Previous research has suggested that adolescent peers influence behavior and provide social support during a critical developmental period, but few studies have addressed the antecedents of adolescent social networks. Research on the collateral consequences of incarceration has explored the implications of parental incarceration for children’s behavioral problems, academic achievement, health, and housing stability, but not their social networks. Using network data from the National Longitudinal Study of Adolescent Health, I find that adolescents with recently incarcerated fathers are in socially marginal positions in their schools and befriend more-marginal peers than other adolescents: their friends are less advantaged, less academically successful, and more delinquent than other adolescents’ friends. Differences in network outcomes are robust to a variety of specifications and are consistent across race and gender subgroups. This study advances the social networks literature by exploring how familial characteristics can shape adolescent social networks and contributes to the collateral consequences of incarceration literature by using network analysis to consider how mass incarceration may promote intergenerational social marginalization.
Social networks are an important context through which information, social norms, and social support flow, particularly during adolescence (Giordano 2003). Research on adolescent social networks has primarily focused on how behaviors and values diffuse through them (e.g., Bearman et al. 2004; Ennett and Bauman 1993; Haynie and Osgood 2005) and how network characteristics shape adolescent outcomes, such as delinquency and educational attainment (e.g., Calvó-Armengol et al. 2009; Ennett et al. 2006; Haynie 2001). Few studies have considered the antecedents of adolescent networks (but see Goodreau et al. 2009; Haas et al. 2010; Harding 2008) despite evidence that race and class pattern adults’ social networks and access to social capital (Desmond 2012; Lin 2000; Marsden 1987; McPherson et al. 2006; Schafer and Vargas 2016).
In light of historically high U.S. incarceration rates that are disproportionately concentrated in minority and impoverished communities (National Research Council 2014; Western and Pettit 2010), this article explores how an increasingly common form of familial disadvantage—parental incarceration—affects adolescent social networks. Although the consequences of incarceration for children and families have been studied extensively (Foster and Hagan 2015b; Johnson and Easterling 2012), prior research has rarely considered that parental incarceration may shape adolescent social networks, potentially fostering disadvantageous social ties early in life.
Several qualitative studies have proposed that children experience social isolation and rejection by peers following parental incarceration (Bernstein 2007; Nesmith and Ruhland 2008; Sack et al. 1976), but these studies have used small samples and cannot account for confounding factors, such as behavior. I use the National Longitudinal Study of Adolescent Health (Add Health) to examine how social network location and friend characteristics differ by recent paternal incarceration. Across a variety of models that adjust for behavior, school fixed effects, and different comparison groups, I find strong and consistent evidence that teenagers with recently incarcerated fathers are embedded in friend groups that are more locationally and compositionally marginal: they are less connected in their schools than other students, and their friends are less advantaged, less academically successful, and more delinquent than other students’ friends. These findings suggest that household-level hardships, such as parental incarceration, can influence adolescent social network characteristics, which prior research has often treated as fixed (Bearman and Moody 2004; Calvó-Armengol et al. 2009). Moreover, like earlier research on the collateral consequences of incarceration, these findings point to the intergenerational inequalities associated with incarceration.
Mass Incarceration and Collateral Consequences
Historically high U.S. imprisonment rates have exposed large numbers of low-income, largely black and Latino, children to parental incarceration. On any given day in 2012, approximately 2.6 million American children had a parent in jail or prison, up from 500,000 in 1980 (Sykes and Pettit 2014). These 2.6 million children represented 11.4 % of all black children, 3.5 % of Hispanic children, and 1.8 % of white children in the United States (Sykes and Pettit 2014). Cumulative exposure estimates indicate that by age 17, 24 % of black children, 11 % of Hispanic children, and 4 % of white children experience some form of parental incarceration, and exposure is markedly higher for children whose parents have not completed high school (Sykes and Pettit 2014; Wildeman 2009).
Prior research has linked parental incarceration to household-level instability and disadvantage (Geller et al. 2009, 2011; Wildeman 2014) as well as child-level behavioral problems, such as increased externalizing behavior, greater delinquency, lower educational attainment, and greater substance abuse (Cho 2011; Foster and Hagan 2013; Hagan and Foster 2012; Murray and Farrington 2005, 2008; Roettger and Swisher 2011; Wildeman 2010). Other research has examined the various ways in which parental incarceration may cause children and young adults to be excluded from important social institutions (Foster and Hagan 2007, 2015b; Murray 2007). The social lives and peer relationships of children of incarcerated parents, however, have received less attention. Qualitative studies have suggested that children of incarcerated parents experience social isolation (Bernstein 2007; Nesmith and Ruhland 2008; Sack et al. 1976), but these studies have used small, unsystematic samples without comparison groups, and “social isolation” is not well-defined.
