Searches

Searching the databases yielded 2886 de-duplicated records (Fig. 1). 2837 records were excluded based on title and abstract and the full texts of 49 potentially eligible studies were examined. Of these, 14 studies met all eligibility criteria. After conducting forward and backward citation searching of the 49 articles that went through to the full-text stage, one further eligible study was found. This culminated in 15 papers eligible for inclusion in the review. Searches were re-run in May 2022 and no new papers were identified. Researchers (HC, EW) reached > 90% agreement at all stages of screening and any uncertainties were resolved with a third researcher with expertise in social network analysis (MC).

Fig 1.

Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) Flow Chart of the Study Selection Process

Risk of Bias

Risk of bias was generally low, with 12 studies scoring low for risk of bias (80%), and three scoring medium for risk of bias (Table 1) (Baller and Richardson 2009; De Luca et al. 2012; Haynie et al. 2006). The same overall conclusion for risk of bias category (high, medium, low) was made for each study using two risk of bias tools (Knox et al., 2019; Sabot et al. 2017). However, the Sabot et al. (2017) tool was preferred as it captured more questions about the potential biases of network studies (Online Resource 4).

Table 1 Summary Characteristics of Included Studies Study Characteristics

Characteristics of each study are summarized in Tables 1 and 2, and Online Resource 5. Twelve studies were from the USA (73%) and three from China (Giletta et al. 2015; You et al. 2013, 2016). Data come from seven independent studies which meant that some samples overlapped: eight studies (53%) used samples from the National Longitudinal Study of Adolescent to Adult Health (which is referred to as ‘Add Health’ moving forward) (Harris et al. 2019), two studies used samples from a study of eight co-educational high schools in China (You et al. 2013, 2016), and five studies (33%) used samples from other data sources of adolescents in the USA and China. Thirteen (86%) of the 15 papers were analyses of longitudinal data with follow-up times ranging from six months to two years, only two were cross-sectional (De Luca et al. 2012; Wyman et al. 2019).

Table 2 Overview of the Peer-friendship Network Metrics Measured in Studies and Associations with Self-injurious Thoughts and Behaviors in Adolescence

Our original inclusion criteria stated ‘adolescents aged 11–18’ years of age. However, studies that used Add Health data analyzed adolescents from 7th -12th grade who were aged 11–19 years old at baseline, and 12–20 years old at their one-year follow up assessment. These studies were still eligible to be included because they analyzed the sociometric friendship networks of adolescents in a school-based setting. Overall, most studies (n = 9) had a sample of adolescents with an age range that spanned across the developmental period of adolescence (e.g., age ranges of 11–19 years of age at baseline) with mean ages of these samples at baseline ranging from 14.93 to 16.21. One study focused on a sample of younger adolescents only, with an age range of 11–14 years old at baseline (Prinstein et al. 2010). Three studies had samples of mid-to-late aged adolescents (age ranges between 14 and 19 years old at baseline) (Giletta et al. 2013; Wyman et al. 2019; Zimmerman et al. 2016). Two studies focused on the period of later adolescence only (i.e., samples of adolescents within the age range of 16–18 only at baseline) (Copeland et al. 2019; Giletta et al. 2015).

Sample sizes ranged from 348 to 13,482 adolescents. All studies were conducted in a school setting (a mix of middle schools and high schools) and used a range of sociometric nomination procedures (e.g., nominate a maximum of five friends, nominate an unlimited number of friends). Table 1 indicates diversity for some samples in terms of demographic characteristics like race/ethnicity and gender. Eleven studies had samples with a mix of ethnic groups, but predominantly White/Caucasian (e.g., White ethnicity in these studies ranged from 45 to 86%). Three studies had samples from high schools of Chinese adolescents only (Giletta et al. 2015; You et al. 2013, 2016). One study analyzed a girl-only Latina friendship network (De Luca et al. 2012). All other studies included both boys and girls with the percentage of girls ranging from 46 to 56%.

Studies examined a range of self-injurious thoughts and behaviors including suicide attempt, suicidal ideation, non-suicidal self-injury (NSSI), and self-cutting behavior. Moving forward, suicide attempts are referred to as ‘attempts,’ suicidal ideation is referred to as ‘ideation’, NSSI and self-cutting behavior(s) are referred to collectively as ‘self-harm’. In the results section only, self-injurious thoughts and behaviors are acronymized to ‘SITBs’ in some places. Six studies measured self-harm, eight studies measured ideation, and eight studies measured attempts.

Peer-network Metrics

Studies assessed multiple network metrics (Tables 2 and 3). To facilitate the synthesis, metrics were grouped in two ways: (1) peer-network structure (i.e., individual position, proximal-peer, network-wide metrics) (2) exposure to friends’ self-injurious thoughts and behaviors (including dynamic processes like socialization and/or selection). Definitions of structural network metrics are provided in Table 3.

