Exploring the mechanistic pathways of how social network influences social norms in adolescent smoking prevention interventions

doi:10.21203/rs.3.rs-1371071/v1

We know little about how smoking prevention interventions might leverage social network structures to enhance protective social norms. In this study we combined statistical and network science methods to explore how social networks influence social norms related to adolescent smoking in school-specific settings in Northern Ireland and Colombia. Pupils (12–15 years old) participated in two smoking prevention interventions in both countries (n = 1344). A Latent Transition Analysis identified three groups characterized by descriptive and injunctive norms towards smoking. We employed a Separable Temporal Random Graph Model to analyze homophily in social norms and conducted a descriptive analysis of the changes in the students’ and their friends’ social norms over time to account for social influence. The results showed that students were more likely to be friends with others who had social norms against smoking. However, students with social norms favorable towards smoking had more friends with similar views than the students with perceived norms against smoking, underlining the importance of network thresholds. Our results support that the ASSIST intervention takes advantage of the friendship network to make a greater change in the students' smoking social norms than the Dead Cool intervention, supporting that social norms are subject to social influence.

Smoking susceptibility

social norms

social network analysis

friendship network

adolescents

social influence.

We explored how adolescents’ social network structures, defined by friendship nominations, influence smoking social norms in the context of two smoking prevention interventions implemented in six schools in Bogotá (Colombia), and six schools in Northern Ireland (United Kingdom). We identified latent groups associated with individual smoking social descriptive and injunctive norms before and after the intervention. We examined whether social influence and homophily were observed among participants according to their friendship nominations and latent groups over time. We found that the social influence on the smoking social norms in adolescents in a prevention context is determined by the type of intervention and the social norms of friendships defined by the friendship threshold.

Globally, around 43.8 million adolescents aged 13–15 years old smoke cigarettes or use smokeless tobacco products ¹. Further, 90% of adult smokers typically start smoking as a teenager (< 19 years old) ² and the tobacco industry has strategically developed new products especially attractive to young people ³. Among adolescents, peer influence over smoking behavior is strongest ^4–6, therefore, this is a critical life stage in the prevention of tobacco consumption.

Behavior change strategies have been useful for reducing health risks among adolescents^7–9. Some behavior change theories argue that individuals' emotions, attitudes, and behaviors can exert interdependent effects on the behaviors of others ¹⁰. However, several studies have demonstrated that social norms can influence a broad range of health behaviors and behavioral changes during public health interventions ^10,11.

Social norms are the standard expectations and rules held by a group of people (e.g. what people in a group believe to be a typical and appropriate action) ¹². Beliefs about what others do are descriptive social norms, and beliefs about what others think one should do are injunctive social norms ^10,11. Social norms depend on context and social networks and operate differently for different behaviors ¹³, like smoking. Therefore, changes in social norms will influence individuals' behaviors and decisions ^14,15. Indeed, Murray et al offered evidence that high-status peers may influence levels of susceptibility to risky behavior such as smoking during adolescence ¹⁶.

There is a growing interest in understanding how we can best use social networks in health behavior change interventions ^17–19. These changes can be attributed to network processes like homophily and peer influence. Homophily refers to the existence of friendships between people who have attributes in common ²⁰. Peer influence refers to a process affecting a person’s behavior that is generated from an overt or subliminal comparison with the behavior of other people with whom that person is connected ²¹. Previous studies have shown that adolescents’ school friends, and their smoking behaviors, exert peer influence on their tendency to smoke ^22,23. Although the importance of social networks and peer influences on health behaviors among adolescents is well established [e.g. ¹⁹], there is limited evidence on how they influence normative behaviors or the mechanisms which underlie social norms-based interventions in different contexts.

Smoking prevention interventions among adolescents leverage peer influence as an important behavior change process ⁷. The Mechanisms of Networks and Norms Influence on Smoking in Schools (MECHANISMS) study ²⁴ used two school-based interventions that aim to prevent adolescent smoking but operate via contrasting theoretical frameworks and subsequent mechanistic pathways. The A Stop Smoking in Schools Trial (ASSIST) leverages social networks and social influence processes to train students to maintain preventive social norms about tobacco and smoking²⁵. By contrast, Dead Cool is a teacher-led intervention where students are taught skills to identify and manage social influences for tobacco use using didactic lessons ²⁶. The interventions targeted students aged 12–15 years old, for whom actual smoking is typically rare, and aimed to maintain the norm of non-smoking. Each intervention was conducted in three schools in Northern Ireland (NI) and three schools in Bogotá, Colombia (BOG), for a total of 12 schools. There are differences between smoking rates between high-income countries (such as NI) and low-middle income countries (such as Colombia). This study takes into account these cultural differences to analyze the two smoking prevention interventions ^24,16.

The study aims to investigate the mechanistic pathways by which social network structures influence social norms before and after the ASSIST and Dead Cool smoking prevention interventions in BOG and NI schools. The analysis included three objectives: (1) to identify transitions of students between the latent groups that are associated with perceived social norms before and after the interventions, (2) to evaluate changes of the friendship network structures over time, and (3) to assess how homophily and influence processes might affect the change in perceived social norms before and after the interventions. To our knowledge, this is the first study to explore social network structure processes (i.e. influence and homophily) on experimentally derived social norms based on game theory experiments ^16,27.

Uncovering the latent groups related to social norms

We identified latent groups and their changes over time, between baseline and follow-up after the smoking prevention interventions, for the smoking-related norms outcomes in the schools. These groups were classified according to descriptive and injunctive social norms favorable towards smoking or against smoking. We conducted an initial classification with a Component-based Feature Saliency for Clustering (CFSC) ²⁸ (SI Appendix) and followed this with a Latent Transition Analysis (LTA) to understand the transitions of participants between groups over time (see Methods for further details including the clustering techniques employed) ^29,30. The LTA identified three latent groups for each time-point (baseline and follow-up) according to the Bayesian information criterion (BIC) for dissolution that was 0.92, indicating a high certainty in the classification (SI Appendix).

