates use the more likely he or she is to develop a severe and persistent addiction.2- 4 Therefore, factors that influence children to start using tobacco represent a significant public health risk.
DiFranza et al5 applied the criteria of Hill6 for assessing causality in epidemiologic studies to examine the marketing of tobacco products and children's initiation of tobacco use. Youth exposed to marketing develop positive attitudes, beliefs, and expectations about tobacco use, which foster the intention to use, and intention leads to initiation. They concluded that the evidence strongly supported a causal inference. Other reviews7- 9 have concluded that children's risk of using tobacco is increased by industry-sponsored tobacco promotion and the depiction of tobacco use in films. This meta-analysis assesses the magnitude of the risk and the stability of the effect across study design characteristics. We use the terms pro-tobacco marketing and media to refer to tobacco advertising, promotions, and samples and to pro-tobacco depictions in films, television, and videos, respectively. A quantified measure of risk will help resolve the tobacco industry–driven controversy about whether their marketing efforts are harmful10- 14 and, if the risk is substantial, will provide evidence to policy makers supporting measures to restrict children's exposure.
Following the Meta-analysis of Observational Studies in Epidemiology standards for reporting meta-analyses,15 we conducted a systematic search of MEDLINE (January 1966 to October 2005), PsychINFO (January 1985 to October 2005), Business Source Premier (January 1965 to October 2005), and ABI/INFORM Global (January 1971 to November 2005) to identify studies that assessed a link between any of 5 tobacco use outcomes (attitudes toward use, intention to use, initiation of use, tobacco use status, and progression to heavier use) and tobacco marketing (advertising, promotions, and samples) and media (depictions of tobacco use in films, television, and videos). We searched using Medical Subject Headings and the text key words advertising or tobacco advertising coupled with adolescent or adolescence and smoking or smoking initiation (or related headings appropriate to the database). Separate searches were conducted using the headings/key words motion pictures or television coupled with adolescent or adolescence and smoking or smoking initiation, or appropriate terms. Searches were limited to studies of humans published in English in peer-reviewed journals. We then searched the reference lists of review articles and all potentially relevant articles to identify additional candidates, and we solicited unpublished studies via a posting on the Society for Research on Nicotine and Tobacco Listserv and by talking with colleagues.
All potentially relevant articles were independently assessed for inclusion by 2 evaluators (R.J.W. and J.R.D.); in cases of disagreement, a third evaluator (D.B.S.) assessed the study and a consensus was reached.16 To be included, a study had to measure both a tobacco use outcome and exposure to marketing or media. Measures of exposure included asking participants to recognize a brand name or logo, recall a brand, identify a favorite brand, express appreciation of advertisements, report whether they had received a sample of tobacco or had received or would use a tobacco promotional item, or report how many actors they had seen smoking in movies or which movies they had seen. Measures of tobacco use outcome included attitudes toward or expectations about the act of using (rather than attitudes about users), intention to use among youths who had never done so, initiation of use, status as a nonuser or user, and progression to heavier use. We excluded articles without original data and those where the primary focus was on (1) tobacco in the context of disease, (2) adult tobacco use, (3) law or policy, (4) anti-tobacco campaigns, (5) effects of advertising on adult tobacco consumption, (6) comparison of smoking rates across nations without controlling for any factor other than advertising, or (7) marketing practices.
Data were extracted independently by 3 of us (R.J.W., J.R.D., and D.B.S.) using a standardized tool. In the rare instances when coding did not match, mismatches were resolved by consensus.16 We noted the first author's affiliation (medicine/public health, behavioral sciences, business/marketing, communications, government, or private industry); year of publication; year of data collection (first year for prospective studies); country and sample breadth (national, regional, single state, or local); sampling strategy (representative or convenience); final sample size; age, sex, and ethnicity of the sample; research design (cross-sectional, prospective, or experimental); and the number of effect size estimates (ESs) (which are unadjusted odds ratios [ORs] with accompanying 95% confidence intervals [CIs]).