Table 3 Structural Peer-network Metric Definitions Study Analyses

Studies used a range of analytical techniques (e.g., different forms of linear, logistic, probit, and lagged regression). One study used a Stochastic Actor-Oriented Model (SAOM), and the results of this study are presented under a separate heading within the ‘exposure to friends’ self-injurious thoughts and behaviors’ section because of the model’s unique ability to tease apart the dynamic processes of socialization and selection compared with other analytical models. Only the SAOM-based study was considered as measuring socialization and selection, all other ‘non-SAOM’ studies that measured exposure to friends’ self-injurious thoughts and behaviors were grouped into a separate category (even if they used the phrases socialization and selection in their paper). Statistical models included a range of covariates including sociodemographic characteristics, depressive symptoms, impulsivity, bullying, connection to the school, peer support, and family suicidality (Online Resource 5).

Eleven studies set out to test gender differences (two studies split their analyses by gender, nine studies assessed gender as an interaction or moderating term within their analyses). Although studies had samples of adolescents from across the range of adolescence, only one study that focused on early adolescence (grades 6–8, ages 11–14) assessed age as an interaction term in their analyses (Prinstein et al. 2010), all other studies either adjusted for age or did not adjust for age due to low variability across the sample (Wyman et al. 2019).

Despite diversity in some samples for demographic characteristics like ethnicity/race, only one study explicitly looked at ethnicity/race as a moderating term within their analyses (Xiao and Lindsey, 2021). One study included an effect for ethnicity in their SAOM model to look at the tendency for same-ethnicity youth to select same-ethnicity friends, but this was not assessed in relation to self-injurious thoughts and behaviors (Giletta et al. 2013). All other studies with diversity in terms of race/ethnicity adjusted for ethnicity in their analyses. Some studies also adjusted for same-sex attraction, but only one study specifically looked at sexual identity as a moderating term in their analyses (Xiao and Lindsey, 2021). One study measured network metrics within a full sample of adolescents and an ‘at-risk’ subset (i.e., where risk was a function of heavy drinking, fighting, rape victimization, same-sex attraction, and obesity) (Baller and Richardson 2009), and one study assessed if network metrics mediated the effect of residential mobility (i.e., moving within the last two years) on self-injurious thoughts and behaviors among adolescents (Haynie et al. 2006).

Effect sizes from the respective studies’ preferred statistical models are reported below, but details about what the models adjusted for are available in Online Resource 5. For studies that used linear regression, they reported standardized (β) and/or unstandardized beta coefficients (b), standard error (SE) and/or 95% confidence intervals (95%CI). For studies that used logistic regression, some reported only the standardized (β) or unstandardized (b) beta coefficients and SE, and others reported odds ratios (OR) and 95% CI’s. Statistics are reported as they were presented in the paper.

Peer-network StructureIndividual Position Metrics

Four studies measured a range of individual positions within a peer-friendship network and their association with self-injurious thoughts and behaviors (Bearman and Moody 2004; Copeland et al. 2019; Haynie et al. 2006; Wyman et al. 2019). Overall, there was evidence that higher sociality (i.e., how many friendship nominations one sends) was negatively associated with self-injurious thoughts and behaviors, but limited evidence or mixed findings for all other position metrics.

Bonacich Centrality

Two longitudinal studies measured Bonacich centrality (i.e., being popular among peers that are also popular across the network) and findings were mixed (Copeland et al. 2019; Haynie et al. 2006). Specifically, one study analyzed a sample of 11,160 adolescents from US high schools and found that Bonacich centrality was negatively associated with self-harm (β: -0.366, SE:0.12) in a simple model adjusted for depressive symptoms and sociodemographic characteristics (Copeland et al. 2019). However, the effect was reduced and no longer independently associated with self-harm in a more complex model with other network facets included (to better understand the simultaneity of network effects) (β: -0.268, SE:0.25). Conversely, a different study analyzed a sample of 9594 adolescents from Add Health assessing multiple network metrics (including Bonacich centrality) as potential mediators on the effect of being a residential mover and attempts among adolescents (Haynie et al. 2006). They found no evidence that Bonacich centrality was associated with attempts among girls (b: 0.111, SE: 0.174) or boys (b: -0.268, SE: 0.296).

Closeness and Betweenness Centrality

Only one study measured these metrics (Copeland et al. 2019). Longitudinally, they found that betweenness centrality (i.e., higher values indicate adolescents that bridge others in a network) was positively associated with self-harm among adolescents (β: 2.316, SE: 0.79) but no evidence that closeness centrality (a measure of how easily one can access others in a network) was associated with self-harm (β: -1.063, SE: 1.11).

