Peer Selection and Influence for Marijuana Use in a Complete Network of First-Year College Students

Abstract

Background:

Marijuana is the most frequently used illicit drug among college students, and there is a need to understand the social processes that contribute to young adults’ marijuana use. Substance use behaviors tend to be more similar (homophilous) among individuals with social ties to one another. However, little is known about whether marijuana use homophily within young adult relationships is due to social selection (seeking out ties with similar marijuana use to one’s own) or social influence (adopting similar marijuana use behaviors as one’s ties), or both.

Methods:

Students (N = 1,489; 54.6% female; M_age = 18.6 years) at one University completed online surveys in their first three semesters of college. Surveys assessed participant characteristics, marijuana use, and ties to up to 10 other students in the whole (sociocentric) network of first-year college students. Stochastic-actor oriented models (SAOMs) were used to analyze the co-evolution of marijuana use and social ties over time.

Results:

Participants were more likely to select peers with similar past 30-day marijuana use as themselves. Concurrently, students’ past 30-day marijuana use became more similar to their peers’ use over time.

Discussion:

Evidence for selection and influence effects were highly significant after controlling for network structure and other covariates indicating these processes may work in tandem to increase marijuana use homophily over the first year of college. This highlights the importance of relationships made early in the first-year of college, as these initial peer ties are likely to reinforce marijuana use behaviors that occur within these relationships.

1.1. Introduction

In the U.S., marijuana use and marijuana use disorders (MUDs) are most prevalent among young adults. Approximately 43% of U.S. college students used marijuana in the past year, and 26% used marijuana in the past month (Schulenberg et al., 2020). This is concerning given recent research suggesting that marijuana use is related to worse academic outcomes, higher dropout rates and graduation delay (Arria, Caldeira, Bugbee, Vincent, & O’Grady, 2015; Suerken et al., 2016). Given that many individuals begin using marijuana in college (Suerken et al., 2014), there is a continued need to uncover factors associated with marijuana use and its consequences.

1.1.1. Social Factors of College Student Marijuana Use

Research generally supports that friends’ marijuana use is a strong predictor of adolescent marijuana use (Ennett et al., 2006; Tucker, De La Haye, Kennedy, Green, & Pollard, 2014). Peer influences appear to be equally as relevant for emerging adults. For example, although use of marijuana and other substances by college students’ parents, siblings and friends has been associated with students’ use, friends’ use of marijuana was the strongest predictor of college students’ own use (Windle, Haardörfer, Lloyd, Foster, & Berg, 2017). Other studies using social network methodology have demonstrated similarities between college students’ marijuana use and their friends’ use in a college dormitory (Barnett et al., 2014).

This tendency for individuals with similar attributes to affiliate with one another (i.e., homophily) is a consistent finding in sociological research (Valente, 2010). Social network theory outlines two mechanistic processes through which homophily occurs: in the case of marijuana use, individuals who use marijuana may form relationships with others who use marijuana (i.e., social selection) and/or the marijuana use of an individual’s social ties may influence that individual’s marijuana use (i.e., social influence) (Kandel, 1978; McPherson, Smith-Lovin, & Cook, 2001). Although both processes may function concurrently, knowing which process has a stronger influence may aid in the understanding of how marijuana and other substance use behaviors develop or spread within a given group. Furthermore, understanding these processes may inform interventions; namely because social selection and social influence processes may require different intervention approaches.

Parsing the effects of selection and influence is methodologically challenging, however, given that relationships and behavior change over time. One method for examining the process of social influence and social selection concurrently is the stochastic actor-oriented model (SAOM). The foundational assumption of SAOMs is that actors within social networks regularly make small changes to relations and behavior to optimize patterns therein at any given point in time. Favorable states within social networks and behavior are inferred through observed changes made by individual actors across waves of longitudinal data collection. The end product of an SAOM is a set of probabilities that describe generalizable patterns in actors’ preferences for social tie formation (e.g., reciprocity, homophily) and/or level of engagement in behaviors of interest (Snijders, 1996; Snijders, Van de Bunt, & Steglich, 2010).

