The MECHANISMS study experimental protocol used financially incentivized ‘co-ordination’ games to elicit injunctive social norms – what people ought to do – as shared perceptions within whole school year groups on the social appropriateness of various smoking and vaping-related scenarios (Krupka and Weber, 2013). Descriptive norms – what people actually do – were elicited as shared perceptions of the rate of acceptance of smoking or vaping behaviors within the school year group (Krupka and Weber, 2013). To encourage participants to consider how most others in their year group would respond (i.e., the norm) instead of providing personal opinions, they were informed that they would receive a payment if their answer matched the most common answer provided in their school year group (Krupka and Weber, 2013; Murray et al., 2020). That is, we asked the pupils to guess how most of their classmates would respond to the smoking-related scenarios, rewarding them if they matched the majority answer. This way, we measured group perceptions rather than just individual opinions. One of the advantages of these experimental methods compared to the traditional self-reports used in public health research is that they mitigate social desirability biases since respondents must report their beliefs about others’ beliefs rather than answering personally (Murray et al., 2020; Mackie et al., 2015). Another advantage is that the method is theoretically intuitive since the existence of such shared “second-order” beliefs is a necessary pre-condition for the existence of a social norm (Bicchieri et al., 2018). On the other hand, self-report methods of measuring social norms have the advantages of simplicity, low cost, and ease of distribution (Murray et al., 2020). Furthermore, the experimental and self-report measures used in our study focus on different aspects of norms. The experiments inquire about the beliefs of the whole school year group whilst the self-report methods ask about influences amongst the respondent’s own family, friends and peers (Murray et al., 2020).

The experimental norms measures can also potentially provide richer insights into intervention mechanisms to better explain variation in individuals’ behaviors within and between different contexts (Kimbrough and Vostroknutov, 2016; Kimbrough and Vostroknutov, 2018; Krupka and Weber, 2013; Murray et al., 2020). For example, the experiments included a measure of individuals’ sensitivities to social norms, or rule-following propensity (Kimbrough and Vostroknutov, 2016; Kimbrough and Vostroknutov, 2018). The task instructed participants to follow an arbitrary rule when doing so imposed explicit monetary penalties directly proportional to the degree of rule-following. In principle, the more a participant cares intrinsically about rule-following the more willing he/she should be to incur the costs of doing so (Kimbrough and Vostroknutov, 2016). Put simply, the task measured how much individual pupils valued following rules and norms in terms of how willing they were to forgo a payment to obey the rule. This norm sensitivity measure has been shown to correlate with willingness to follow norms of cooperation, reciprocity, and pro-social behavior in different decision contexts (Kimbrough and Vostroknutov, 2016). According to the behavioral economics theory, individuals with higher norms sensitivities are more likely to conform to the norms within their social context (Kimbrough and Vostroknutov, 2016; Kimbrough and Vostroknutov, 2018; Krupka and Weber, 2013; Murray et al., 2025a).

The MECHANISMS study is the first to use coordination games to measure norms around adolescent smoking and vaping, and to investigate how individuals’ conformity to social norms varies with this norm-sensitivity measure (Hunter et al., 2020).

Tobacco consumption is the leading preventable risk factor for chronic disease and mortality worldwide, responsible for over seven million annual deaths from direct consumption and 1.2 million from second-hand smoke (World Health Organization, 2025). Most adult smokers start smoking during adolescence (Institute of Medicine, 2015), a developmental stage when susceptibility to social influences is heightened (Foulkes and Blakemore, 2016). During adolescence, many young people take their cues consciously or subconsciously from observing the beliefs, actions, and attitudes of friends, peers and family (Littlecott et al., 2019; Allen and Feigl, 2017; Vitória et al., 2011). Whether it is due to peer influence – a social process where an individual’s behavior or attitudes are affected by observing peers within social networks – or through selection homophily processes – the tendency for individuals to form friendships with others who share similar characteristics and behaviors – research consistently shows that adolescent smokers usually have more smoking friends, whilst non-smokers have more non-smoking friends (Liu et al., 2017; Steglich et al., 2012; Krupka et al., 2016; Montgomery et al., 2020). In other words, smokers often end up surrounded by other smokers, either because they are drawn to friends who already smoke, or because they are influenced by smoker friends to start smoking.

Numerous studies have highlighted the importance of peer influence and peer selection homophily in shaping adolescent smoking outcomes, mostly focusing on smoking behavior, intentions, and susceptibility (Montgomery et al., 2020; Watts et al., 2024; Hoffman et al., 2007; Mercken et al., 2012; Lahiri et al., 2024; Trucco, 2020). Within the MECHANISMS study, we have contributed to this literature by investigating these social network processes with respect to our study’s novel experimental measures of adolescent smoking norms and a range of psychosocial antecedents of smoking (e.g., attitudes, self-efficacy, and perceived risks and benefits) (Montes et al., 2023; Murray et al., 2023). We also compared peer influence estimates derived from Simulation Investigation of Empirical Network Analysis (SIENA) models with conventional regression-based methods that are more common in public health and behavioral economics research (Hoffman et al., 2007; Mercken et al., 2012; Flashman and Gambetta, 2014; Fowler and Christakis, 2008; Go et al., 2012; Miething et al., 2016; Parkinson et al., 2018; Rohrer et al., 2021; Mercken et al., 2009; Ripley et al., 2025; Steglich et al., 2010; Snijders et al., 2010; Mercken et al., 2010; Ragan et al., 2019). One advantage of the regression-based approach was that it allowed us to distinguish the influence of proximal peers (e.g., close friends) from more distal peers (e.g., pupils in the same school class or year group), and to assess both lagged and contemporaneous peer influence effects (Murray et al., 2023). Specifically, we tested whether participants’ outcomes at follow-up were predicted by the average responses of their nominated friends, school classes, or year groups, finding positive peer influence effects across most study outcomes (Murray et al., 2023).

