Selection homophily and peer influence for adolescents' smoking and vaping norms and outcomes in high and middle-income settings

The study addresses how adolescent smoking and vaping norms and behaviors spread through school friendship networks, focusing on two mechanisms: selection homophily (adolescents befriend similar peers) and peer influence (peers shape behaviors and attitudes). Against a backdrop of high global tobacco use, rising adolescent vaping, and the effectiveness of school-based, social norms interventions, the research examines these mechanisms in both a high-income (Northern Ireland, UK) and an upper-middle-income (Bogotá, Colombia) context. It aims to test whether experimentally elicited (behavioral economics-based) injunctive and descriptive norms, as well as self-reported smoking-related outcomes and objective smoking measures, are shaped by peer selection and influence, and whether these processes differ by setting and intervention type (ASSIST vs Dead Cool).

Prior work shows smokers cluster together via both selection and influence, with cultural context moderating effect sizes (stronger in collectivistic settings). School-based prevention with social influence components is often effective, yet evidence from LMICs is limited and mixed. Earlier studies used self-reports and traditional models (e.g., mixed-effects, cross-lagged panel models) and sometimes found selection stronger than influence. Experimental norm elicitation via coordination games reduces social desirability biases and deepens understanding of norms. Network methods such as SIENA disentangle co-evolving selection and influence while accounting for network structure; however, few studies have combined experimental norms elicitation with both regression and SIENA approaches across different income settings, nor contrasted proximal (friends) and distal (class/year) peer influences.

Design: Pre–post quasi-experimental study in 12 secondary schools (6 Northern Ireland; 6 Bogotá), full year cohorts (target age 12–13). Participation: 1344/1444 (93.1%) between Jan–Nov 2019. Schools received one of two proven prevention programs over one semester: ASSIST (peer-led, leveraging influence) or Dead Cool (classroom pedagogy). Ethics approvals from Queen’s University Belfast (18:43) and Universidad de los Andes (937/2018). Culturally adapted Spanish materials for Bogotá. Measures: - Experimental norms elicitation (behavioral economics/game theory): • Part 1: Rule-following task (norm sensitivity) – number of balls placed in the rule-following bucket (0–50). • Part 2: Injunctive norms via incentivized coordination ratings (6-point scale mapped to −1 to +1) for 8 scenarios (P2S2–P2S9), e.g., adult smoking near children, students vaping or posting vaping images, chewing tobacco. • Part 3: Descriptive norms via incentivized coordination on perceived peer acceptance of friend smoking/vaping (P3Q1–P3Q2; −1 to +1). • Part 4: Willingness to pay (0–10 tokens) to support the program (donation to ASSIST/Dead Cool). Incentives: fixed participation plus performance-based payments; modal responses within the school year group define norms. - Survey: demographics, self-reported injunctive norms (IN1–IN7), descriptive norms (DN1.1–DN1.5; DN2.1–DN2.3), smoking behavior, intentions, susceptibility, knowledge, attitudes, self-efficacy (emotional, friends, opportunity), perceived risks (physical, social, addiction), perceived benefits, perceived behavioral control (PBC: easy to quit; avoid smoking). Validated items used. - Objective smoking: exhaled CO (ppm) via Smokerlyzer. - Friendship networks: each student nominated up to 10 closest friends in the year group; nominations matched to IDs (automated + manual), enabling network construction at baseline and follow-up. Statistical analysis: - Objective 1 (Selection homophily): Mixed-effects logistic regressions predicting (1) baseline friend nominations, (2) adding friends, (3) dropping friends, from absolute differences in focal vs potential friend outcomes (original scales and re-scaled 0–10), with individual random intercepts and clustered SEs; also models for matching smoking susceptibility. - Objective 2 (Peer influence): OLS regressions with robust SEs predicting focal outcomes at follow-up from average peers’ outcomes (friends, class, year) at baseline (lagged) and follow-up (contemporaneous), controlling for gender, age, intervention, ethnicity, SES, and baseline outcome. Smoking susceptibility modeled via logistic regression (odds for a 10% increase in susceptible peers). Sensitivity analyses included ordered logit for ordinal outcomes and reciprocated friends only; models adjusted for setting. - Objective 3: Cross-lagged panel models (CLPMs) using lavaan to estimate cross-lagged paths between focal outcomes and average friends’ outcomes (peer influence: friends_baseline → focal_follow-up; selection: focal_baseline → friends_follow-up), with robust SEs and FIML; model fit via CFI, RMSEA, SRMR. - Objective 4: SIENA stochastic actor-oriented models (RSiena) for each school modeling co-evolution of friendships and behavior (experimental norms scales, donation, self-report scales, behaviors, CO, susceptibility), including selection homophily and peer influence effects plus network structural and covariate controls. Estimates combined via meta-analysis (Snijders & Baerveldt method) with Fisher’s one-sided tests (p≤0.005) and heterogeneity tests; subgroup comparisons (setting; intervention). Simulations (500 per model per school) decomposed Moran’s I to quantify proportions of similarity due to selection, influence, undetermined, and control mechanisms. Significance threshold p≤0.01 due to multiple testing; Holm-Bonferroni noted in tables.