Research studying marijuana use using SAOM with adolescent samples has generally supported that adolescents select friends based on similarity in marijuana use (De La Haye, Green, Kennedy, Pollard, & Tucker, 2013; De La Haye, Green, Pollard, Kennedy, & Tucker, 2015). There is also good evidence for influence (Wang, Hipp, Butts, & Lakon, 2018), although in some cases evidence for influence was found only in subgroups, as De La Haye et al. (2013) found it in a subset of students at one of the two schools they studied, and Tucker et al. (2014) found that influence was stronger for reciprocated (mutual) relationships and among friends who were relatively popular.

The experience of first-year college students is distinct from adolescents; greater independence from parents, exploration of new behaviors and identities, and exposure to social norms, beliefs, and contexts may all contribute to increased substance use (Arnett, 2005; Schulenberg & Maggs, 2002; White et al., 2006). Despite the known importance of peer influences and social contexts, research on the social networks of college students and their influence on substance use is still minimal. Furthermore, the use of sociocentric methods—in which the majority of individuals within a defined community are surveyed and their ties to other individuals within that community are measured—is rare in college student substance use research (Rinker, Krieger, & Neighbors, 2016). Compared to a personal (egocentric) network approach, a whole, or sociocentric approach considers the overall structure and composition of a network and its consequences for individual behavior (Perry, Pescosolido, & Borgatti, 2018).

1.1.2. The Current Study

The current study addressed two primary research questions: 1) To what extent do first-year college students select peers with similar marijuana use behavior? and 2) To what extent do peers influence college students’ marijuana use? The primary behavior of interest was the frequency of marijuana use in the past 30 days. We hypothesized that participants would nominate peers who had similar past 30-day marijuana use as themselves, and participants’ marijuana use would become more similar to their peers’ use over time. We controlled for several demographic and individual difference variables, and included typical controls for network structure suggested in SAOM models (Ripley, Snijders, Boda, Voros, & Preciado, 2020; Snijders et al., 2010).

2.1. Material and Methods

2.1.1. Design of Parent Study

Data were from a controlled intervention trial with the purpose of determining whether behavior change following a brief alcohol intervention would diffuse through a first-year class network. The class was segmented into two groups (Brief Motivational Intervention; BMI and Natural History Control) according to dorm location, and 25% of heavy drinkers in the BMI group (n = 61) received the intervention (Barnett et al., 2019); receiving the intervention was controlled for in these analyses. All methods were approved by the University Institutional Review Board.

2.1.2. Participants

All first-year students enrolled in the fall of 2016 at a mid-sized, private university in the northeast were eligible to participate, with the exception of non-traditional age students and students not living on campus (n = 32).

2.1.3. Procedures

Using a roster provided by the university, eligible students were contacted with postcards and through email ((Barnett et al., 2019). The email contained a project description and a web link to the consent form. Students who chose not to consent could allow their name to remain on the nomination list or opt out of having their name on the list. Students who did not consent or opt out remained on the list for others to select. Students who were under the age of 18 provided an email or mailing address for a parent who was sent project information and a consent form link.

Surveys were administered six weeks into the semester in the fall (T1) and spring semester (T2) of the first year, and in the fall of the second year (T3). Surveys were administered using web-based software and were available for two weeks. Gift cards to an online vendor were provided as compensation for survey completion ($50, $55, and $60 for the three surveys, respectively), with a bonus of $20 if all three surveys were completed. Enrollment at T1 was 1,342; students who had not enrolled at T1 were allowed to enroll at T2 (n = 101) or T3 (n = 46), resulting in a total sample size of 1,489 (90% enrollment rate) providing one or more observations. Retention was high, with follow-up rates of 97.8% at T2 and 95.5% at T3. Following data cleaning (e.g., for implausible response patterns; (Barnett et al., 2019), the total number of participants available for analysis was 1,329 at T1, 1,402 at T2 and 1,389 at T3.