Given the overwhelming evidence that social influences are central in determining adolescent smoking uptake, many smoking prevention programs target young adolescents (typically aged 12–13 years) and use social norms or peer influence approaches (Campbell et al., 2008; Thomas et al., 2015; Ahmed et al., 2018). Social influence-based interventions have been effective for preventing adolescent smoking uptake in high-income settings, but reviews have highlighted a lack of evidence from low and middle-income countries (LMICs) (Thomas et al., 2015; Munabi-Babigumira et al., 2012; Huriah and Dwi, 2020; Ba-Break et al., 2023; Nishio et al., 2018; Macarthur et al., 2016; Brown et al., 2014). They recommend that high-quality studies should be conducted in LMICs, incorporating successful strategies from high income settings, appropriately adapted for local culture and conditions (Munabi-Babigumira et al., 2012). This is particularly important given that smoking rates are declining in high-income countries but remain high in LMICs, which now account for over 80% of the world’s 1.3 billion tobacco users (World Health Organization, 2025).

Measuring individuals’ norm sensitivities as part of our experimental protocol reflects an important component of social influence theory. Namely, the extent of attitude or behavior change that occurs through social influence is largely due to variation in individuals’ susceptibilities to social influences (Stacy et al., 1992; McGuire, 1968). Social influences may only have a strong impact on the behavior of individuals and groups with characteristics (e.g., personality, contextual, cultural, and environmental traits) that make them susceptible to social influences (Stacy et al., 1992).

For ‘socially contagious’ behaviors like smoking (Littlecott et al., 2019; Allen and Feigl, 2017), it is important to establish how certain moderating variables may interact to vary the impact of social influence on behavior. Moderators are defined as qualitative (e.g., setting) or quantitative (e.g., rule-following) variables that affect the direction and/or strength of the relationship between a predictor variable and an outcome (Baron and Kenny, 1986). In the current study, a moderating effect implies that the effect of peer influence on smoking norms or outcomes varies at different values of the moderator and is indicated by a significant interaction effect between the moderator and predictor over and above the additive effects of the two variables (Cohen and Cohen, 1983). In simple terms, if individuals’ norm sensitivities act as a moderator, this means individuals who are more sensitive to social norms are more likely to be influenced by their peers’ smoking behaviors.

Previous studies have investigated factors like gender (McCoy et al., 2017; Lansford et al., 2009), social network positions and structure (Lansford et al., 2009; Haynie, 2015), personality traits (Slagt et al., 2015; Marschall-Lévesque et al., 2014), and context (Marschall-Lévesque et al., 2014), as potential moderators of peer influence in adolescent risk behaviors. One study conducted in the United States in 1992 investigated personal characteristics such as gender, self-efficacy, self-esteem, parental supervision, and perceived stress, as potential moderators of peer influence in adolescent smoking and found significant moderating effects of self-efficacy (Stacy et al., 1992). Having higher levels of self-efficacy (i.e., greater belief in your own ability to resist social influence for smoking) reduced the strength of social influence from friends on individuals’ smoking behavior (Stacy et al., 1992). A recent review of moderators of peer influence for adolescent substance use identified that ten of the 43 included studies investigated tobacco use (Rodríguez-Ruiz and Espejo-Siles, 2025). The authors found evidence that peer influences for adolescent substance use are moderated by a range of individual, family, school, peer, and community factors including emotional control and anxiety, peer proximity and reciprocity, closeness to parents, siblings’ willingness to use substances, school disapproval and school troubles, peer support, and neighborhood characteristics.

Moderation analyses can provide answers for important public health questions, including which participants benefit most from intervention strategies, and in what contexts (Sheeran et al., 2017). Investigating potential moderators of social influences for adolescent smoking could have important implications in terms of how smoking prevention programs are designed and implemented. For example, knowledge of effective moderating variables can help with identifying individuals and groups who are most susceptible to social influences for smoking or identifying the most efficient individuals within social networks to recruit as peer leaders to spread anti-smoking messages. Medical Research Council guidance on developing complex interventions highlights that the context within which an intervention operates is a critical factor that can act as a barrier or facilitator to its implementation or effectiveness (O’Cathain et al., 2019; Moore et al., 2015). Recent research also suggests that peer processes in adolescent smoking may vary as the tobacco control context and societal norms change when countries introduce tobacco control legislation (Littlecott et al., 2023; Littlecott et al., 2022). Having been designed to compare norms and social network-based intervention mechanisms between schools in a high-income setting (NI) with schools in a middle-income setting (Bogotá), the MECHANISMS study provides a unique opportunity to investigate the moderating effects of context on peer influences for adolescent smoking. The experimental protocol also proposes that peer influences should be moderated by individuals’ norms sensitivities (Hunter et al., 2020).