Note: Nodes represent individual participants; Ties represent the friendship nomination between participants; average in degree represents the mean number of friends nominated per individual participant; density is the ratio of the number of ties observed to the number of possible ties for a given network. Jaccard Index measures the structural change between follow-up and baseline, thus there is only one coefficient per school over time. shows students’ membership of the latent groups according to the intervention. The first latent group included 116 students (11.4%) at baseline and 109 (10.7%) students at follow-up. This group was characterized by students with descriptive social norms favorable towards smoking and injunctive social norms against smoking. This suggests that students in this group believe that other people around them smoke, but they believe that people around them believe that smoking is not socially acceptable. The second latent group included 878 students (86.2%) at baseline and 885 (86.9%) at follow-up. This group was characterized by students with descriptive and injunctive social norms against smoking. This suggests that students in this group believe that other people around them do not smoke and that they believe that smoking is not socially acceptable. Finally, the third latent group included 24 students (2.4%) at baseline and 24 (2.4%) at follow-up. This group was characterized by students with injunctive social norms more favorable towards smoking. This suggests that students in this group perceive social acceptance toward smoking, regardless of whether they believe that people around them smoke or not. There was similar group membership for the two interventions.

According to the LTA results, transitions (Figure 1) reflected a change in the students who had social norms favorable towards smoking, in general. In fact, 55 students changed their social norms from being favorable towards smoking to being against smoking after the intervention (50 students representing 43% of the students at baseline in the group with descriptive social norms favorable towards smoking, and 5 students representing 21% of the students at baseline in the group with injunctive social norms favorable towards smoking). In contrast, 48 students changed their social norms to be favorable towards smoking after the intervention (5% of students who were in the group with injunctive and descriptive social norms against smoking at baseline).

By comparing transitions for the ASSIST and Dead Cool intervention classes, with a 5% significance level chi-square test, we observed that for both settings the students who received the ASSIST intervention, which is based on peer influence and purposefully utilizes the social network structure for intervention implementation, showed a greater group transition to having both social norms against smoking than students who received the Dead Cool intervention. Of the total number of students who experienced this group transition in NI, approximately 86% participated in the ASSIST intervention, while among those who made this transition in BOG, 59% participated in the ASSIST intervention.

Finally, we conducted a multinomial logistic regression (Table 2) to estimate the association of latent groups at baseline with self-reported outcomes from socio-demographic variables (age, sex, home composition, ethnic status) and psychosocial traits like self-efficacy, pro-sociality (need to belong, fear of negative valuation, pro-sociality), and personality (openness, extraversion, agreeableness, conscientiousness, stability) ²⁴. We found that age, sex, and setting were associated with group assignment (p-value < 0.05). Older students were more likely to be assigned to a group with injunctive and/or descriptive social norms favorable toward smoking (groups 1 and 3) than to group 2, in comparison to younger students (11 years old, Table 2). Girls were more likely than boys to be assigned to group 1 than to group 2 (odds ratio [OR]: 1.70, confidence interval [CI]: [1.05;2.74]). This suggested that girls were more likely to be in a group with descriptive social norms favorable towards smoking than boys. Students from NI were more likely than students from BOG to be assigned to group 1 over group 2 (OR:1.88, CI: [1.06;3.31]). This suggested that students from NI were more likely to be in a group with descriptive social norms favorable toward smoking. There were no significant group associations with home composition or ethnic status variables. We also found that the intervention received was associated with belonging to group 3 compared to group 2. Students who participated in the Dead Cool intervention were less likely to belong to a group with injunctive social norms favorable toward smoking than the students who participated in the ASSIST intervention (OR: 0.27, CI: [0.08;0.81]).

Regarding the psychological characteristics, we found that self-efficacy and three of the five personality traits (extraversion, agreeableness, and conscientiousness) were associated with group assignment, with a p-value < 0.05 (Table 2). For an increase of 0.1 on the self-efficacy scale, students were less likely to be assigned to the group with descriptive social norms favorable towards smoking (group 1) than to the group with descriptive and injunctive social norms against smoking (group 2, OR: 0.49, CI: [0.37;0.65]). For an increase on the extraversion scale, students were more likely to be assigned to the group with descriptive social norms favorable toward smoking (group 1) than to the group with descriptive and injunctive social norms against smoking (group 2, OR: 1.59, CI: [1.11;2.28]). For an increase on the agreeableness scale, students were less likely to be assigned to the group with injunctive social norms favorable toward smoking (group 3) than to the group with descriptive and injunctive social norms against smoking (group 2, OR: 0.67, CI: [0.42;1.05]). For an increase on the conscientiousness scale, students were less likely to be assigned to the group with descriptive social norms favorable toward smoking (group 1) than to the group with descriptive and injunctive social norms against smoking (group 2, OR: 0.59, CI: [0.38;0.94]). Pro-sociality characteristics (need to belong, fear of negative evaluation, pro-sociality) and two of the five personality trials (openness and emotional stability) were not significantly associated with group assignment.

Assessing the change in network structure over time

We assessed 12 directed networks, two per school, one at baseline and the other at follow-up (Figure 2). Networks were defined by the friendship nomination (ties) between the students (nodes). Social networks were similar in size across schools and settings over time (Table 3). Of the 12 school networks, the number of nominated friendship nodes decreased on average by 1.75% from baseline to follow-up (nodes ranged from 83 to 177 at baseline and from 77 to 176 at follow-up), while the average in-degree, representing the number of received friendship nominations per node, remained at 6 friends on average after the intervention.

Then, we measured the Jaccard index to assess changes in the network structure over time. The Jaccard index measures the diversity between samples and was used for its simplicity in handling continuous and categorical variables. The resulting values ranged from 0.26 to 0. (Table 3). All but one school was found to have significant changes in the structure of friendship ties over time.

Assessing the relationship between the network structure and the social norms latent groups

Given that the networks showed structural changes over time, we analyzed homophily and social influence network mechanisms related to changes in social norms. We implemented a Separable Temporal Exponential Random Graph Model (STERGM) to analyze homophily according to social norms and used descriptive analysis to analyze social influence among participants.

Formation and dissolution of ties related to social norms (homophily effects)

The STERGMs estimate the likelihood of the formation and dissolution of ties related to social variables over time ³¹. By applying the STERGM, overall, we found that the formation and dissolution of friendships had no significant association with being assigned to any particular social norms group. This implies that there is little homophily around smoking-related social norms among students. In the STERGM model outcomes, we observed that transitivity (i.e. that friends of friends are also friends) and reciprocity (i.e. that a friendship nomination is returned) measures were significant and positive concerning the formation and dissolution of friendship ties in the 12 networks (Table 4). The transitivity results suggest that students were more likely to form friendships with students who had mutual friends and to preserve friendships when there were mutual friends. The reciprocity results tell us that students were more likely to form and preserve reciprocal friendships. Social norms groups from the LTA at baseline had a significant bearing on the formation of friendship ties in 3 of the 12 networks, and social norms group at follow-up had a significant effect on the dissolution of friendship ties in 2 of the 12 networks (Table 4).