Before the resolution of mismatches, raters coded categorical study characteristics identically 98.0% of the time on average (range, 93.9%-100%). For continuous variables, the mean correlation was 0.98 (range, 0.95-1.00). Before the resolution of mismatches, the raters coded exposure measures identically 97.3% of the time and outcome measures 97.9% of the time. Because of a high proportion of missing data, several variables (year of data collection, age, sex, and ethnicity) could not be included in analyses of moderator variables.
For each ES we determined the operational definition of the exposure variable and classified it within a conceptual category: awareness (eg, recognition of a brand name or logo, ability to name a brand, or expressed awareness of tobacco promotions), advertising exposure (eg, having seen advertisements or the number of advertisements seen), movie exposure (eg, the number of smoking scenes viewed in movies), appreciation of advertisements (eg, naming a favorite advertisement or rating advertisements as appealing), receipt of a tobacco promotional item or a sample of tobacco, participation in a promotional campaign (eg, willingness to use or actual use of a promotional item), or a scale of receptivity to marketing and media that combined aspects of the other measures. Likewise, we determined the operational definition of the outcome variable and classified it as follows: attitudes toward use, intention to use (typically measured either by a yes answer to a question about intention or by a score on the standardized susceptibility measure of Pierce et al17,18), initiation of use, status as a user or nonuser, or progression of use (ie, heavier use or earlier age at initiation).
In cases in which multiple studies used the same data to investigate conceptually equivalent exposure measures and outcomes, we used the study that most clearly showed the relationship. For example, if one study reported a dichotomous effect19 and a second reported a dose-response relationship,20 we chose the dichotomous effect. In cases in which a single article reported data derived from separate samples (eg, 2 different age groups21 or data collected in different years22) or from separate designs (eg, cross-sectional and longitudinal23), the sets of data were treated as separate studies.
To avoid overweighting of individual studies, when multiple ESs in a single study addressed the same conceptual categories of exposure and outcome (eg, comparing smokers' and nonsmokers' recognition of 6 brand names and logos, all of which would address awareness24), we averaged the ESs using DSTAT version 1.11.25
The primary analyses were conducted using Comprehensive Meta-analysis version 2.2.023.26 Exposures and outcomes were treated as dichotomous variables. Weighted averages are reported as ORs and 95% CIs; study weights were assigned by means of inverse variance weighting. The overall ES and ESs for the analyses of moderator variables were calculated using a linear mixed-effects model to account for differences between studies.27,28 The consistency of effects across studies is reported as I2, the percentage of total variance attributable to heterogeneity among studies rather than to chance.29 Values of I2 less than 25% are considered low, those around 50% are moderate, and those greater than 75% indicate high heterogeneity.
We assessed the potential for publication bias in 6 ways. We calculated Begg and Mazumdar rank correlations30 and the regression intercept of Egger et al.31 Because the Egger et al regression test can yield false-positive results when ESs are reported as ORs, we also performed the alternative regression test of Peters et al.32 We calculated fail-safe Ns to determine how many additional null findings would be needed to render the results nonsignificant at α = .05, using the Rosenthal methods,33,34 or to yield a “trivial” OR of 1.05, using the Orwin method.35 Rosenthal suggested that meta-analysts calculate a tolerance level around a fail-safe N equal to 5 times the number of effects included in the meta-analysis (symbolized by k) plus 10 (the “5k + 10” benchmark)34,36,37; we used this benchmark to assess the observed fail-safe Ns. Finally, to assess whether smaller studies distorted the overall ES we conducted a cumulative meta-analysis, adding studies sequentially from largest to smallest sample size.38
To assess the relationship between design quality and ES, we analyzed differences in ES between 5 subgroups created to represent a continuum of increasing design quality: cross-sectional design/convenience sample, cross-sectional/representative, experimental/convenience, prospective/convenience, and prospective/representative. Ordinarily, an experimental design would be considered strongest for demonstrating causality, but because of ethical considerations, experimental studies in this field are limited to assessments of attitudes. Prospective studies, which can track the change from nonuse to use, were, therefore, ranked highest.
Long ago, the Federal Trade Commission raised concerns that tobacco industry marketing was targeted at young people.39 Similar concerns have been raised about depictions of tobacco use in movies.40 In 1998, the Master Settlement Agreement between the industry and 46 US states banned advertising on billboards and in youth-oriented magazines.41 To determine whether there was a decrease in ESs across time that might be attributed to changes in marketing and media, we conducted a second cumulative meta-analysis in which studies were sorted from oldest to most recent.