In the current paper, we investigated hypothesized moderators of peer influence effects for our adolescent smoking and vaping outcomes, by including interactions between peer-group means (nominated friends, school classes, and school year groups) of the outcome variables and the moderators in regression models (Hayes, 2013). Our outcomes included experimental smoking and vaping norms, self-report smoking norms, self-report and objectively measured smoking behavior, intentions, susceptibility, knowledge, attitudes, self-efficacy, perceived risks and benefits, and perceived behavioral control (PBC). Our hypothesized moderators included setting (NI versus Bogotá), intervention (ASSIST versus Dead Cool), gender, school socio-economic status (SES), individuals’ norm sensitivities and other socially oriented personality characteristics, the ‘Big Five’ personality traits (Morizot, 2014; Ortet et al., 2017), social network parameters, and self-efficacy to resist smoking.

The MECHANISMS study is a pre-post quasi-experimental study (Hunter et al., 2020). Twelve schools (N = 6 NI, N = 6 Bogotá; n = 1344/1444 pupils, participation = 93.1%) participated in the MECHANISMS study between January and November 2019 (Hunter et al., 2020). Study procedures have previously been described (Hunter et al., 2020; Murray et al., 2020; Murray et al., 2023; Sánchez-Franco et al., 2021). Full school year groups were recruited in each school (NI Year 9, Bogotá Year 7; target age 12–13 years). Schools were assigned to the ASSIST or Dead Cool programs (Campbell et al., 2008; Thurston et al., 2019). During one school semester, pupils received the smoking prevention programs, and completed incentivized (monetary) norms elicitation experiments (Kimbrough and Vostroknutov, 2016; Kimbrough and Vostroknutov, 2018; Krupka and Weber, 2013) and self-report surveys, before and after the programs. All data collection was conducted via Qualtrics (Qualtrics, Provo, Utah, USA) on tablet computers. Prior to implementation in Bogotá, all study materials and intervention programs were thoroughly culturally adapted, including translation into Spanish language (Sánchez-Franco et al., 2021). Ethics approval was granted from Queen’s University Belfast (21 September 2018, reference 18:43) and Universidad de los Andes (30 July 2018, reference 937/2018). Participants and parents provided informed consent, and data collection procedures complied with institutional guidelines. Further details on study procedures are available in Supplementary File 1. The study flow diagram is shown in Supplementary File 1: Figure S1.1, and participants’ baseline characteristics are shown in Supplementary File 1: Table S1.1.

NI is a high-income country in the UK (The World Bank, 2020b), with around 2 million inhabitants (Northern Ireland Statistics and Research Agency, 2019). In 2022, current cigarette consumption rates were 2.2% for adolescents aged 11–16 years in NI (1.0% reported smoking at least once a week and 7.6% reported having smoked tobacco at least once before). Current e-cigarette consumption rates were 9.2% (6.3% reported vaping at least once a week and 21.3% reported having vaped at least once before) (Northern Ireland Statistics and Research Agency, 2023). In the UK, it is illegal to sell tobacco products to or buy tobacco products for minors under the age of 18 years (UK Government, 2014). Tobacco education is a formal part of the UK school curriculum (National Institute for Health and Clinical Excellence (NICE), 2023). The UK has comprehensive tobacco control legislation regulating tobacco advertising, sale to minors, packaging, smoke-free public places, and in-door smoking (UK Government Office for Health Improvement and Disparities, 2021; UK Government Department of Health, 2017; Action on Smoking and Health (ASH), 2017). It is also illegal to sell e-cigarettes to under 18s. Whilst the UK government has sought to maximize the potential of e-cigarettes as a smoking cessation aid for adults, they are currently introducing stricter legislation like the tobacco control measures, to regulate sale to minors and ban advertising and sponsorship of e-cigarettes and other nicotine products (UK Government Department of Health, 2024; UK Government Department of Health, 2022).

Bogotá is the capital city of Colombia, an upper-middle-income country (The World Bank, 2020b), with over 7 million inhabitants (National Administrative Department of Statistics, 2019). In 2022, current cigarette consumption rates were 4.5% for adolescents aged 12–18 years across Colombia (11.1% reported having used tobacco or cigarettes at least once before). In Bogotá, current cigarette consumption rates were 6.2%. Current e-cigarette consumption rates were 11.2% for adolescents aged 12–18 years across Colombia (22.7% reported having used e-cigarettes at least once before). In Bogotá, current e-cigarette consumption rates were 12.1% (Ministry of Justice and Law – Colombian Drug Observatory, Ministry of National Education, 2022). Colombia adopted the WHO-FCTC in 2009 which includes legislation regulating tobacco advertising, packaging, sale to minors, and smoke-free public places (Ministerio de Salud y Protección Social, 2009). In Colombia, selling tobacco products to minors under the age of 18 years is illegal. This has a high level of compliance within the regulatory sphere. The legislation also includes a complete ban on tobacco advertisements, sponsorships, and promotions, which is highly implemented in television, cinemas, and banners. However, adolescents can still access contraband cigarettes or purchase them from street vendors (Colombia Ombudsman Office, 2017). At the time of MECHANISMS data collection, Colombia’s public policy for tobacco control did not include e-cigarettes, which were unregulated until 2024 (Malagón-Rojas, 2024).