Changes in social norms associated with peer influence (Social influence effects)

We analyzed the change in the students’ social norms group over time, according to the group changes of their nominated friends. The analyzed data is shown in Figure 3. We described the group changes for friends who remained friends at both time points (n=3579 ties at baseline and follow-up), the newly nominated friends at follow-up (n=2029 ties only at follow-up), and the friends that were nominated only at baseline but not at follow-up (n=2392 ties only at baseline). We separately assessed the students according to their social norms changes. Depending on their own change (i.e. ego), we measured the average percentage of friends who changed or maintained their social norms (graphs a, b, c, and d in Figure 3Figure 2). This analysis was carried out differentiating by setting (NI or BOG) and intervention (represented by bars in the graphs of Figure 3Figure 2).

Regardless of the change in the students' own social norms, most of their friends changed towards or remained in, the group of social norms against smoking. In fact, the average proportion of friends with injunctive and descriptive social norms against smoking at follow-up ranged between 55% and 94% (average 78%) for those who were friends at both time points, 64% and 100% (average 80%) for those who were friends only at follow-up, and 50% and 100% (average 100%) for those who were friends only at baseline. This suggested that students are more likely to nominate more friends who remained in the social norms group against smoking, regardless of their social norms change, intervention, and setting.

Finally, we measured the percentage of friends at baseline that had social norms towards and against smoking, for each possible scenario of social norms change. In respect of the students who maintained or changed their social norms to be favorable towards smoking, the average percentage of friends with norms that were favorable towards smoking at baseline was 22% and 21% respectively. In contrast, in respect of the students who maintained or changed their social norms to be against smoking, the percentage of friends with norms that were favorable towards smoking at baseline was 12% and 19%, respectively. We did not find differences by setting or intervention. This suggested that students who maintained or changed their social norms to be in favor of smoking were likely to have more friends in favor of smoking at baseline than the students who sustained or changed their perceived social norms to be against smoking, regardless of their setting and intervention.

Using MECHANISMS study data ²⁴ we identified three latent groups of participants classified according to descriptive and injunctive social norms favorable towards or against smoking. We analyzed how influence and homophily regarding social norms affected friendship networks among high school students in two countries' settings with different rates of adolescent smoking; namely, Northern Ireland and Colombia. To our knowledge, this is the first study to explore the influence of social network structures on experimentally derived social norms based on game theory experiments yielding measures less susceptible to social desirability bias ³².

We identified three latent groups associated with descriptive and injunctive social norms around smoking. The first group included students with descriptive social norms favorable towards smoking and injunctive social norms against smoking. The second group included students with descriptive and injunctive social norms against smoking. The third, and by far the smallest group, included students with injunctive social norms favorable towards smoking (irrespective of their perceived descriptive norms). These findings further illustrate the potential of LTA to identify latent groups of social norms for smoking in different contexts. This study is novel in that it uses LTA to identify groups related to smoking social norms in prevention interventions and not for already smoking population. Other studies have used LTA to identify groups related to smoking exposure and smoking status ³³, transitions through the stages of change for smoking cessation ³⁴, and studies related to smoking in adolescents and young adults to distinguish subgroups with different smoking prevalence ³⁵ and to characterize diverse groups according to tobacco and nicotine product use ³⁶.

We found that slightly older students were more likely to belong to a group with injunctive and descriptive social norms favorable towards smoking. This result is similar to previous studies that have shown that susceptibility towards smoking and smoking rates increase rapidly with age³⁷. Furthermore, we found that girls were more likely to be assigned to a group with descriptive social norms favorable towards smoking than boys. This may be because since the mid-1980 smoking has become more socially acceptable for females and adolescent girls are more exposed to tobacco products, while the smoking social norm has remained similar for men ^38,39. We found that self-efficacy, agreeableness, and conscientiousness were inversely associated with belonging to a group with social norms favorable toward smoking, while extraversion was positively associated. Our results are coherent with previous evidence showing that improvements in self-efficacy are related to reductions in smoking susceptibility ⁴⁰, and suggest that higher levels of self-efficacy can reduce the probability of belonging to a group with descriptive social norms favorable towards smoking. Our results were inconclusive regarding the relationship between the pro-sociality traits (need to belong, fear of negative evaluation, and pro-sociality) and belonging to any particular smoking social norms groups, even though other studies have shown a negative relationship between pro-sociality traits and smoking susceptibility ^41,42 and a negative relation between pro-social influence and age ⁴³.

After the ASSIST and Dead Cool school-based interventions, most students remained or transitioned to the group with social norms against smoking though approximately 5% of the students transitioned to a group with norms favorable towards smoking. We found differences in transitions between groups according to the smoking prevention intervention delivered in the schools. Regarding the influence process (i.e. change in the person’s behavior generated by her or his friendships) in the schools, we found that the ASSIST intervention seems to promote more transitions to groups with social norms against smoking than the Dead Cool intervention. This outcome reflects the intended underlying mechanism of the ASSIST intervention, which was intentionally designed to use the social network structure of peers to diffuse anti-smoking norms, while Dead Cool is based upon traditional classroom pedagogy. Our results support the idea that social norms are subject to a type of social contagion within social networks ⁴⁴. This finding illustrates the potential of social influence acting in friendship networks to contribute to smoking prevention, even in culturally different networks⁴⁵. However, further research is needed to understand the optimal network intervention strategy to influence social norms. In fact, there are many different types of network intervention approaches^{17,18,46−48} and ASSIST uses only one type of network targeting approach (i.e. nominated influential students from the school year) ²⁴, ⁴⁹. Also, we found that students who retain or change to social norms favorable towards smoking are likely to have more friends with similar views at baseline than the students with perceived norms against smoking. This result is observable regardless of their setting and intervention. This finding underlines the importance of network thresholds to behavior adoption and change ^46–48,50.

