Tobacco use remains a significant global health concern, causing millions of deaths annually. While smoking rates are declining in high-income countries, they persist in low- and middle-income countries (LMICs). Adolescence is a critical period for smoking initiation, heavily influenced by social factors like peer behavior and attitudes. The rise of e-cigarettes adds another layer of complexity, as they are often used for experimentation by adolescents and may serve as a gateway to traditional cigarette smoking. Smoking prevention programs frequently target younger adolescents (12–13 years) and often employ social norms-based approaches or aim to leverage peer influence. Peer influence, the process where an individual's behavior is shaped by peers, needs to be considered alongside selection homophily, the tendency to form friendships with similar individuals. Research consistently shows that adolescent smokers tend to have more smoking friends, and vice-versa. This correlation, however, can vary across cultures, potentially being stronger in collectivistic cultures compared to individualistic ones. High-income countries (e.g., UK, US) are generally more individualistic, while LMICs (e.g., many in Latin America) tend to be more collectivistic. Schools provide an ideal setting for interventions aiming to modify health behaviors by influencing peer norms and interactions. Many interventions have proven successful in high-income settings, but evidence from LMICs remains limited. Interventions targeting groups and social networks may prove more effective than those focusing solely on individuals. The ASSIST intervention, for instance, uses influential students to promote anti-smoking norms within their peer groups. These interventions often operate by changing social norms—shared rules and standards guiding behavior within a social group. Injunctive norms reflect perceptions of socially approved behaviors, while descriptive norms concern perceptions of behaviors actually performed by others. Social network structures influence how these norms spread, and vice versa. Traditional self-report assessments of norms are often susceptible to social desirability bias. Behavioral economics methods, however, provide incentivized approaches to measure norms more objectively, mitigating this bias. The MECHANISMS study is the first to apply these methods to adolescent smoking and vaping behaviors. This study investigates the mechanisms behind social norm transmission within school friendship networks by examining selection homophily and peer influence effects on experimental measures of smoking and vaping norms and other smoking-related outcomes in schools in Northern Ireland (high-income) and Bogotá (middle-income). The study also uses a novel approach by including a wide range of smoking-related psychosocial antecedents in addition to behavioral data.

Existing research on peer influence and selection homophily in adolescent smoking has largely focused on behavioral outcomes (smoking behavior, intentions, or susceptibility). Fewer studies have incorporated psychological characteristics. Go et al. (2012) used mixed-effects logistic regression and found both processes explained the association between peer smoking and adolescent smoking initiation. Hoffman et al. (2007), using cross-lagged panel structural equation models (CLPMs), showed that peer influence was a stronger predictor of 'ever smoking' than peer selection. However, a longitudinal social network analysis in the original ASSIST trial (Mercken et al., 2012) found that selection homophily explained a greater proportion of smoking behavior similarity than peer influence. This led to the recommendation that future research consider both processes. Chu et al. (2020) used agent-based modeling, considering social contagion but not selection homophily and peer influence explicitly. Disentangling these two processes is a challenge. Krupka et al. (2016) studied their effects on university freshmen's economic preferences. This paper aims to build on this work by exploring the behavioral mechanisms underlying the influence of social norms on adolescent smoking and vaping. This study uses a range of statistical methods: mixed-effects logistic regressions to investigate selection homophily (whether similarity in outcomes increases the likelihood of friendship), OLS regressions to assess peer influence (effects of friends', class', and year group's average responses), CLPMs to examine cross-lagged effects simultaneously, and SIENA models which account for network dynamics. Previous research often used regression-based approaches to study these effects separately, potentially leading to inflated estimates of peer influence (Ragan et al., 2019). The study's methodology has broader relevance for studying health-related behaviors.