The ASSIST and Dead Cool programs have previously been shown to effectively reduce rates of adolescent smoking initiation (Campbell et al., 2008; Thurston et al., 2019). The ASSIST program is based on the diffusion of innovations theory (Rogers, 2003), and works on the principles of peer education and diffusion. It is designed to train the most influential pupils in the school year group, nominated in a ‘Peer Questionnaire’ at baseline, to use informal contacts with their peers – other pupils in their school year group – to encourage them not to smoke (Campbell et al., 2008). Dead Cool is a skills-based program based on the theory of planned behavior (Ajzen, 1991). It includes training for schoolteachers and provision of program resources – lesson plans, pupil workbooks, fact sheets and a DVD – to enhance pupils’ knowledge of potential influences on smoking from family, friends, and the media (Thurston et al., 2019; Dunne et al., 2016). Pupils participate in eight classroom-based sessions during which they watch DVD clips of adolescents discussing smoking-related issues, and complete various workbook and group activities (Thurston et al., 2019; Dunne et al., 2016).

The behavioral economics and game theory experiments included several tasks that used monetary incentives as part of the experimental design (Kimbrough and Vostroknutov, 2016; Kimbrough and Vostroknutov, 2018; Krupka and Weber, 2013). Part 1 included a rule-following task measuring individuals’ sensitivities to the effects of social norms (Kimbrough and Vostroknutov, 2016; Kimbrough and Vostroknutov, 2018). Participants were given 5 min to sequentially allocate 50 balls across two buckets (one blue and one yellow), following an arbitrary rule with explicit monetary costs. Participants were told that “The rule is to put the balls in the blue bucket” (Kimbrough and Vostroknutov, 2018). They were also informed that they would receive £0.05 (NI; COP $100 Bogotá) for every ball they put in the blue bucket and £0.10 (NI; COP $200 Bogotá) for every ball they put in the yellow bucket. Individuals’ norm sensitivities were elicited as the number of balls allocated to the rule-following bucket (‘rule-following’). Parts 2 and 3 included incentivized co-ordination games measuring injunctive and descriptive norms for smoking and vaping in whole school year groups (Krupka and Weber, 2013). Participants were informed they would receive a payment if their response to a randomly selected question matched the most common answer in their school year group. Injunctive norms, defined as shared beliefs about what actions people ought to take (Krupka and Weber, 2013), were assessed by asking participants to rate the social appropriateness of eight smoking and vaping-related scenarios (“extremely socially inappropriate” to “extremely socially appropriate”; P2S2–9). Descriptive norms, defined as shared beliefs about what actions people actually do take (Krupka and Weber, 2013), were assessed by asking participants to estimate the proportion of peers in the year group who would be accepting of a close friend smoking or vaping (“none of my peers” to “all of my peers”; P3Q1–2). Part 4 assessed participants’ willingness to pay to support prevention interventions that promote anti-smoking norms. Participants were given ten virtual tokens of equal monetary value, asked how many they wanted to donate to the smoking prevention program delivered in their school, and informed they would receive a payment equal to the amount not donated (‘Donation to ASSIST/Dead Cool’). Participants received participation fees of £5.00 (NI; COP$5.000 Bogotá) and could earn money in each part of the experiment (maximum £30 NI, COP$50.000 Bogotá) depending on their answers and answers provided by their peers. Payments were received after the follow-up experiment.

Further details on the experimental protocol and outcomes are available in the Supplementary File 1: Methods and Supplementary File 1: Table S1.2. Higher numerical values represented greater norm sensitivities (rule-following), more pro-smoking injunctive and descriptive norms, and higher donations to ASSIST/Dead Cool. Table 1 shows the smoking and vaping scenarios.

Our survey collected socio-demographic information (gender, age, ethnicity, SES), social networks, self-report smoking outcomes, and personality characteristics. In NI, SES for schools and individual pupils was based on the Northern Ireland Multiple Deprivation Measure (NIMDM2017), which ranks postcodes based on seven domains of deprivation (1 = most deprived to 890 = least deprived) (Northern Ireland Statistics and Research Agency, 2017). In Bogotá, SES for schools was based on the socio-economic level index for educational institutions provided by the Colombian Institute for the Promotion of Higher Education (1 = lower to 4 = higher) (Colombian Institute for the Evaluation of Education, 2017). In Bogotá, SES for individual pupils was determined as the socio-economic level index of the household provided by the Colombian National Administrative Department of Statistics (0 = informal settlement, 1 = lowest to 6 = high) (National Administrative Department of Statistics, 2021).

Social networks were assessed by asking pupils to name up to ten of their closest friends in their school year group (Murray et al., 2023; Dunne et al., 2016). Pupils were provided with class rosters and asked to provide the first name, surname, and form class of nominated friends (data were anonymized by the study team).