Regarding homophily (i.e. friendships between people who have attributes in common), friendship ties are formed and maintained in association with common and reciprocal friendships, but we did not find changes in friendship ties related to smoking norms groups. This result is consistent with the discussion by Valente and Pitts ⁵⁰, where natural network evolution is purported to connect friends in common by unobservable variables. The authors named this phenomenon 'latent homophily', and it describes how students may be connected for unobserved reasons in a study (e.g. behaviors or activities outside the school, like physical activity, specific sports, or participating in music groups). This result suggests that belonging to the identified social norm groups is not significant for homophily among students, however, other characteristics incentivize homophily on school students like having reciprocal relationships.

Our findings should be interpreted considering the following limitations. First, 1407 students were analyzed in the network, 1394 at baseline and 1370 at follow-up, but only 1018 students had complete data on social norms and sociodemographic attributes for both time points. However, the analyzed data corresponds to 72% of the students in the networks and this subgraph preserves the same in-degree distribution as that of the entire network, according to a 5% significance t-test conducted among both degree distributions. This means that the analyzed subsample preserves the original network structure. Despite the reduction in the sample, the sub-sample remains diverse in terms of sociodemographic attributes, thus our results permit us to make robust inferences about differences in transitions based on age, gender, setting, and intervention. Second, there is no control group to assess the effect of the interventions in terms of changes in the friendship network structure and changes in the social norms – the study was not designed to test intervention effectiveness, but to provide a detailed exploration of mechanistic pathways. Nevertheless, we can compare the effect of peer influence and homophily among students offered the ASSIST and Dead Cool interventions according to their different designs. This is a relevant outcome for the MECHANISMS study, which was not designed to offer head-to-head comparisons of the effectiveness of the interventions, which have both been previously tested, but was designed to compare and contrast their intervention mechanisms ²⁴. Third, even if we were able to measure some changes in social norms and the network structures during our short study period, a longer period of implementation and observation of the programs might be needed to observe prominent changes attributable to peer influence and homophily effects, and social norms. Finally, social desirability bias can be affected by personal ⁵¹ and social network characteristics ⁵² and can affect the way that norms are reported ^51,52. This in turn could obscure the identification of social network effects on social influence and even the nature of the social influence effects that are occurring. This study is unique in distinguishing and accounting for both self-report social norms and experimentally derived incentivized social norms. Previously, Murray et al ¹⁶ showed evidence for the construct validity of eliciting adolescent smoking norms through self-report and experimentally derived incentivized methods. Furthermore, it is important to identify social norms separately from behavior in a longitudinal analysis that can distinguish the determinants of change in each ^16,27.

In conclusion, the ASSIST intervention, which is designed to leverage peer influence in school friendship networks, can help prevent adolescent smoking uptake by increasing the proportion of students that change to social norms against smoking. This effect was higher for students receiving ASSIST than for the Dead Cool intervention, which does not purposefully leverage friendship networks or personal attributes. Intervention design features that are important for network-based interventions are the threshold proportion of friends at baseline with social norms favorable towards smoking. This suggests the presence of an influence process to change social norms, determined by the type of intervention and friendship threshold. Finally, we did not find homophily driven by social norms group, however, there are social network measures associated with homophily like transitivity and reciprocity. Deciding which students should be peer supporters, which have been further investigated in a series of simulation studies^48–50, concerning the students' attributes and diverse network measures may help reach and impact the students with social norms more favorable toward smoking and potentially increase the effectiveness of the program.

4.1 Study design and participants

We conducted behavioral economic experiments on two data collection time-points (baseline and follow-up) in six schools in NI and six in BOG. For the 12 schools, we delivered two different smoking prevention interventions (ASSIST and Dead Cool). Schools were randomly assigned to receive the ASSIST intervention (three in NI and three in BOG) or the Dead Cool intervention (three in NI and three in BOG). A total of 1444 students participated in the study, including n=796 who received the ASSIST intervention (n=423 in NI; n=373 in BOG), and n=648 who received the Dead Cool intervention (n=295 in NI; n=353 in BOG). All students in a whole school year group (target age 12-15 years) were invited to participate (Year 9 in NI and Year 7 in BOG). In total, 91.7% of those eligible participated in the study (94.3% in NI and 87.9% in BOG). The research protocol of this study was approved by the Queen’s University Belfast, School of Medicine, Dentistry and Biomedical Sciences ethics committee (reference number 18.43; v3 Sept 2018), and by the Ethics committee of the Universidad de Los Andes, Bogotá (937 - July 30, 2018). Written informed consent to participate in this study was provided by the participant's legal guardian/next of kin. All research was performed in accordance with relevant guidelines/regulations of Northern Ireland and Colombia.

In the ASSIST intervention, all students nominated up to 15 classmates across the 3 questions whom they view as influential (who do you respect / who are good leaders / who do you look up to in your year at your school?). The top 18% of nominated students were invited to a two-day training course where they were trained to be peer supporters and were asked to intervene informally in everyday conversations to encourage their peers to not smoke ²⁵. For the Dead Cool intervention, teachers were trained to deliver eight weekly lesson plans in their class groups. The lessons examined the influences on smoking behavior from friends, family members, and the media ²⁶. Since both interventions were developed in the UK, extensive cultural adaptation was undertaken to implement the interventions in BOG ⁵³.

Data were collected before and after the delivery of the interventions using two instruments (more detailed information can be obtained in the published protocol paper ²⁴). First, a self-administered survey was used to collect data including sociodemographic information, self-efficacy, self-report social norms, and pro-sociality. The survey also included a social network questionnaire asking participants about peer relationships in the same year group, including closest school friends, peers with whom they would talk about something that was upsetting you, peers with whom they spend time outside of school, and influential peers ²⁴. For this analysis, the closest school friends’ relationship was used (“Please name up to ten of your closest friends in your school year”).

Second, as an experimentally derived measure of social norms, the incentivized experiments questionnaire (based on Game Theory) was used to collect a measure of general norms sensitivity, injunctive norms of altruism, injunctive norms about smoking behavior, descriptive norms about smoking behavior, and willingness to pay to support anti-smoking norms [13], ²⁴, ⁴².

Qualtrics software was used to collect all data (Qualtrics, Provo, Utah, USA Version Jan. 2019). All the instruments were translated and back-translated by bilingual speakers/translators. The study was approved by the School of Medicine, Dentistry, and Biomedical Sciences Ethics Committee at Queen's University Belfast on September 21 of 2018 (ref: 18:43) and from the Research Ethics Committee at Universidad de los Andes on July 30 of 2018 (ref 937/2018).