- Selection homophily (Objective 1): Greater absolute differences between pupils on many outcomes reduced odds of nominating someone as a friend at baseline and of adding them between waves; differences increased odds of dropping friends (ORs typically 0.87–0.99 per unit difference for nomination/adding; 1.03–1.19 for dropping; p≤0.01). Matching on smoking susceptibility increased odds of nomination/addition (e.g., OR=1.20 at baseline; OR=1.16 adding; OR=1.26 at follow-up; p≤0.001). - Peer influence (Objective 2): Significant positive associations between focal outcomes and peers’ averages across friends, classes, and year groups, both lagged and contemporaneous, for numerous outcomes. Standardized betas ranged approximately 0.07–0.55; for susceptibility, a 10% increase in susceptible peers increased odds at follow-up (friends OR=1.14; class OR=1.17; year OR=1.31; all p≤0.01). Objectively measured CO showed strong influence from friends/class/year. - CLPMs (Objective 3): Both peer influence and selection paths were often significant simultaneously for experimental injunctive norms (notably P2S2, P2S5, P2S7, P2S8 and the injunctive scale) and several self-report outcomes. Some outcomes showed only influence (e.g., experimental descriptive norms P3Q2, injunctive P2S6, attitudes, perceived addiction risks, CO), while others showed only selection (e.g., self-report smoking behavior, self-efficacy friends, PBC avoid, susceptibility); in several of these, the other path approached significance (p≈0.02). - SIENA meta-analyses (Objective 4): Significant peer influence effects for experimental injunctive norms (b=3.95, SE=1.03, p<0.0001), donations (b=4.13, SE=0.43, p<0.0001), intentions (b=5.50, SE=3.72, p=0.0023), and objectively measured CO (b=8.12, SE=1.48, p<0.0001). Peer selection homophily significant for smoking susceptibility (b=0.17, SE=0.06, p=0.0017); approached significance for self-report descriptive norms scale 2 (b=0.38, p=0.0176), self-report smoking behavior (b=0.30, p=0.0074), and self-efficacy opportunity (b=0.48, p=0.0111). Limited heterogeneity across schools overall. - Mechanism contributions (Moran’s I decomposition): On average across 21 outcomes, similarity between friends attributed 32.84% to selection and 39.22% to influence (undetermined 1.08%, control 26.86%). For experimental injunctive norms, influence dominated (~89%). For CO, influence ~90%. For susceptibility, selection ~54%. - Subgroups: ASSIST schools showed higher combined selection/influence contributions to similarity than Dead Cool (74.6% vs 53.8%). Selection tended to be stronger in Bogotá (e.g., CO selection higher in Bogotá), while influence tended to be stronger in Northern Ireland for some outcomes (e.g., intentions). - Pre–post changes (Table 1): Several experimental norms items/descriptive norms and knowledge improved (e.g., descriptive norms became less accepting; knowledge increased), with some decreases in self-efficacy and increases in CO readings; see detailed z-tests with multiple-testing notes.

Findings demonstrate that both selection homophily and peer influence shape adolescents’ smoking/vaping-related norms and behaviors within school networks. Experimentally elicited injunctive norms—measured via incentivized coordination games capturing shared beliefs—were especially sensitive to peer influence, aligning with theories that norms are inherently social and diffuse through networks. Self-reported outcomes showed mixed patterns, reflecting measurement differences and specific referent groups. Objective CO measures evidenced strong peer influence, indicating behavioral convergence through networks. Comparing regression, CLPM, and SIENA approaches showed convergent evidence for both mechanisms, while also highlighting method-specific sensitivities (e.g., SIENA’s focus on similarity to friends vs regressions capturing distal peer contexts). Differences by intervention (stronger network-mediated similarity in ASSIST) and setting (selection stronger in Bogotá; influence stronger in NI for selected outcomes) suggest context and program design moderate mechanisms of diffusion. Overall, the results support the utility of social norms strategies in prevention and the need to consider both proximal and distal peer contexts and both selection and influence mechanisms.

The study provides multi-method evidence that adolescent smoking/vaping norms and behaviors are jointly shaped by selection homophily and peer influence within school networks. Experimentally elicited injunctive norms and objective smoking behavior show robust peer influence, while susceptibility displays stronger selection. Comparable proportions of friend similarity were attributable to selection and influence overall, with larger network-mediated effects in ASSIST schools than Dead Cool and differing dominant mechanisms between Northern Ireland and Bogotá. These findings endorse social norms-based prevention strategies and underscore the importance of modeling both mechanisms together. Future research should test moderators (setting, intervention, personality, network position), assess generalizability to other cultural and income contexts, and further integrate experimental norm elicitation with advanced network modeling.

- Quasi-experimental pre–post design without randomized assignment of interventions at the study level; small number of schools limits generalizability. - Multiple testing risks; stricter p≤0.01 threshold applied and Holm-Bonferroni noted, but type I/II trade-offs remain. - Regression models cannot fully account for endogenous network processes; contemporaneous influence models lack temporal precedence. SIENA addresses network dynamics but does not capture distal peer (class/year) influence and defines influence via similarity rather than peer averages. - Some SIENA models required parameter constraints or excluded schools due to non-convergence. - Complete case analyses may exclude some nominations with missing attributes, though participation and completion rates were high. - Cultural adaptation may not capture all contextual nuances despite careful translation and tailoring.