The initial search yielded 331 nonduplicated citations addressing marketing and 70 addressing media (Figure 1). After reviewing the abstracts we excluded 317 articles and retained 84 full-text articles for detailed evaluation; review of their reference lists yielded 9 additional published studies, and data from 2 unpublished studies were provided by colleagues. Via e-mail we requested additional data from the authors of 20 studies with insufficient published data. Up to 3 requests were made at monthly intervals; 8 authors met the request, providing data from 10 studies.12,17,19,80- 84 Full review of the 95 studies resulted in excluding 44 for the following reasons: unavailable data (n = 10),42- 51 no outcome measure (n = 8),52- 59 no exposure measure (n = 8),18,60- 66 duplicated another article (n = 6),20,67- 71 no original data (n = 7),8,10,11,13,72- 74 only time series data from which an ES could not be calculated (n = 2),75,76 included only tobacco users (n = 2),77,78 and studied only antismoking campaigns (n = 1).79 The 51 studies retained for the meta-analysis included 141 949 participants (Table 1) and presented 146 ESs, 89 of which were included in the analysis after averaging those estimates that were conceptually equivalent. Figures 2, 3, and 4 present the ESs and 95% CIs for all the studies, divided by decade of publication.
The overall ESs from the fixed- and mixed-effects models did not differ (fixed-effects: OR, 2.17; 95% CI, 2.12-2.22; mixed-effects: OR, 2.19; 95% CI, 1.93-2.47; P<.001 for both), and there was considerable heterogeneity among studies (I2 = 96.9). Evidence of publication bias was not observed with either the Begg and Mazumdar30 rank correlation (P = .89, 2-tailed) or the Egger et al31 regression intercept (P = .57, 2-tailed), although the Peters et al32 alternative regression method yielded a significant result (slope = −67.34; P<.001). The Rosenthal fail-safe N indicated that 86 514 additional null effects would be needed to render the overall ES nonsignificant at P = .05; the Orwin fail-safe N indicated that 1324 null effects would be needed to yield an OR of 1.05 or less. Both fail-safe Ns greatly exceed the 455 ESs representing the 5k +10 threshold. Together, these tests suggest little likelihood of publication bias and, given the strength of the effect, little likelihood that the exclusion of studies with negative results (due to a file drawer effect34 or unavailability of data) could have led to this outcome. The cumulative meta-analysis revealed no shift in ES as smaller studies were included. The ES after addition of the 2 largest studies,92,103 which accounted for 6 ESs, was 1.92 (95% CI, 1.28-2.88), and the ES after all the studies were included was 2.19. Because the overall ES is well within the 95% CI for that of the largest studies, there is no evidence of a “small study effect.”38 There was no correlation between sample size and ES (slope = 0.00; P = .29).
We next examined the effects of design quality and sampling breadth (Table 2). Overall, quality of design was unrelated to ES. In the initial analysis, studies with national samples tended to yield lower ESs than those with regional, state, or local samples (P = .06). However, 2 ESs from 1 study86 were significantly different from the other ESs in that study and from those derived from the other 5 studies with national samples.22,98,102,112 When the 2 outlying ESs were removed from the analysis there was no difference among studies on the basis of sampling breadth (P = .67).
A preliminary analysis revealed that 56 of the 89 ESs had tobacco use status (nonuser vs user) as the outcome variable, followed by initiation of use (n = 13), intention to use (n = 10), progression to heavier use (n = 5), and attitudes toward use (n = 5). Among exposure measures, advertising exposure (n = 17) and awareness of marketing (n = 17) were most frequently represented, followed by use of a promotional item (n = 14), approval of advertising (n = 14), movie exposure (n = 10), receptivity (n = 9), receipt of a promotional item (n = 6), and receipt of a tobacco sample (n = 2). The small number of ESs in each subclassification precluded a fine-grained analysis. Consequently, we classified exposure measures by the degree of psychological involvement required: low engagement (awareness of marketing, advertising exposure, movie exposure, and receipt of a promotional item or tobacco sample) or high engagement (appreciation of advertising, willingness to use a promotional item, and receptivity). We classified outcome measures as either cognitive (attitudes toward use and intention to use) or behavioral (initiation of use, status, and progression). Overall, high-engagement exposures produced higher ESs than did low-engagement exposures, and ESs for behavioral outcomes were greater than those for cognitive outcomes (Table 3). The ES for progression to heavier use (OR, 1.42) was significantly lower than those for initiation of use (OR, 2.23) and tobacco use status (OR, 2.46).