Self-report smoking outcomes included: injunctive norms (Cremers et al., 2012), descriptive norms (Cremers et al., 2012), past/current smoking behavior (Dunne et al., 2016; Fuller and Hawkins, 2012), intentions and susceptibility (Dunne et al., 2016; Mazanov and Byrne, 2007; Pierce et al., 1998), knowledge of the health effects of smoking (Cremers et al., 2012), attitudes towards smoking (Ganley and Rosario, 2013), self-efficacy to resist smoking (emotional, friends, and opportunity subscales) (Condiotte and Lichtenstein, 1981; Lawrance, 1989), perceived risks of smoking (physical, social, and addiction subscales) (Halpern-Felsher et al., 2004; Song et al., 2009; Aryal et al., 2013), perceived benefits of smoking (Halpern-Felsher et al., 2004; Song et al., 2009; Aryal et al., 2013), and PBC (‘easy to quit smoking’, and ‘to avoid smoking’) (Smith et al., 2006). Self-report injunctive norms were assessed with seven items enquiring about perceived approval of smoking from groups of important others (e.g., “most of the people who are important to me,” “mother,” “father”; IN1-7) (Cremers et al., 2012). Self-report descriptive norms were assessed with five items enquiring about how often groups of important others engaged in smoking behavior, and three items enquiring about the proportion of groups of important others who are smokers (DN1.1–1.5, DN2.1–2.3) (Cremers et al., 2012).

Personality variables collected at baseline included: pro-sociality (Goodman et al., 2003; Bevelander et al., 2018), FNE (Bevelander et al., 2018; Leary, 1983; Collins et al., 2005), NTB (Bevelander et al., 2018; Leary et al., 2013), and the ‘Big Five’ personality traits (Morizot, 2014; Ortet et al., 2017).

Smoking behavior in the last 24 h was objectively measured using hand-held carbon monoxide monitors measuring expelled air carbon monoxide in parts per million (PICOAdvantage Smokerlyzer, Bedfont) (Thurston et al., 2019; Bedfont Scientific Ltd, 2018). Supplementary File 1: Table S1.2 shows details of all measurement instruments. Higher numerical values represented more anti-smoking outcomes on the self-report survey (e.g., more anti-smoking behavior, intentions, attitudes, and higher knowledge of the effects of smoking), apart from perceived benefits and smoking susceptibility (0 = not susceptible, 1 = susceptible). Higher numerical values also represented greater objectively measured smoking behavior.

The following variables were examined as moderators:

All moderator variables were measured at baseline. Social network parameters at baseline and follow-up were both considered as moderators depending on whether the peer influence effects were measured at baseline or follow-up.

Higher numerical values represented greater SES and norms sensitivities, higher levels of the personality characteristics, and greater self-efficacy to resist smoking. They also represented higher clustering, centrality, and Gini degree coefficients, corresponding to more heterogeneous networks (Supplementary File 1: Table S1.2). Definitions of the social network parameters are provided in the ‘Social network parameter definitions’ subsection of Supplementary File 1.

Analyses were conducted using Stata 13 (StataCorp) (StataCorp, 2013). Descriptive statistics, and Cronbach’s alpha coefficients for individual scales were calculated (Table 1).

Peer influence effects were examined for the following smoking and vaping outcomes which were targeted by the ASSIST and Dead Cool interventions (1): experimentally measured injunctive norms, experimentally measured descriptive norms, experimental donations to ASSIST/Dead Cool, self-report injunctive norms, self-report descriptive norms, self-report smoking behavior, intentions, knowledge, attitudes, self-efficacy, perceived risks, perceived benefits, PBC, objectively measured smoking behavior, and smoking susceptibility.

To examine influence effects from friendship networks, school classes, and school year groups, variables were computed for each outcome at baseline and follow-up, containing: (1) the average responses of each focal participant’s (i) friendship network; (2) the average responses of i’s school class, excluding i; and (3) the average responses of i’s school year group, excluding i (Krupka et al., 2016). Ordinary least square (OLS) regressions with robust (Huber-White) standard errors (Huber, 1967; White, 1980) were used to examine influence effects, and moderation of influence effects. In addition to the ‘average peer’ predictor variable, the moderator variable and its interaction with the predictor variable were included in each model. We considered a model to show a significant moderating effect if the interaction term reached statistical significance at p ≤ 0.01.

Logistic regressions were run with focal participants’ smoking susceptibility at follow-up as the outcome, and robust (Huber White) standard errors (Huber, 1967; White, 1980). Variables were computed at baseline and follow-up containing the percentage of i’s friendship network, school class, and school year group, that were susceptible to commencing smoking. In addition to the ‘percentage peer’ predictor variable, the moderator and its interaction with the predictor were included.

Baseline covariates included in each model were gender, age, intervention, ethnicity, individuals’ SES, and baseline values (Murray et al., 2023). Continuous predictor variables, moderator variables, and baseline outcomes were mean-centered. The models were repeated to examine influence effects from average peer responses at baseline and average peer responses at follow-up (Murray et al., 2023).

Variance inflation factors (VIFs) were calculated to examine multi-collinearity (>5 or 10 usually indicates problematic amounts of collinearity). VIFs for ‘setting’ were high for many of the models with school class or year group responses as predictors (Murray et al., 2023). Where high VIFs affected one of the models with setting as a moderator, this is indicated in the footnotes of Supplementary Files 2: Table S2.1 and Supplementary File 3: Table S3.1. VIFs were satisfactory for all other analyses.