The MECHANISMS study employed a pre-post quasi-experimental design, involving 12 schools (six in Northern Ireland and six in Bogotá). Full school year groups (target age 12-13) participated. Schools were assigned to one of two smoking prevention programs: ASSIST (peer-focused) or Dead Cool (conventional classroom-based). Data were collected over one semester, using incentivized experiments and self-report surveys. Incentivized experiments: The game theory experiments included several incentivized tasks. Part 1 assessed social norms sensitivity. Parts 2-3 elicited injunctive (beliefs about what actions are appropriate) and descriptive (beliefs about actions taken by peers) norms for smoking and vaping using coordination games. Participants received payments if their responses matched the modal answer in their school year group. Part 4 measured willingness to pay to support anti-smoking norms. Self-report survey and carbon monoxide measurements: A validated survey collected socio-demographics, friendship networks, self-report smoking outcomes, and psychosocial antecedents. Carbon monoxide monitors objectively measured smoking behavior. Social networks data: Participants named up to ten closest friends. Statistical analysis: Analyses used Stata and R. Selection homophily was examined using mixed-effects logistic regressions with binary outcome variables indicating friendship nominations, adding friends, or dropping friends. Predictor variables included absolute differences in outcome scores between focal participants and potential friends. Peer influence effects were assessed using OLS regressions, examining the impact of average peer group responses (friends, classmates, year group) on focal participants' outcomes. CLPMs examined cross-lagged and autoregressive effects between participants' and friends' outcomes. Finally, SIENA models were used to simultaneously estimate selection homophily and peer influence, accounting for network dynamics.

Mixed-effects logistic regressions revealed significant selection homophily effects, with reduced odds of friendship nominations/adding friends for greater differences in several outcomes (experimental injunctive and descriptive norms, donations, self-report norms, smoking behavior, etc.). The odds of adding/dropping friends were influenced by matching smoking susceptibility status. OLS regressions demonstrated positive peer influence effects from baseline and follow-up responses of friends, classmates, and year groups on various outcomes (experimental norms, self-report norms, smoking behavior, intentions, knowledge, attitudes, self-efficacy, perceived risks, etc.). Effects from friends were generally larger than from more distal peers. Adjusting models for setting affected results due to multicollinearity. CLPMs indicated both peer influence and selection homophily operated simultaneously for various outcomes. Some outcomes showed only significant peer influence or selection homophily effects. SIENA models, via meta-analysis, showed significant peer influence effects for experimental injunctive norms, donations, intentions, and objectively measured smoking behavior. Selection homophily effects approached significance for self-report descriptive norms scale 2, self-report smoking behavior, and self-efficacy opportunity. There were no significant differences in effect estimates across all schools. Subgroup analyses revealed significant differences in selection homophily and/or peer influence across settings and interventions for several outcomes (e.g., higher peer selection in Bogotá, higher peer influence in NI). Moran's I decomposition showed comparable contributions of selection homophily (32.8%) and peer influence (39.2%) to similarity in smoking/vaping outcomes between friends, with higher proportions for ASSIST schools.

The study's findings highlight the importance of both selection homophily and peer influence in shaping adolescent smoking and vaping norms and behaviors. The use of experimental methods to measure norms provided valuable insights into shared beliefs and expectations among peers. The consistent observation of peer influence effects across different analytical methods emphasizes its significance. The higher percentage of similarity between friends attributable to selection homophily and/or peer influence in ASSIST schools aligns with the program's design and prior research. The differences found across settings suggest the need for culturally tailored interventions, considering the varying importance of selection homophily and peer influence across different contexts. The relatively even distribution of peer influence effects between baseline and follow-up suggests the importance of both lagged and contemporaneous influences on adolescent smoking outcomes. The study's comparison of various statistical methods strengthens its conclusions, offering insights into the limitations and strengths of each approach. Although regression models can assess distal peer influence, SIENA models offer superior control for network dynamics.

This study demonstrates the substantial influence of peer selection and peer influence on adolescent smoking and vaping outcomes. The findings support social norms approaches in interventions and highlight the need for culturally tailored programs that address both mechanisms. Future research should explore moderators of peer influence and investigate generalizability across diverse settings and populations. Further exploration of alternative models like the Causal Attitude Network (CAN) model may deepen understanding of attitude formation and behavior in this context.

The relatively small number of schools included in the study limits the generalizability of the findings. The use of complete case analysis might have introduced bias. The large number of statistical tests conducted requires careful interpretation of results, even after accounting for multiple testing. While efforts were made to address these limitations, they warrant consideration when interpreting the results.