Popularity (in-degree)

Two studies measured popularity (i.e., how many friendship nominations one receives) and findings were mixed (Copeland et al. 2019; Wyman et al. 2019). Cross-sectionally, one study analyzed the friendship networks of 10,291 adolescents from 38 US high-schools and found evidence of a negative association between popularity and attempts [vs no SITBs] (OR: 0.95, 95% CI: 0.92, 0.98), ideation [vs no SITBs] (OR: 0.95, 95% CI: 0.92, 0.97), and attempts [vs. ideation] (OR: 0.98, CI: 0.94, 1.02) in models adjusted for sex, age, and ethnicity (Wyman et al. 2019). Longitudinally, a different study found no evidence that popularity was associated with self-harm among adolescents (β: 0.047, SE: 0.06) (Copeland et al. 2019).

Sociality (out-degree)

Two studies measured this metric, and both found evidence that sociality (i.e., how many friendship nominations one sends) was associated with lower self-injurious thoughts and behaviors (Copeland et al. 2019; Wyman et al. 2019). Longitudinally, one study found that sociality was negatively associated with self-harm among adolescents (β: -0.167, SE: 0.05) and this effect was stronger among boys (Male X Sociality β: -0.269, SE: 0.11) (Copeland et al. 2019). Cross-sectionally, a different study found that sociality was negatively associated with attempts [vs no SITBs] (OR: 0.87, 95%CI: 0.85, 0.90), ideation [vs no SITBs] (OR: 0.93, 95% CI: 0.91, 0.96) and for ideation [vs attempts] (OR: 0.92, 95%CI: 0.88, 0.96) among adolescents in models adjusted for sex, age, and ethnicity (Wyman et al. 2019). The effect for sociality was stronger among girls for ideation [vs no SITBs] (OR: 0.89, 95%CI: 0.85, 0.92) compared with boys (OR: 0.97, 95%CI: 0.93, 1.01).

Isolation

There were mixed findings for isolation: three studies (two longitudinal, one-cross-sectional) measured this metric, two of which found evidence that being isolated (i.e., neither sending nor receiving any nominations) was associated with higher self-injurious thoughts and behavior (both using Add Health samples) but with gender differences (Bearman and Moody 2004; Haynie et al. 2006; Wyman et al. 2019). Longitudinally, one study analyzed the friendship networks of 13,465 adolescents from Add Health and found that isolation (vs. not) was positively associated with ideation (but not attempts) among girls (OR: 2.010, 95%CI: 1.073, 3.765) in models adjusted for sociodemographic characteristics (Bearman and Moody 2004). In boys, they found no evidence isolation was associated with ideation (OR: 0.665, 95%CI: 0.307, 1.445) or attempts (OR: 0.767, 95%CI: 0.159, 3.707). A different study analyzed a sample of 9594 adolescents from Add Health and found evidence that isolation (vs. not) was positively associated with attempts among girls (β: 0.700, SE: 0.345), but no evidence it was associated with attempts among boys (β: 0.458, SE:0.539) (Haynie et al. 2006). A cross-sectional study found no evidence of an association between isolation (vs. not) and attempts [vs no SITBs] (OR: 1.22, 95%CI: 0.74, 1.99), ideation [vs no SITBs] (OR: 1.25, 95%CI: 0.79, 1.97) or for attempts [vs. ideation] (OR: 1.64, 95%CI: 0.75, 3.61) among adolescents from a different sample of US High schools (Wyman et al. 2019).

Proximal-peer Metrics

Seven studies measured proximal-peer metrics (i.e., features of an adolescent’s immediate peer-group within the larger network) and their association with self-injurious thoughts and behaviors among adolescents (Baller and Richardson 2009; Bearman and Moody 2004; Copeland et al. 2019; De Luca et al. 2012; Giletta et al. 2015; Wyman et al. 2019; Xiao and Lindsey 2021). There was evidence that the intransitivity index was positively associated with self-injurious thoughts and behaviors, but limited evidence or mixed findings for all other proximal-peer metrics.

Intransitivity Index

Two longitudinal studies (using samples from Add Health) measured the intransitivity index (i.e., a measure of the proportion of an adolescent’s friends’ friends who were not also the adolescent’s friend) and both found that scoring higher on this index was positively associated with self-injurious thoughts and behaviors among adolescents (Baller and Richardson 2009; Bearman and Moody 2004). Specifically, one study found the index was positively associated with ideation among their whole sample of adolescents (n = 2084, β: 1.208, SE: 0.342) and among an at-risk subset (n = 1300, β: 1.246, SE: 0.621) (Baller and Richardson 2009). A different study found the index was positively associated with ideation (but not attempts) among girls (OR: 2.198, 95%CI: 1.221–3.956) but no evidence it was associated with ideation or attempts among boys (Bearman and Moody 2004).