Significant interactions (p ≤ 0.01) were probed using the simple slopes and Johnson-Neyman techniques (Hayes, 2013; Johnson and Neyman, 1936). For dichotomous moderator variables, the simple slopes technique was used to calculate marginal effects at each level of the moderator (NI versus Bogotá, ASSIST versus Dead Cool, and boys versus girls/PNTS). For continuous moderator variables, the marginal effects were calculated at one standard deviation below (‘low’) and above (‘high’) the mean value of the moderator. Conditional effects were graphed with their 95% confidence intervals (CIs), showing the relationship between the predictor (average ‘peer’ variable) and the predicted value of the outcome at follow-up, as a function of the moderator. Regions of significance were calculated, using the Johnson-Neyman technique (Johnson and Neyman, 1936). These methods allowed us to visualize whether the strength or direction of the peer influence effects changed depending on the value of the moderator.

Due to the large number of tests for each moderator, we adopted a fourfold approach to adjusting for multiple testing. Firstly, we have discussed our results with reference to a significance criterion of p ≤ 0.01. Secondly, in the results tables we have highlighted which results would have attained statistical significance at the p ≤ 0.05 level after using the Holm-Bonferroni procedure to adjust the p-values for multiple testing (Holm, 1979). Thirdly, we used binomial tests to determine whether the number of significant interactions (p ≤ 0.01) observed per moderator was greater than expected by chance. Fourthly, we have used multiverse-style analyses to summarize the distribution of p-values and standardized regression coefficients for interaction effects across the models for each moderator (Steegen et al., 2016; Del and Gangestad, 2021).

For the multiverse-style analyses, we used histograms to show the distribution of p-values for interaction effects for each moderator. Volcano plots were constructed showing the relationship between the p-values and standardized regression coefficients. Figures were constructed to visualize the distribution of significant interaction effects according to outcome (experimental norms outcomes, self-report norms outcomes, or other smoking outcomes), peer group (friends, school class, or school year group), and peer influence measurement time-point (baseline or follow-up) (Steegen et al., 2016; Del and Gangestad, 2021). Finally, we constructed summary tables and heatmaps to facilitate the narrative synthesis of results for each moderator.

A more detailed description of the statistical methods is provided in the Supplementary File 1: Methods.

Table 2 shows a summary of the main results including the binomial tests and our overall conclusions about each moderator. The binomial tests showed evidence that setting was a significant moderator of the peer influence effects, but they provided no indication of whether peer influence was stronger in NI or Bogotá overall. Intervention was a significant moderator, and there was some indication that the peer influence effects were stronger in ASSIST schools compared to Dead Cool overall. Gender was a significant moderator and there was evidence that the peer influence effects were stronger for girls/PNTS compared to boys.

School SES was a significant moderator across all schools, with stronger peer influence effects in schools with lower SES. When we repeated the models examining peer influence from friends separately in NI and Bogotá, we found the peer influence effects were stronger in NI schools with lower SES. Whilst there was some evidence that school SES was a significant moderator in Bogotá, there was little indication of whether the peer influence effects were stronger at higher or lower school SES overall.

Overall, we found evidence that peer influence effects were stronger for pupils with higher levels of pro-sociality, FNE, and extraversion. There was evidence that peer influence effects were stronger for pupils with higher eigenvector and closeness centralities. There were also stronger peer influence effects in school networks with less heterogeneous degree distributions, as defined by the school network Gini degree coefficient (i.e., networks in which individuals had similar numbers of connections).

There was some evidence that self-efficacy was a significant moderator, but little indication of whether the peer influence effects were stronger at higher or lower levels of self-efficacy overall.

We found no overall evidence that pupils’ experimentally measured norm sensitivities, NTB, openness, agreeableness, conscientiousness, emotional stability, clustering coefficients, or betweenness centralities were significant moderators.

The significant moderating effects were distributed evenly across models of peer influence from friends, school classes, and year groups, and across models measuring peer influence at baseline versus follow-up.

Figure 2 shows a heatmap summarizing the results for each moderator according to smoking outcomes. The significant moderation effects appear dispersed across moderators and outcomes, with no single moderator consistently dominating. The strongest moderation signals were observed for setting, intervention, gender, pro-sociality, extraversion, network centrality parameters, Gini degree coefficients, and self-efficacy. These effects were most evident for models with objectively measured smoking behavior, intentions, and perceived physical risks as the outcomes. This suggests that the research context, gender, individual dispositions, and network structures all shape how peer influence impacts adolescent smoking.

The results for models with setting, intervention, and gender as moderators are reported in Figures 3–7, Supplementary File 2: Tables S2.1–S2.3, Supplementary File 3: Table S3.1, Supplementary File 4: Figures S4.1–S4.49, and Supplementary File 5: Figures S5.1–S5.12.

Setting was a statistically significant (p ≤ 0.01) moderator in 20/276 models. Three were for influence effects from friends, six were for school classes, and 11 were for school year groups. Ten models showed more positive peer influence effects in Bogotá and ten showed more positive peer influence effects in NI.