4.2 Students attributes

Sociodemographic variables

We included the following demographic variables: age, sex, home composition, and ethnic status. Details of the categories for each variable can be found in the SI Appendix.

Psychosocial characteristics

We obtained self-report outcomes, including self-efficacy, pro-sociality, and personality traits, from the students in the baseline self-administered survey. Self-efficacy was assessed on a single scale. Pro-sociality was assessed with three variables: Need to Belong Scale, Fear of Negative Evaluation Scale, and Pro-Social Behavior Scale. The personality traits were assessed with the five subscales of the 'Big Five' personality questionnaire: Openness, Extraversion, Agreeableness, Conscientiousness, and Emotional Stability. Further details about the survey and outcomes can be found in the MECHANISMS study protocol ²⁴.

Intervention

For each student, a categorical variable indicating the intervention in which they participated (ASSIST or Dead Cool) was assigned, and a binary variable representing if the student was selected as a peer supporter in the ASSIST intervention was assigned.

Social norms measures

Both injunctive and descriptive smoking social norms were measured with a self-report survey and an incentivized experiment (SI Appendix, section 1.2) with a total of 25 variables ²⁴ (SI Appendix, section 1.1). The self-report survey collected information about family injunctive norms, peer injunctive norms, family descriptive norms, and peer descriptive norms ⁵⁴. Self-report injunctive norms were measured with a seven-point Liker scale, assessing individuals' beliefs about whether important others (e.g. parents, family, and friends) would approve of their smoking. The self-report descriptive norm was collected using a six-point Liker scale that assesses the adolescents' perceived smoking behavior of important others (e.g. parents, family, and friends).

The incentivized experiment used coordination games to elicit injunctive and descriptive social norms for smoking and vaping behaviors. In this experiment, participants were provided with financial incentives to match their answers to other participants' in their school year group ²⁷. The incentivized descriptive norms items measured beliefs about the proportion of the school year group who would be accepting of a close friend smoking or vaping. The incentivized injunctive norms assessed beliefs about how most others in the school year group would rate the social appropriateness of eight smoking-related scenarios.

4.3 Statistical Analysis

Uncovering the latent groups related to social norms

To reduce the dimensionality of the social norms variables, we conducted two confirmatory factor analysis (CFA) models, one to obtain a latent injunctive social norm variable, and the other to obtain a latent descriptive social norm variable (SI Appendix, section 1.2). The input variables for both CFAs were the self-report survey indicators and the incentivized experiment indicators standardized by the maximum value and changing the directionality of the metric to align them ¹⁶.

First, we conducted an initial exploration of how students could be clustered with a CFSC. Second, to understand how the latent groups changed over time, we conducted an LTA. The LTA uncovers the patterns that characterize the clusters of students according to their descriptive and injunctive social norms and estimates how the students transition across the groups over time. The CFSC is a solution to the multi-feature selection problem in clustering. We addressed the special characteristics of each sample in distinguishing a cluster rather than focusing on selecting a subset of samples. To do this, we quantitatively measured the clustering relevance of each sample to a cluster, which we refer to as component-based feature saliency. To achieve this, we assumed that the probability distribution of the samples can be modeled as a Gaussian mixture model (GMM). To estimate the feature saliency we used mixture models, and to optimize the clustering results we used Bayesian parameter estimation with Markov Chain Monte Carlo (MCMC) sampling for those cases where no analytical solution was possible ²⁸. Furthermore, the LTA is a longitudinal extension of Latent Class Analysis (LCA), a multivariate approach that allows the uncovering of hidden grouping variables. This methodology has been extensively studied due to its importance in identifying latent status membership probabilities, and transition probabilities to capture transitions between latent classes over time ⁵⁵ (SI Appendix).

Students from both settings with no missing data on the descriptive and injunctive norms items were included (n=1018). When the LTA modeling exhibits substantive separation in the group membership probabilities of each student, covariates do not have to be included in the estimation of the latent groups. The estimation of the latent groups and the LTA was performed using the MPlus program ⁵⁶. To estimate the association of the covariates (sociodemographic and psychosocial traits) with the latent groups at baseline, a multinomial logistic regression was performed using the latent groups from the LTA as the dependent variable and covariates as group predictors⁵⁷. The estimations were performed using the 'nnet' package in R ⁵⁸.

Assessing the change in network structure over time

Next, we evaluated if there were structural changes over time in the network of friends, undertaking a temporal analysis. For this, we calculated the Jaccard index J_i ⁵⁹ that measures the amount of change in the ties of the network over time using the following equation.

Where A_(1,1) represents the number of ties that existed before and after the intervention, A_(0,1) represents the number of ties that did not exist before the intervention but existed after the intervention, and A_(1,0) represents the number of ties that existed before the intervention, but ceased to exist after the intervention. According to the literature, to conclude that there were significant structural changes in the network over time, values between 0.3 and 0.6 should be obtained ⁵⁹.

Assessing the relationship between the network structure and the social norms group

To determine the presence of homophily or influence processes, we examined the association between the relationship changes in the structure of the friendship network with the changes in friends’ social norms. Homophily refers to the linkage of students by attributes in common. Influences refer to imitating behaviors of students with which the student is connected. For this, we conducted STERGMs and descriptive analysis.

A STERGM allows us to conclude if there is a greater likelihood of preserving or dissolving ties than that which can be attributed to chance ^31,60. For each network, a STERGM was conducted, including reciprocity and transitivity as network structural variables and the students’ social norms groups as individual attributes. STERGM simulations were conducted using the 'stergm' package in R ³¹ (SI Appendix).

The descriptive analysis allowed us identifying if when a student changed their social norms toward or against smoking (or if they stayed in the same group over time) this corresponded to similar changes in the social norms of their friends. Also, we analyzed the friendship threshold according to the proportion of friends (out-degree, i.e. nominations made) that have social norms favorable towards smoking. The analysis was conducted using only the students with complete social norms data at baseline and follow-up.

Acknowledgments and funding sources

Funding: Medical Research Council Population and Systems Medicine Board (reference number MR/R011176/1).