Chi-square and t-test analyses were conducted to examine differences between those who completed at least one survey (n = 1,489) and those who did not complete any surveys (n = 171) using roster information provided by the university. The following variables were examined: sex (male vs female), minor (vs adult) status on first day of survey, race (white vs not white), Hispanic (yes vs no), citizenship (Foreign citizen, permanent resident, and US citizen), financial aid status (yes vs no), athlete status (yes vs no), intervention vs control dorm, and residence in a substance free dorm at baseline (yes vs no). Those who participated in the study were less likely to be men, (χ²(2) = 5.65, p = .017). more likely to identify as Hispanic (χ²(2) = 6.18, p = .013), more likely to be on financial aid (χ²(2) = 6.18, p = .013), and less likely to be intercollegiate athletes (χ²(2) = 20.32, p < .001). There were no significant differences on the other variables.

2.1.4. Measures

2.1.4.1. Marijuana use.

At T1, lifetime marijuana use was assessed by asking “Have you ever used marijuana / cannabis / hash?” (yes/no). In subsequent surveys, this item assessed participants’ use since the last survey. Participants who endorsed lifetime/recent use were asked “In the past 30 days, on how many days have you used marijuana?” Per recommendations (Ripley et al., 2020), this item was reduced to five categories: 0 days, 1 day, 2–3 days, 4–8 days, and 9 or more days, with no recent use recoded to 0 days.

2.1.4.2. Peer nominations.

Modeled after the Important People Instrument (Longabaugh & Zywiak, 2002), participants were asked to identify individuals in the first-year class “who have been important to you in the past month.” Participants were presented with a pull-down menu containing all students in the class (except those who had opted out) that had an auto-complete function, allowing them to rapidly select their classmates (up to 10 nominations).

2.1.4.3. Control variables.

Birth sex, race (dichotomized as white/nonwhite), Hispanic/Latino ethnicity, and intercollegiate athlete status (“Are you a member of a varsity athletic team at [university]?”) were included as controls. Parent education was used as a proxy for socioeconomic status, with options from less than a high school diploma to doctorate degrees recoded into 0 (neither parent has bachelor’s degree), 1 (one parent has bachelor’s degree), or 2 (both parents have bachelor’s degrees) (Weitzman, 2000). First-generation college status was measured with the item “Do you identify as a first-generation student?” The university provided dorm/room location for T1 and T2 and participants self-reported at T3. Some floors in the residence halls were designated as substance free; the university provided the location of substance-free floors.

Average number of drinks per week was calculated by multiplying participants’ drinking frequency (“In the past 30 days on how many days did you have at least one drink of any alcoholic beverage?”) and number of drinks per drinking day (“In the past 30 days, on the days when you drank, how many drinks did you drink on average?”) Including alcohol use as a time-varying control allowed us to determine whether the relationships between peer selection and influence were a function of substance use in general (with alcohol being the most commonly used substance) and not specific to marijuana use.

Resistance to Peer Influence (RPI; Steinberg & Monahan, 2007) measures a person’s ability to resist peer pressure; greater susceptibility to peer influence is associated with greater risky behaviors (Allen, Porter, & McFarland, 2006; Santor, Messervey, & Kusumakar, 2000), so this individual difference could account for conformity to peer behavior. The RPI scale has 10 items; respondents indicate which of two different types of people they are more like, with higher scores reflecting greater RPI. This measure was administered at the first observation for all participants and showed good internal consistency (α = 0.73).

2.1.5. Data Analysis

2.1.5.1. Stochastic Actor-Oriented Model.

We make use of the capability within the SAOM framework to simultaneously assess the influence of demographic, behavioral variables and social ties over time, while also assessing the influence of network structure on marijuana use. Thus, some model parameters from the SAOM developed within this paper represent changes in the tendency to form a social tie (network/selection terms) while others represent changes in the tendency for marijuana use (behavior/influence terms).

In order to assess the evidence for our hypothesis that social network factors will influence marijuana use behavior, we included a behavior/influence term ‘Marijuana Use: Average Alter’, which estimates the effect of the average level of marijuana use among a given person’s friends upon that person’s own marijuana use. Conversely, in order to assess evidence for our hypothesis that the level of marijuana use of an individual will have an effect on the particular friendship choices they make, we included the network/selection term ‘Marijuana Use: Similarity’. This term represents the tendency for ties to form among individuals with similar levels of marijuana use. Positive and statistically significant coefficients on these terms indicate support for each of these hypotheses, respectively.