Coreness

Only one study measured this metric (i.e., size of friendship group based on one’s own and immediate friendship nominations) (Wyman et al. 2019). Cross-sectionally, they found that coreness was negatively associated with attempts [vs no SITBs] (OR: 0.84, 95%CI: 0.80, 0.87), ideation [vs no SITBs] (OR: 0.90, 95%CI: 0.86, 0.93), and attempts [vs. ideation] (OR: 0.89, 95% CI: 0.84, 0.95) among adolescents in models adjusted for sex, age, and ethnicity. For ideation, the effect was stronger among girls (OR: 0.83, 95%CI: 0.79, 0.87) compared with boys (OR: 0.95, 95%CI: 0.89, 1.00).

Peer-network Density and Size

Only one study measured these metrics (i.e., measures of the number of connections within an adolescent’s immediate friend group) and associations with belonging to trajectory classes of ideation and attempts among a sample of 9421 adolescents from Add Health (Xiao and Lindsey 2021). Longitudinally, they found no evidence of an association between these metrics and belonging to trajectory classes of attempts. However, in supplementary moderator analyses looking at race/ethnicity and sexual identity, they found that sexual minority youth were more likely to be in a ‘high-decreasing’ ideation trajectory compared with a ‘low stable’ ideation trajectory when having a densely connected peer-network (OR 1.19, 95% CI 1.02–1.40). They found no strong evidence for ethnicity as a moderator of these effects.

Reciprocity

There were mixed findings for reciprocity and self-injurious thoughts and behaviors. Of the three studies that measured reciprocity (two longitudinal, one cross-sectional), two found no evidence of an association with self-injurious thoughts and behaviors (Copeland et al. 2019; De Luca et al. 2012) and one found evidence that a lack of reciprocal friends (vs. having friends with/without self-harm) was associated with being in a moderate compared with low self-harm trajectory (Giletta et al. 2015)Footnote 1. Specifically, one study analyzed the friendship networks of 565 adolescents from two schools in China. They assessed whether having no reciprocal friends compared with having reciprocal friends (with and without self-harm) was associated with belonging to latent trajectory groups of self-harm and/or ideation (low, moderate, high). After adjusting for gender and depressive symptoms they found that adolescents were less likely to be in the moderate trajectory group for self-harm (compared with low) if they had reciprocal friends who self-harm (OR: 0.46, 95%CI: 0.21, 1.00) or reciprocal friends who did not self-harm (OR: 0.35, 95%CI: 0.16, 0.76) compared with having no reciprocal friends at all. No friendship type effects were found for ideation trajectory groups (Giletta et al. 2015). Conversely, a different study measured the mere presence of reciprocity (i.e., the proportion of adolescent’s ties in which both parties nominated each other) and found no evidence of an association with self-harm among adolescents (β: 0.030, SE: 0.22) (Copeland et al. 2019). Cross-sectionally, another study analyzed a Latina only friendship network (n = 1618) from Add Health and found no evidence of an association between reciprocity and attempts or ideation (statistics not reported) (De Luca et al. 2012).

Network-wide Metrics

Network-wide refers to measuring the way an entire school functions (e.g., a climate or context related to self-injurious thoughts and behaviors) compared with structural positions or proximal-peer metrics. Two studies measured network-wide density (Bearman and Moody 2004; Wyman et al. 2019) and only one study measured a range of network-wide metrics and associations with self-injurious thoughts and behaviors (Wyman et al. 2019). There was more evidence that density (i.e., belonging to an overall dense school friendship network) was negatively associated with self-injurious thoughts and behaviors, but limited evidence for all other network-wide metrics. Overall, these two studies provide evidence that peers relate to self-injurious thoughts and behaviors not just through characteristics of direct friends or structural positions among direct friends, but the overall patterns of networks in the school relate to overall rates of self-injurious thoughts and behaviors.

Density

Two studies measured density (i.e., a measure of the ratio of actual friendship ties among adolescents in the network to all possible ties) and found that schools with higher density were associated with lower self-injurious thoughts and behaviors (Bearman and Moody 2004; Wyman et al. 2019). Longitudinally, one study found evidence that for girls, density was negatively associated with ideation (OR: 0.333, 95%CI: 0.142, 0.783) but not for attempts. For boys, there was no evidence density was associated with ideation (OR: 1.061, 95%CI: 0.375, 2.999), but they found evidence it was negatively associated with attempts [among boys with past year ideation] (OR: 0.049, 95%CI: 0.005, 0.521) (Bearman and Moody 2004). Cross-sectionally, one study found that density was negatively associated with rates of ideation [vs no SITBs] (β: -1.61, 95%CI: -2.61, -0.61) in a simple model adjusted for network size, sex, and ethnicity, but this effect was reduced and no longer independently associated in a more complex model with other similar network facets included (β: -0.82, 95%CI: -6.06, 4.41). They found no evidence of an association with rates of attempts in any models (Wyman et al.