Author contributions

F.M., O.L.S, F.K., and R.F.H. conceived the study; F.M., M.B, A.U., S.S., H.Z., F.K., and R.F.H. designed the work; Analyses were performed by M.B., A.U., S.S., C.C., L.T. and J.L.; Interpretation of the data was performed by F.M., M.B., A.U., S.S., C.C., L.T. J.L., H.Z., F.K., J.M., O.L.S., and R.F.H; all authors wrote the manuscript and revised it

Data availability

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Code availability

The code used to generate the results of the CFA and STERGM reported in the current study is available for replication at https://github.com/mmblanco/Social-network-and-social-norms. The code used to generate the results of the CFSC has been cited and is available at https://github.com/XHongs/Component-based-Feature-Saliency-for-Clustering. The code used to generate the results of the LTA (MPlus software) is available from the corresponding author on reasonable request.

Competing interests

The author(s) declare no competing interests.

World Health Organization. WHO global report on trends in prevalence of tobacco use 2000–2025. World Health Organisation (2019).
Bonnie, R. J., Stratton, K. & Kwan, L. Y. Public Health Implications of Raising the Minimum Age for Purchasing Tobacco Products. (2015).
Chen, Y. et al. Adolescents’ behavioral and neural responses to e-cigarette advertising. Addict. Biol. 23, 761–771 (2018).
Abroms, L., Simons-Morton, B., Haynie, D. L. & Chen, R. Psychosocial predictors of smoking trajectories during middle and high school. Addiction 100, 852–861 (2005).
Defoe, I. N., Dubas, J. S., Somerville, L. H., Lugtig, P. & van Aken, M. A. G. The unique roles of intrapersonal and social factors in adolescent smoking development. Dev. Psychol. 52, 2044–2056 (2016).
Vitória, P. D., Salgueiro, M. F., Silva, S. A. & De Vries, H. Social influence, intention to smoke, and adolescent smoking behaviour longitudinal relations. Br. J. Health Psychol. 16, 779–798 (2011).
Reducing Adolescent Risk: Toward an Integrated Approach. Reducing Adolescent Risk: Toward an Integrated Approach (SAGE Publications, Inc., 2003). doi:10.4135/9781452233611.
Michie, S., Atkins, L. & West, R. The behaviour change wheel: a guide to designing interventions LK - https://univdelosandes.on.worldcat.org/oclc/984816374. TA - TT - (Silverback Publishing, 2014).
Huriah, T. & Dwi Lestari, V. School-based Smoking Prevention in Adolescents in Developing Countries: A Literature Review. Open Access Maced. J. Med. Sci. 8, 84–89 (2020).
Stok, M. F. & de Ridder, D. T. D. The focus theory of normative conduct. in Social Psychology in Action: Evidence-Based Interventions from Theory to Practice (eds. Scheepers, D. & Ellemers, N.) 95–110 (2019). doi:10.1007/978-3-030-13788-5_9.
Cialdini, R. B. & Goldstein, N. J. Social Influence: Compliance and Conformity. Annu. Rev. Psychol. 55, 591–621 (2004).
Mackie, G., Moneti, F., Shakya, H. & Denny, E. What are social norms? How are they measured? 1–100 (2015).
Cislaghi, B. & Heise, L. Four avenues of normative influence: A research agenda for health promotion in low and mid-income countries. Heal. Psychol. 37, 562–573 (2018).
Nyborg, K. et al. Social norms as solutions. Science (80-.). 354, 42–43 (2016).
Fowler, J. H. & Christakis, N. A. Cooperative behavior cascades in human social networks. Proc. Natl. Acad. Sci. U. S. A. 107, 5334–5338 (2010).
Murray, J. M. et al. Confirmatory factor analysis comparing incentivized experiments with self-report methods to elicit adolescent smoking and vaping social norms. Sci. Rep. 10, 1–17 (2020).
Valente, T. W. Network Interventions. Science (80-.). 337, 49 LP – 53 (2012).
Hunter, R. F. et al. Social network interventions for health behaviours and outcomes: A systematic review and meta-analysis. PLOS Med. 16, e1002890 (2019).
Montgomery, S. C. et al. Peer social network processes and adolescent health behaviors: A systematic review. Prev. Med. (Baltim). 130, 105900 (2020).
McPherson, M., Smith-Lovin, L. & Cook, J. M. Birds of a Feather: Homophily in Social Networks. Annu. Rev. Sociol. 27, 415–444 (2001).
La Fond, T. & Neville, J. Randomization Tests for Distinguishing Social Influence and Homophily Effects. in Proceedings of the 19th International Conference on World Wide Web 601–610 (Association for Computing Machinery, 2010). doi:10.1145/1772690.1772752.
Valente, T. W., Hoffman, B. R., Ritt-Olson, A., Lichtman, K. & Johnson, C. A. Effects of a social-network method for group assignment strategies on peer-led tobacco prevention programs in schools. Am. J. Public Health 93, 1837–1843 (2003).
Steglich, C., Sinclair, P., Holliday, J. & Moore, L. Actor-based analysis of peer influence in A Stop Smoking In Schools Trial (ASSIST). Soc. Networks 34, 359–369 (2012).
Hunter, R. F. et al. MECHANISMS Study: using Game Theory to assess the effects of social norms and social networks on adolescent smoking in schools—study protocol. Front. Public Heal. 8, 377 (2020).
Campbell, R. et al. An informal school-based peer-led intervention for smoking prevention in adolescence (ASSIST): a cluster randomised trial. Lancet 371, 1595–1602 (2008).
Thurston, A. et al. A randomized controlled efficacy trial of a smoking prevention programme with Grade 8 students in high schools. Int. J. Educ. Res. 93, 23–32 (2019).
Krupka, E. L. & Weber, R. A. Identifying social norms using coordination games: Why does dictator game sharing vary? J. Eur. Econ. Assoc. 11, 495–524 (2013).