2.1.5.2. SAOM control variables.

Consistent with best practice recommendations for SAOMs (Ripley et al., 2020; Snijders et al., 2010), variables were included to assess and control for network selection effects, including terms for transitivity (3-cycles, transitive triplets, GWESP), reciprocity, indegree and outdegree popularity. Other controls assess the effects of demographic factors on tendencies to make friendship nominations (denoted in the supplemental materials by the pairing of a demographic variable together with the word ‘alter’), to receive friendship nominations (‘ego’), and for homophily. We similarly operationalized both the effect of lifetime/recent marijuana use and the effect of residence within a substance-free dorm, respectively, with ego, alter, and homophily terms. We also included a general term to account for the effect of spatial propinquity stemming from living in the same dorm building. Finally, we included marijuana use alter and ego terms within the selection model as controls necessary to interpret our key network/selection variable of interest, Marijuana Use: Similarity.

Control variables pertaining to the marijuana use behavior included generic effects on marijuana use from number of incoming ties, number of outgoing ties, all demographic variables named above, resistance to peer influence, alcohol consumption, lifetime/recent marijuana use, and residence in a substance-free dorm. While the network composition and behavior choices of individual actors within our study are assumed to change over time, we modeled the data under the assumption that effects governing such changes are approximately constant over the study period and do not vary systematically between data collection waves.

All statistical modeling was completed using the R statistical computing environment, together with the ‘data.table’, ‘network’, and ‘RSiena’ packages (Butts, 2015; Dowle & Srinivasan, 2019; Ripley et al., 2020). A glossary of relevant SAOM terms is included in supplemental materials.

3.1. Results

Participant characteristics are in Table 1. Descriptive statistics for marijuana use and social network variables are displayed in Table 2. At T1, just under half of participants (48.3%) reported lifetime use of marijuana. This increased to 58.5% at T2 and 67.0% at T3. Past 30-day use was lower but increased over time with 34.6% of students reporting past 30-day use at T1, and 39.1% and 41.7% at the T2 and T3 surveys.

Table 1.

Sample descriptive information (N = 1,489).

Variable	N (%) or M(SD)
Age at T1	18.6 (0.51)
Birth sex
Female	813 (54.6%)
Male	675 (45.4%)
Race
White	837 (57.1%)
Asian	341 (23.3%)
Black/African American	106 (7.2%)
American Indian/Alaskan Native	14 (1.0%)
Multi-racial	149 (10.2%)
Other	8 (0.5%)
Hispanic/Latino/a Ethnicity	219 (14.7%)
Intercollegiate athlete	224 (15.0%)
First generation college	243 (16.3%)
Parent education
Low	117 (8.2%)
Middle	196 (13.7%)
High	1117 (78.1%)
Substance-free floor first year	201 (13.9%)
Substance-free floor second year	40 (2.8%)
Drinks per week at T1a	4.8 (6.3)
Drinks per week at T2a	4.5 (6.0)
Drinks per week at T3a	4.4 (5.7)
Resistance to Peer Influence	2.9 (0.41)

Notes. Age at T1 was missing for participants enrolled at T2 and T3 but was collected at these later time points. Birth sex was missing for one participant; race was missing for 34 (2.3%); ethnicity was missing for one participant; socioeconomic status was missing for 59 (4.0%). Age, substance-free floor status, and drinks per week were time-varying for analyses. All other observations were measured once and used in all analyses. All tabled variables were included as control variables in analysis.

^aLog transformed for analyses.

Table 2.

Marijuana use and network descriptives over waves.