The mean standardized regression coefficient for the interaction effect across the 276 models with setting as the moderator was −0.003 (7.2% at p ≤ 0.01, 11.6% at p ≤ 0.05, Figures 4–7).

Intervention was a statistically significant moderator in 11/276 models. Four were for influence effects from friends, four were for school classes, and three were for school year groups. Seven models showed more positive peer influence effects in ASSIST schools and four showed more positive peer influence effects in Dead Cool schools. The mean standardized regression coefficient was 0.00008 (4.0% at p ≤ 0.01, 7.6% at p ≤ 0.05, Supplementary File 5: Figures S5.5–S5.8).

Gender was a statistically significant moderator in 18/276 models. Five were for influence effects from friends, six were for school classes, and seven were for school year groups. Twelve models showed more positive peer influence effects for girls/PNTS and six showed more positive peer influence effects for boys. The mean standardized regression coefficient was 0.02 (6.5% at p ≤ 0.01, 13.0% at p ≤ 0.05, Supplementary File 5: Figures S5.9–S5.12).

The results for models with school SES as the moderator are reported in Supplementary File 2: Tables S2.4–S2.6, Supplementary File 3: Table S3.2, Supplementary File 4: Figures S4.50–S4.76, and Supplementary File 5: Figures S5.13–S5.24.

School SES (all schools) was a statistically significant moderator in 15/276 models. Five were for influence effects from friends, six were for school classes, and four were for school year groups. Nine models showed decreasing peer influence effects as school SES increased. The mean standardized regression coefficient was −0.003 (5.4% at p ≤ 0.01, 8.3% at p ≤ 0.05).

In NI, School SES was a statistically significant moderator in 8/92 models examining peer influence effects from friends. Seven showed decreasing peer influence effects as school SES increased. The mean standardized regression coefficient was −0.02 (8.7% at p ≤ 0.01, 15.2% at p ≤ 0.05).

In Bogotá, School SES was a statistically significant moderator in 4/92 models examining peer influence effects from friends. One showed decreasing peer influence effects as school SES increased. The mean standardized regression coefficient was 0.02 (4.3% at p ≤ 0.01, 9.8% at p ≤ 0.05).

The results for models with norm sensitivities and related personality characteristics as moderators are reported in Figure 8, Supplementary File 2: Tables S2.7–S2.10, Supplementary File 3: Table S3.3, Supplementary File 4: Figures S4.77–S4.109, and Supplementary File 5: Figures S5.25–S5.40.

Pro-sociality was a statistically significant moderator in 14/276 models. Five were for influence effects from friends, four were for school classes, and five were for school year groups. Eleven models showed increasing peer influence effects as pro-sociality increased. The mean standardized regression coefficient was 0.009 (5.1% at p ≤ 0.01, 12.7% at p ≤ 0.05).

FNE was a statistically significant moderator in 8/276 models. Four were for influence effects from friends, two were for school classes, and two were for school year groups. All eight models showed increasing peer influence effects as FNE increased. The mean standardized regression coefficient was 0.01 (2.9% at p ≤ 0.01, 16.3% at p ≤ 0.05).

We found little evidence that peer influence effects were significantly moderated by rule-following (individuals’ norm sensitivities) or NTB, which showed statistically significant interactions in only 5 and 6 out of 276 models, respectively.

The results for models with the ‘Big Five’ personality traits as moderators are reported in Supplementary File 2: Tables S2.11–S2.15, Supplementary File 3: Table S3.4, Supplementary File 4: Figures S4.110–S4.134, and Supplementary File 5: Figures S5.41–S5.60.

Extraversion was a statistically significant moderator in 10/276 models. Two were for influence effects from friends, four were for school classes, and four were for school year groups. Seven models showed increasing peer influence effects as extraversion increased. The mean standardized regression coefficient was −0.0001 (3.6% at p ≤ 0.01, 8.7% at p ≤ 0.05).

We found little evidence that peer influence effects were significantly moderated by the ‘Big Five’ personality traits of openness, agreeableness, conscientiousness, or emotional stability, which showed statistically significant interactions in only 6, 4, 4, and 1 out of 276 models, respectively.

The results for models with social network parameters as moderators are reported in Supplementary File 2: Tables S2.16–S2.20, Supplementary File 3: Table S3.5, Supplementary File 4: Figures S4.135–S4.176, and Supplementary File 5: Figures S5.61–S5.80.

Individuals’ eigenvector centralities were statistically significant moderators in 11/276 models. Five were for influence effects from friends, two were for school classes, and four were for school year groups. All 11 models showed increasing peer influence effects as eigenvector centralities increased. The mean standardized regression coefficient was 0.009 (4.0% at p ≤ 0.01, 12.0% at p ≤ 0.05).

Individuals’ closeness centralities were statistically significant moderators in 13/276 models. Five were for influence effects from friends, four were for school classes, and four were for school year groups. Twelve models showed increasing peer influence effects as closeness centralities increased. The mean standardized regression was 0.02 (4.7% at p ≤ 0.01, 10.9% at p ≤ 0.05).

School network Gini degree coefficients were statistically significant moderators of peer influence effects in 11/276 models. Four were for influence effects from friends, one was for school classes, and six were for school year groups. Nine models showed decreasing peer influence effects as Gini degree coefficients increased. The mean standardized regression coefficient was −0.02 (4.0% at p ≤ 0.01, 7.6% at p ≤ 0.05).