The classification scheme gave us 4 possible combinations of exposure and outcome (see Table 1 for the classification of each ES). The overall ES for studies using a cognitive outcome was 1.51 (95% CI, 1.08-2.13; P = .02); the ES for low-engagement exposure was smaller than for high-engagement exposure with a cognitive outcome (Table 4). Because of the small number of included studies, we assessed the potential for publication bias in studies with a cognitive outcome. The Rosenthal fail-safe N indicated that 340 null effects would be needed to render the ES nonsignificant at P<.05. The Orwin fail-safe N indicated that 179 null effects would be necessary to yield a trivial OR of 1.05. These fail-safe Ns greatly exceed the 5k +10 threshold of 85 null effects. Given that the ratio of excluded null effects to included effects is greater than 5:1, there is little likelihood that the exclusion of studies led to the observed outcome. The overall ES for studies using a behavioral outcome was 2.33 (95% CI, 2.03-2.67; P<.001); low- and high-engagement exposures with behavioral outcomes produced equivalent ESs (Table 4).
The cumulative meta-analysis by publication date did not reveal a trend toward decreasing ESs across time. There was no overall difference in ESs derived from studies conducted between 1981 and 1990 (8 studies and 12 ESs), 1991 and 2000 (26 studies and 37 ESs), or 2001 and 2005 (21 studies and 40 ESs) (P = .15). However, studies conducted since 2001 tended to yield larger ESs (overall ES, 2.45; 95% CI, 2.05-2.95) than did those conducted in the 1990s (overall ES, 1.89; 95% CI, 1.52-2.35; P = .07).
DiFranza et al5 concluded that exposure to marketing causes children to initiate tobacco use. We quantified this risk; the odds of becoming a tobacco user are more than doubled by exposure to marketing and media. This relationship is robust, with similar effects observed across time in different countries, in cross-sectional and prospective designs using a variety of measures of exposure, and whether the outcome is initiation or tobacco use status.
Exposure that does not actively engage the recipient has a substantial impact, increasing the odds of tobacco use by approximately 90%. This suggests that youth are highly susceptible to the pervasive presence of marketing and media. Once a recipient is psychologically involved, the odds increase almost 3-fold, suggesting that the interaction between exposure and an individual's psychological processes is powerful. The impact of marketing and media remains powerful even after the initiation of use; exposure increased the odds of progression to heavier use by 42%.
A variety of psychosocial factors (eg, age, ethnicity, family structure, parental socioeconomic status, personal income, parental attitudes and tobacco use, sibling and peer tobacco use, peer attitudes and norms, family environment, engagement in other risky behaviors, stress, depression, and self-esteem) raise the odds of initiation.122- 124 The appearance of symptoms of addiction seems to be the primary factor driving progression to heavier use.125 Thus, it is not surprising that marketing and media have a greater impact on initiation of use than on progression to heavier use.
Compared with unexposed youth, exposed youth had approximately 50% greater odds of holding positive attitudes toward tobacco use or expressing an intention to use in the future and more than twice the odds of initiating use. For years the tobacco industry has claimed that a cross-sectional association between exposure to marketing and tobacco use is found because users are more likely than nonusers to attend to marketing.13 If use causes exposure rather than vice versa, ESs in cross-sectional studies should exceed those in prospective studies of nonusers, in which assessment of exposure precedes the onset of use. The present findings refute the industry's argument; the ESs in prospective studies are equivalent to those in cross-sectional studies.