	Time 1 (N = 1,329)	Time 2 (N = 1,402)	Time 3 (N = 1,389)
Past 30-day number of marijuana use days
0	65.4%	60.9%	58.3%
1	8.6%	8.3%	10.3%
2 to 3	9.5%	11.4%	10.5%
4 to 8	8.7%	9.8%	10.3%
9+	7.8%	9.6%	10.5%
Social network variables
Total number of ties	8,144	6,246	5,508
Network Density	0.0004	0.0003	0.0003
Average Degree	5.9	4.4	3.9
Degree Centralization	0.0064	0.0049	0.0049
Jaccard Index	Time 1 to Time 2: 0.309		Time 2 to Time 3: 0.369

Note. Number of days used marijuana was missing for 13 participants at T1, 31 at T2, and 37 at T3.

The overall density of the network (the number of actual ties divided by all possible ties) at each survey time point was quite low (less than 1 percent). This is common among large networks, since it is exceedingly difficult for each individual to know a large proportion of other individuals in the network (Valente, 2010). Participants also tended to name fewer important people over time, decreasing from an average of 6 alters at T1, to 4 alters at T3. Network ties were sufficiently stable to detect network change from T1 to T2 and T2 to T3, as indicated by Jaccard indices of 0.30 or higher (Ripley et al., 2020).

3.1.1. Stochastic Actor-Oriented Model

Main SAOM results of interest are displayed in Table 3. Full model results, including parameter estimates for controls, are in supplemental materials. After controlling for the above described endogenous network and individual factors, SAOM results suggest a relatively strong effect of previous 30-day marijuana use on selection and maintenance of friendship ties. Odds of forming or maintaining a tie to another participant were 1.79 times greater (OR = 1.79, 95%CI [1.59–2.01], p < 0.001) for those with a similar level of reported marijuana use in the previous 30 days. Frequency of marijuana use in the past 30 days was not associated with a tendency to nominate more peers (Marijuana Use Ego) or to receive more nominations (Marijuana Use Alter).

Table 3.

SAOM selection and influence results.

	Parameter	Estimate (SE)	Odds Ratio (95% CI)	p-value
Effects on Friendship Tie Formation and Maintenance (Selection)	Marijuana Use Similarity	0.58 (0.06)	1.79 (1.59 – 2.01)	<0.001
	Marijuana Use Alter	0.03 (0.02)	1.03 (1 – 1.07)	0.064
	Marijuana Use Ego	−0.01 (0.02)	0.99 (0.96 – 1.02)	0.567
	Marijuana Lifetime at T1 Alter	0.04 (0.03)	1.04 (0.99 – 1.1)	0.148
	Marijuana Lifetime at T1 Ego	−0.03 (0.03)	0.97 (0.9 – 1.03)	0.297
	Marijuana Lifetime at T1 Homophily	0.04 (0.02)	1.04 (1 – 1.08)	0.050
Effects on Marijuana Use Behaviors (Influence)	Marijuana Use Average Alter	0.31 (0.05)	1.36 (1.24 – 1.50)	<0.001
	Marijuana Use Indegree	0.05 (0.02)	1.05 (1.01 – 1.08)	0.009
	Marijuana Use Outdegree	0 (0.01)	1.00 (0.97 – 1.03)	0.842
	Marijuana Use: Effect from Lifetime at T1 Marijuana Use	0.4 (0.08)	1.49 (1.28 – 1.73)	<0.001

Notes. Marijuana use reflects past 30-day use. SAOM results including control variables can be found in Tables 1 and 2 in Supplemental Materials. Terms are further explained in Supplemental Materials.

Results from the behavior/influence portion of the SAOM suggest a general linear decrease in students’ frequency of marijuana use over time (OR = 0.32, 95%CI [0.26–0.39], p < 0.001). This was accompanied by a significantly positive quadratic shape (OR = 1.23, 95%CI [1.19–1.27], p < 0.001), indicating that changes in students’ marijuana use trajectories become more pronounced over time (whether increasing or decreasing). A consistent association was found between the average level of past 30-day marijuana use reported among one’s peers, and one’s own subsequent reported level of marijuana use. The odds of participants changing their level of marijuana use in the direction of the average of their peers was 1.36 times greater (OR = 1.36, 95%CI [1.24 – 1.50], p < 0.001) than making no change or moving in the opposite direction.