Hong, X. et al. Component-Based Feature Saliency for Clustering. IEEE Trans. Knowl. Data Eng. 33, 882–896 (2021).
Muthén, L. K. & Muthén, B. O. Mplus: statistical analysis with latent variables : user’s guide. 7th, 594 p. (2012).
Bartolucci, F., Farcomeni, A. & Pennoni, F. Latent Markov models: A review of a general framework for the analysis of longitudinal data with covariates. Test vol. 23 (2014).
Krivitsky, P. N. & Goodreau, S. M. STERGM - Separable Temporal ERGMs for modeling discrete relational dynamics with statnet. Development 1–31 (2016).
Christakis, N. A. & Fowler, J. H. The collective dynamics of smoking in a large social network. N. Engl. J. Med. 358, 2249–2258 (2008).
Bray, B. C., Smith, R. A., Piper, M. E., Roberts, L. J. & Baker, T. B. Transitions in Smokers’ Social Networks After Quit Attempts: A Latent Transition Analysis. Nicotine Tob. Res. 18, 2243–2251 (2016).
Martin, R. A., Velicer, W. F. & Fava, J. L. Latent transition analysis to the stages of change for smoking cessation. Addict. Behav. 21, 67–80 (1996).
Mohammadpoorasl, A., Nedjat, S., Fakhari, A., Yazdani, K. & Fotouhi, A. Predictors of transition in smoking stages in Iranian adolescents: latent transition analysis.
Clendennen, S. L., Loukas, A., Creamer, M. R., Pasch, K. E. & Perry, C. L. Longitudinal Patterns of Multiple Tobacco and Nicotine Product Use Among Texas College Students: a Latent Transition Analysis. Prev. Sci. 20, 1031–1042 (2019).
Breslau, N. & Peterson, E. L. Smoking cessation in young adults: age at initiation of cigarette smoking and other suspected influences. Am. J. Public Health 86, 214–220 (1996).
Women and Smoking: A Report of the Surgeon General. MMWR Morb. Mortal. Wkly. Rep. 51, 1 (2002).
Mackay, J. & Amos, A. Women and tobacco. Respirology 8, 123–130 (2003).
Santano-Mogena, E., Franco-Antonio, C., Chimento-Díaz, S., Rico-Martín, S. & Cordovilla-Guardia, S. Factors associated with smoking susceptibility among high school students in western Spain. Sci. Rep. 11, 1988 (2021).
Mat Hassan, N., Abdul Aziz, A., Husain, R., Daud, N. & Juhari, S. N. Association of prosocial behavior with ever smoking and alcohol drinking among school-going adolescents. Heliyon 6, e04530 (2020).
Kimbrough, E. O. & Vostroknutov, A. Norms Make Preferences Social. J. Eur. Econ. Assoc. 14, 608–638 (2016).
Ahmed, S. et al. Susceptibility to prosocial and antisocial influence in adolescence. J. Adolesc. 84, 56–68 (2020).
Davis, C. A., Heiman, J. R. & Menczer, F. A role for network science in social norms intervention. Procedia Comput. Sci. 51, 2217–2226 (2015).
Go, M.-H., Tucker, J. S., Green, H. D., Pollard, M. & Kennedy, D. Social distance and homophily in adolescent smoking initiation. Drug Alcohol Depend. 124, 347–354 (2012).
Badham, J., Kee, F. & Hunter, R. F. Simulating network intervention strategies: Implications for adoption of behaviour. Netw. Sci. 6, 265–280 (2018).
Badham, J., Kee, F. & Hunter, R. F. Effectiveness variation in simulated school-based network interventions. Appl. Netw. Sci. 4, 70 (2019).
Badham, J., Kee, F. & Hunter, R. F. Network structure influence on simulated network interventions for behaviour change. Soc. Networks 64, 55–62 (2021).
Starkey, F. et al. Rationale, design and conduct of a comprehensive evaluation of a school-based peer-led anti-smoking intervention in the UK: the ASSIST cluster randomised trial [ISRCTN55572965]. BMC Public Health 5, 43 (2005).
Valente, T. W. & Pitts, S. R. An Appraisal of Social Network Theory and Analysis as Applied to Public Health: Challenges and Opportunities. Annu. Rev. Public Health 38, 103–118 (2017).
Hebert, J. R., Clemow, L., Pbert, L., Ockene, I. S. & Ockene, J. K. Social desirability bias in dietary self-report may compromise the validity of dietary intake measures. Int. J. Epidemiol. 24, 389–398 (1995).
Latkin, C. A., Edwards, C., Davey-Rothwell, M. A. & Tobin, K. E. The relationship between social desirability bias and self-reports of health, substance use, and social network factors among urban substance users in Baltimore, Maryland. Addict. Behav. 73, 133–136 (2017).
Sánchez-Franco, S. et al. Cultural adaptation of two school-based smoking prevention programs in Bogotá, Colombia. Transl. Behav. Med. (2021) doi:10.1093/tbm/ibab019.
Cremers, H. P., Mercken, L., Candel, M., De Vries, H. & Oenema, A. A web-based, computer-tailored smoking prevention program to prevent children from starting to smoke after transferring to secondary school: Randomized controlled trial. J. Med. Internet Res. 17, (2015).
Lanza, S. T., Patrick, M. E. & Maggs, J. L. Latent Transition Analysis: Benefits of a Latent Variable Approach to Modeling Transitions in Substance Use. J. Drug Issues 40, 93–120 (2010).
Muthén, L. K. & Muthén, B. O. Mplus User’s Guide. Sixth Edition. (1998).
Asparouhov, T. & Muthén, B. Auxiliary Variables in Mixture Modeling: Three-Step Approaches Using Mplus. Struct. Equ. Model. A Multidiscip. J. 21, 329–341 (2014).
Venables, W. N. & Ripley, B. D. Modern Applied Statistics with S. (Springer, New York, 2002).
Snijders, T. A. B., Van De Bunt, G. G. & Steglich, C. E. G. Introduction to stochastic actor-based models for network dynamics. Soc. Networks 32, 44–60 (2010).
Guerra, A. M. et al. Effects of a Physical Activity Program Potentiated with ICTs on the Formation and Dissolution of Friendship Networks of Children in a Middle-Income Country. International Journal of Environmental Research and Public Health vol. 17 (2020).