We found little evidence that peer influence effects were significantly moderated by individuals’ clustering coefficients or betweenness centralities, which showed statistically significant interactions in only 4 and 3 out of 276 models, respectively.

The results for models with the self-efficacy subscales as moderators are reported in Supplementary File 2: Tables S2.21–S2.23, Supplementary File 3: Table S3.6, Supplementary File 4: Figures S4.177–S4.195, and Supplementary File 5: Figures S5.81–S5.92.

The self-efficacy subscales were statistically significant moderators in 19/774 models. Seven were for influence effects from friends, three were for school classes, and nine were for school year groups. Seven models showed increasing peer influence effects as self-efficacy decreased.

Study strengths include the large sample size, and comparison of results between two settings (one high-income, one middle-income) that have different norms, cultural traits, regulatory contexts, network structures, and health behavior patterns. This is important given the lack of relevant research and high smoking rates in LMICs (World Health Organization, 2025; Thomas et al., 2015; Munabi-Babigumira et al., 2012; Huriah and Dwi, 2020). All study materials underwent a thorough cultural adaptation process at the start of the study in Bogotá (Sánchez-Franco et al., 2021). Our study includes a broad range of smoking and vaping-related outcomes to provide richer insights into the working mechanisms of the interventions. These include self-report and objective measures of smoking behavior, and smoking/vaping norms assessed by self-report and experimental methods. This is the first study to apply experimental methods from behavioral economics, which mitigate social desirability biases, to study norms for adolescent smoking and vaping (Hunter et al., 2020; Murray et al., 2020). We have also examined moderation of peer influence effects by a broad range of potential moderating variables.

Limitations include the relatively small sample of schools, exclusion of nominated friends with missing attribute data, and multiple testing. In each setting, we endeavored to recruit schools with a range of deprivation levels and mixed-gender. There was a high participation rate across the schools, and rates of completion for the experiments and survey were high at both time-points. Therefore, the impact of missing data should be minimal. We have accounted for multiple testing by discussing our results with reference to a significance criterion of p ≤ 0.01, highlighting which results would have attained statistical significance (p ≤ 0.05) after using the Holm-Bonferroni procedure to correct the p-values for multiple testing (Holm, 1979), and using multiverse-style analysis strategies (Steegen et al., 2016; Del and Gangestad, 2021). The issue of making adjustments for multiple testing within a study is a widely debated issue. Whilst there are no established rules or guidance, several prominent academics have made a strong case for why it is not always desirable, or even correct, to adjust for multiple testing (Feise, 2002; Perneger, 1998; Rothman, 1990). Whilst adjustments for multiple testing reduce type one error rates (the rate of falsely finding a significant result), they simultaneously increase type two error rates (the rate of falsely concluding a null result), which increases the likelihood of missing important findings. Given that our analyses aimed to test the theoretically justifiable hypotheses outlined in the introduction section, we adopted the approach of discussing all results meeting the p ≤ 0.01 criterion. This approach permitted us to discuss all of the potential implications for future norms-based intervention research, whilst highlighting which results would not have attained statistical significance using the stricter adjustment criterion. VIFs were high for several models with setting as a moderator. However, we have followed standard recommendations for conducting moderation analyses – examining interaction effects (Hayes, 2013) – and have indicated which models had high VIFs. Only one model with a significant interaction showed potentially problematic amounts of collinearity.

Our results highlight important opportunities for refining future social norms and social network-based interventions, particularly for adolescent smoking prevention in LMICs. Social norms intervention approaches that align descriptive norms – providing accurate information on how many peers smoke – with injunctive norms – emphasizing that most peers disapprove of smoking and vaping – may be especially effective in LMICs (Cialdini, 2003). There may also be stronger social influence mechanisms in how individuals evaluate potential social sanctions for norm violations in LMICs (‘metanorms’) (Eriksson et al., 2021; Axelrod, 1986). Leveraging peer influences to highlight the social consequences of smoking could be an effective strategy in LMICs. Providing gender-tailored norms information should also be explored in future prevention efforts (Lewis and Neighbors, 2015; Neighbors et al., 2010). Our findings highlight that peer-led programs like ASSIST and teacher-led programs like Dead Cool operate through distinct mechanisms (network diffusion versus direct instruction). Recognizing these pathways can inform the design of tailored interventions that combine or adapt elements of each approach to maximize impact and sustainability in different settings. Social influence-based interventions like ASSIST may be particularly beneficial in more deprived areas or in LMICs lacking strong tobacco control legislation, where smoking remains largely normalized (Littlecott et al., 2023; Littlecott et al., 2022). Finally, our results underscore the importance of adolescents’ social identity processes, sensitivity to peer evaluation, and the structural properties of school networks in shaping peer influence. Selecting peer leaders based on eigenvector or closeness centralities may enhance diffusion effects, while alternative network intervention strategies like ‘segmentation’ may be beneficial for more heterogenous networks (Hunter et al., 2020; Lansford et al., 2009; Haynie, 2015; Panter-Brick et al., 2006; Robalino and Macy, 2018; Hunter et al., 2019).