4.1. Discussion

This is the first known study to use longitudinal social network analysis to investigate the dynamic co-evolution of the processes of social influence and social selection in an entire cohort of college students at one university and thus adds to the body of work investigating these processes in younger age groups. As hypothesized, we found evidence that both selection and influence mechanisms contributed to marijuana use homophily after controlling for network structure, individual factors, and other dyadic covariates. Students were more likely to select others with similar lifetime and past 30-day marijuana use behaviors and, after controlling for selection effects, students’ frequency of marijuana use became more similar to their peers’ use over time. Thus, both of our hypotheses were supported.

These results contrast with previous research among adolescents in a few ways. Using Add Health data collected in the mid-1990’s, De La Haye and colleagues’ (2013) found that adolescents selected friends with similar lifetime and past-month marijuana use in two different high schools, but friends’ use did not influence adolescents’ use in either school. Using the same sample, De La Haye and colleagues (2015) reported selection effects after controlling for multiple risk factors. We also found that the tendency to select other students with similar marijuana use persisted after controlling for alcohol use and individual characteristics but also found strong influence effects. Differences in findings with previous research could be explained by the age of the data in the prior research (collected in the 1990s, with many changes in risk perception and legalization since), or the age of the participants (with college students more likely to have experience with marijuana than high school students).

It is important to determine whether individual risk and protective factors specific to college students moderate selection and influence processes. For example, certain subgroups of college students (e.g., those living on campus) show a higher likelihood of initiating marijuana use during the first year of college (Suerken et al., 2014). Future research of this kind may also test network and dyadic relationship-related characteristics as moderators of selection and influence, that is, whether these processes differ depending on network centrality or relationship strength (e.g., perceived closeness, romantic versus friend relationship, etc.). Tucker et al. (2014) determined that adolescents were more likely to adopt the marijuana use behaviors of friends if the friendship ties were reciprocated (indicating a stronger relationship), and in one school site, adolescents were more likely to adopt the marijuana use behavior of popular friends. In the present study, since indegree centrality (i.e., popularity) was positively associated with marijuana use, it could be that ties to popular individuals would increase one’s own marijuana use.

Finally, future work applying SAOM to questions regarding marijuana use may benefit from exploring the possibility of changes in the values of model parameters across successive periods of network data. Any study being conducted in the midst of major changes to the regulation of marijuana use (as have been carried out recently in many US states) may expect to find notable changes in the dynamics surrounding influence and selection as they relate to marijuana use.

4.1.1. Strengths and Limitations

Strengths include the high enrollment (90% of the class with one or more observation) and retention (>95%), and inclusion of new participants at all time points. The primary limitation is that we did not analyze potentially influential relationships outside of the first-year class, and thus cannot draw conclusions about the influence of people outside of the class network. Participants’ self-report of marijuana use may be inaccurate. We did not assess marijuana use dependence or problems; it will be important to consider whether the social processes that lead to use result in academic or social problems. As shown in our data, participants nominated fewer network members over time. We cannot be certain whether participant networks are reducing in size, or if participants nominated fewer people to reduce the survey length. We asked participants to identify important peers; participants may not have named those who were most influential with regard to marijuana use. There a few demographic categories that were underrepresented, despite our high enrollment rate. Finally, the model did not consider change between specific waves of data; it may not have been accurate to assume there is a general process that occurs throughout the waves as a single and consistent process.

4.1.2. Implications

College is a period of considerable social network change and increase in the initiation of and maintenance in marijuana use, and as such, it is an optimal time for intervention. Reducing marijuana use during the first semester of college may reduce the likelihood of selecting into substance-using peer networks and subsequent escalation of use as a consequence of affiliating with such peers. In addition, environmental interventions may address the prevalence of marijuana use during college as the higher the prevalence at college, the more intense the source of influence effects. Valente (2012) provides an overview of different social network interventions that may be efficacious in reducing marijuana use risks; those that might be appropriate for early college include approaches that identify individuals who have prominence in the network to act as behavior change leaders, or that facilitate behavior change in clusters of individuals who show the target behavior. Few such interventions have been developed, and none have focused on the substance use of young adults (Hunter et al., 2019), so this area is ripe for future research development.