Table 1. The proportion of participants aggregated for both settings belonging to each latent group by intervention.

Social norms latent group	ASSIST		Dead Cool
Social norms latent group	Baseline	Follow-up	Baseline	Follow-up
Group 1. Descriptive social norms favorable towards smoking	15%	13%	7%	8%
Group 2. Both social norms against smoking	82%	85%	91%	89%
Group 3. Injunctive social norms favorable towards smoking	3%	2%	2%	3%
Total	100%	100%	100%	100%

Table 2. Multinomial odds ratio estimates for the association between the sociodemographic and psychosocial traits variables with the latent group assignment at baseline

Variable type	Variable	Group 1	Group 3
-	Intercept (A)	9.61E-10***	4.34E-5***
Sociodemographic	Sex = Girl	1.70*	1.00
	Sex = Prefer not to say	0.55	9.12E-17
	Age = 12 years	1.74E10***	2.25E4***
	Age = 13 years	2.72E10***	2.43E4***
	Age = 14 years	3.42E10***	2.12E4***
	Age = 15 years or more	4.04E10***	2.19E4***
	Ethnic minority = No	1.10	2.06
	Home composition = both parents	1.00	0.81
	Home composition = other adults	1.23	2.31
	Setting = NI	1.88*	1.13
Intervention	Dead Cool	0.76	0.27*
Psychosocial traits	Self-efficacy	0.49***	1.27
	Need to Belong	1.11	0.54
	Fear of Negative Evaluation	0.96	1.52
	Prosocial Behavior	0.89	1.06
	Openness (Big Five’s subscale)	1.08	0.72
	Extraversion (Big Five’s subscale)	1.59*	1.18
	Agreeableness (Big Five’s subscale)	0.67	0.35*
	Conscientiousness (Big Five’s subscale)	0.59*	0.40
	Emotional stability (Big Five’s subscale)	0.78	0.79

Note: The reference categories for each variable are as follows: group 2 for latent groups, boy for sex, 11 years for age, yes for ethnic minority, “single parent” for home composition, BOG for setting, and ASSIST for intervention. *<0.05 **<0.01 ***<0.001,

Table 3. Friendship networks characteristics.

Schools	Time Point	Nodes	Ties	Average in degree	Density	Jaccard Index
NI-Dead Cool	Baseline	83	537	6.5	0.08	0.44
	Follow-up	82	524	6.4	0.08	0.44
	Baseline	111	851	7.7	0.07	0.46
	Follow-up	112	756	6.8	0.06	0.46
	Baseline	97	706	7.3	0.08	0.56
	Follow-up	97	787	8.1	0.08	0.56
NI-ASSIST	Baseline	119	849	7.1	0.06	0.40
	Follow-up	117	755	6.5	0.06	0.40
	Baseline	122	787	6.5	0.05	0.40
	Follow-up	121	740	6.1	0.05	0.40
	Baseline	177	1230	6.9	0.04	0.42
	Follow-up	176	1015	5.8	0.03	0.42
BOG-Dead Cool	Baseline	106	703	6.6	0.06	0.42
	Follow-up	103	651	6.3	0.06	0.42
	Baseline	99	743	7.5	0.08	0.41
	Follow-up	99	641	6.5	0.07	0.41
	Baseline	126	739	5.9	0.05	0.38
	Follow-up	121	590	4.9	0.04	0.38
BOG-ASSIST	Baseline	144	957	6.6	0.05	0.34
	Follow-up	141	796	5.6	0.04	0.34
	Baseline	126	828	6.6	0.05	0.33
	Follow-up	124	775	6.3	0.05	0.33
	Baseline	84	473	5.6	0.07	0.26
	Follow-up	77	374	4.9	0.06	0.26

Note: Nodes represent individual participants; Ties represent the friendship nomination between participants; average in degree represents the mean number of friends nominated per individual participant; density is the ratio of the number of ties observed to the number of possible ties for a given network. Jaccard Index measures the structural change between follow-up and baseline, thus there is only one coefficient per school over time.

Table 4. Coefficients and p-values resulting from the STERGM model for each network.

School		Formation			Dissolution
School		Reciprocity	Transitivity	Social norms group on baseline	Reciprocity	Transitivity	Social norms group on follow-up
NI- Dead Cool	Estimate	3.09	3.00	-0.15	1.01	0.59	-0.11
	p-Value	0.00***	0.00***	0.27	0.00***	0.00***	0.52
	Estimate	3.64	3.32	0.02	0.93	0.70	0.10
	p-Value	0.00***	0.00***	0.89	0.00***	0.00***	0.46
	Estimate	3.77	4.48	0.04	0.81	0.83	-0.08
	p-Value	0.00***	0.00***	0.73	0.00***	0.00***	0.66
NI- ASSIST	Estimate	3.43	2.54	0.19	1.13	0.72	0.27
	p-Value	0.00***	0.00***	0.10*	0.00***	0.00***	0.04**
	Estimate	2.77	3.84	0.00	0.83	0.78	0.02
	p-Value	0.00***	0.00***	1.00	0.00***	0.00***	0.90
	Estimate	3.38	3.74	0.08	0.66	0.79	0.21
	p-Value	0.00***	0.00***	0.46	0.00***	0.00***	0.07*
BOG- Dead Cool	Estimate	3.88	5.08	-0.37	1.22	0.59	0.04
	p-Value	0.00***	0.00***	0.00***	0.00***	0.00***	0.79
	Estimate	3.06	5.25	0.41	1.67	0.52	-0.11
	p-Value	0.00***	0.00***	0.03**	0.00***	0.00***	0.51
	Estimate	3.66	4.63	0.14	1.18	0.68	N/A
	p-Value	0.0001***	0.0001***	0.386	0.0001***	0.0001***	N/A
BOG- ASSIST	Estimate	3.82	4.33	0.06	1.60	0.85	-0.13
	p-Value	0.00***	0.00***	0.65	0.00***	0.00***	0.33
	Estimate	3.73	4.92	0.07	1.60	0.69	N/A
	p-Value	0.0001***	0.0001***	0.51	0.0001***	0.0001	N/A
	Estimate	3.83	6.11	-0.13	1.06	0.85	0.11
	p-Value	0.00***	0.00***	0.42	0.00***	0.00***	0.54

Note: Transitivity represents the probability that friends of friends are also friends; Reciprocity represents the percentage of friendship nomination that is returned; N/A did not converge; p < 0.1; * p < 0.05; *** p < 0.00

No competing interests reported.

SNASocialNormsSupplementaryInformation.pdf

Exploring the mechanistic pathways of how social network influences social norms in adolescent smoking prevention interventions

Status:

Journal Publication

Version 1

Abstract

Figures

Significance Statement

1. Introduction

2. Results

Uncovering the latent groups related to social norms

Assessing the change in network structure over time

Assessing the relationship between the network structure and the social norms latent groups

3. Discussion

4. Methods

4.1 Study design and participants

4.2 Students attributes

Sociodemographic variables

Psychosocial characteristics

Intervention

Social norms measures

4.3 Statistical Analysis

Uncovering the latent groups related to social norms

Assessing the change in network structure over time

Assessing the relationship between the network structure and the social norms group

Declarations

Acknowledgments and funding sources

Author contributions

Data availability

Code availability

Competing interests

References

Tables

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1