If attempts to mitigate the harm posed by marketing and media, embodied in voluntary and mandatory restrictions instituted during the past 2 decades, were effective, we would expect to observe a decrease in ESs across time. We did not. Indeed, the ES shows a trend toward increasing since 2001, suggesting that the risk to youth has increased. These measures have not had an evident effect on exposure. The tobacco industry has been remarkably adept at shifting strategies to avoid the restrictions imposed by the Master Settlement Agreement by preferentially placing their advertisements in magazines whose youth readership exceeds 20%,126,127 and tobacco use in current movies is comparable with that seen in the 1950s, when it was rampant.128
The strengths of this meta-analysis include the large number of ESs available for analysis, lack of evidence for publication bias, large overall sample size, the time during which data were collected, and diversity among studies in design, geographic distribution, and exposure and outcome measures. Potential limitations include the inability to assess possible moderation by demographic or psychosocial factors because of a lack of data or to conduct a fine-grained analysis of the interaction between exposure and outcome measures because of a paucity of studies in particular categories.
Despite the industry-fueled controversy surrounding the relationship between exposure to pro-tobacco marketing and media and initiation of use,10- 14 the present findings offer considerable evidence for a causal link. First, the overall ES is strong and robust, showing consistency between cross-sectional and prospective studies. Second, the 3 experimental studies reviewed demonstrated that greater exposure led to more positive attitudes toward use. Third, the failure of design-related moderator variables to impact the ES suggests that such alternative explanations are limited. Most important, in 13 prospective studies, exposure more than doubled the odds of initiating tobacco use.
Ethical considerations preclude experimental studies that would expose youth to marketing and media to see whether they would initiate tobacco use. However, laboratory and field-based experiments have shown that media portrayals of violence in films (low engagement) and video games (high engagement) produce subsequent violent attitudes and behaviors in children and adolescents.129,130 The causal relationship between violent media and behavior makes it plausible that tobacco marketing and media also affect behavior.
Given the money spent by the tobacco industry on worldwide marketing and the global distribution of American movies, it is unlikely that many children escape exposure. Our findings suggest that half of all children who initiate tobacco use do so because of exposure to marketing and media. Half the youth who initiate smoking will likely continue for more than 15 years,3 making themselves vulnerable to serious harm. Smoking causes 36 diseases,131 which kill almost half of all smokers.132 Current and future schemes that expose children to pro-tobacco messages, products, or images are a public health menace. More definitive measures to curtail exposure, such as a ban on all tobacco promotional activities, as called for in the World Health Organization Framework Convention on Tobacco Control,133 are needed and warranted.
Correspondence: Robert J. Wellman, PhD, Department of Family Medicine and Community Health, University of Massachusetts Medical School, 55 Lake Ave N, Worcester, MA 01655 (Robert.Wellman@umassmed.edu).
Accepted for Publication: June 27, 2006.
Author Contributions: Dr Wellman had full access to all the data presented in this study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Wellman, Sugarman, and DiFranza. Acquisition of data: Wellman, Sugarman, DiFranza, and Winickoff. Analysis and interpretation of data: Wellman, Sugarman, DiFranza, and Winickoff. Drafting of the manuscript: Wellman, Sugarman, and DiFranza. Critical revision of the manuscript for important intellectual content: Wellman, Sugarman, DiFranza, and Winickoff. Statistical analysis: Wellman and Sugarman. Obtained funding: Wellman and DiFranza. Administrative, technical, and material support: Wellman, Sugarman, DiFranza, and Winickoff. Study supervision: Wellman and Winickoff.
Financial Disclosure: None reported.
Funding/Support: This study was supported by grant 53081 from the Robert Wood Johnson Foundation (Dr Wellman).
Role of the Sponsor: The funding agency played no role in designing or conducting the study; in collecting, managing, analyzing, or interpreting the data; or in preparing, reviewing, or approving the manuscript.
Disclaimer: The ideas expressed herein are those of the authors and not necessarily those of the funding agency.
Acknowledgment: We thank Bethany J. Hipple, MPH (Center for Child and Adolescent Health Policy, Massachusetts General Hospital), for assistance in searching the business literature and for reviewing the first draft of the manuscript and William Shadel, PhD (Rand Corp) for reviewing the first draft.