This study extends research on the collateral consequences of parental incarceration by investigating whether and how adolescents with recently incarcerated fathers in a nationally representative data set are disadvantaged in their social lives, exploring in particular their social location and their friends’ characteristics. This is an important step because networks not only affect well-being during adolescence but also are a critical pathway by which parental incarceration could negatively affect children’s behavior and life chances.
Adolescent Social Networks and Peer Groups
Adolescent peer groups differ greatly in their level of delinquency, academic orientation, and structure (Coleman 1961; Ryan 2001), which are all strongly associated with later academic achievement and deviance. As hypothesized by differential association and differential reinforcement theory (Akers 1985; Burgess and Akers 1966; Sutherland and Cressey 1955), empirical analyses have found that friends influence adolescents’ engagement in deviant and delinquent activities, including alcohol and drug use and criminal behavior (Clark and Lohéac 2007; Haynie 2002; Ingram et al. 2007; Kandel 1978). Delinquency among one’s friends also contributes to children’s school failure both directly and indirectly through diminished academic achievement (Battin-Pearson et al. 2000). Moreover, having low-achieving friends can reduce an adolescent’s own academic aspirations and achievement (Davies and Kandel 1981; Flashman 2014; Ryan 2001).
Location within peer social networks may also influence adolescents’ behavior and academic success. Centrality—that is, being well connected within a network—predicts academic achievement: better-connected adolescents perform better in school (Calvó-Armengol et al. 2009). Conversely, social isolation and rejection by peers during childhood predict subsequent school failure and delinquency (Ollendick et al. 1992; Parker and Asher 1987). Social location also conditions the effect of friends’ delinquency on an adolescent’s own delinquency. Popularity and centrality are associated with lower delinquency for adolescents embedded in nondelinquent friend groups, while these characteristics are associated with increased delinquency for adolescents embedded in delinquent friend groups (Haynie 2001).
In the context of historically high incarceration rates, parental incarceration may shape adolescent social networks, ultimately contributing to the delinquency and diminished achievement commonly observed among children of incarcerated parents (e.g., Aaron and Dallaire 2010; Cho 2011; Hagan and Foster 2012; Murray and Farrington 2005; Roettger and Swisher 2011). I illuminate this potential pathway by examining the social location of adolescents experiencing paternal incarceration and the types of peers they befriend.
Mechanisms and Hypotheses
Networks of children with incarcerated parents are likely to differ from other adolescents’ networks for two main reasons: behavioral differences and stigma.
Parental incarceration is associated with internalizing behaviors and depression (Foster and Hagan 2013; Murray and Farrington 2008; Wakefield and Wildeman 2011). Such behaviors can lead children to withdraw from social networks, making them more socially isolated, with fewer friends than their peers (Laursen et al. 2007). Parental incarceration also increases children’s aggressive and antisocial behaviors (Geller et al. 2012; Murray and Farrington 2005; Murray et al. 2012; Wildeman 2010), which could increase a child’s risk of social isolation (Cairns et al. 1988; Laursen et al. 2007). Additionally, aggressive adolescents tend to associate with other aggressive adolescents, and early aggression is associated with delinquency during adolescence (Cairns et al. 1988; Vitaro et al. 1997). Thus, adolescents who experience parental incarceration are likely to be socially isolated, with more antisocial, delinquent friends than other adolescents.
In addition to behavioral differences, stigma may mediate the association between parental incarceration and adolescents’ social network characteristics. Stigma is fundamentally a social status, a socially conferred judgment of moral contamination that attaches to one’s biography, appearance, or social connections (Goffman 1963). Accordingly, peer groups are a particularly important domain in which adolescents are likely to experience and ascribe stigma (Moses 2010).
Although many scholars have hypothesized that the poor outcomes often observed among children of incarcerated parents are at least partly due to stigma (Besemer et al. 2011; Gabel 1992; Hagan and Dinovitzer 1999; Murray and Farrington 2005; Murray and Murray 2010; Phillips and Gates 2011; Western and Wildeman 2009), few studies have examined stigma directly. Qualitative studies have suggested that children and families of incarcerated individuals experience stigma and shame (Bernstein 2007; Braman 2004; Nesmith and Ruhland 2008; Sack et al. 1976), and experimental findings have suggested that teachers stigmatize children with incarcerated parents (Dallaire et al. 2010). However, no one has yet sought evidence of parental incarceration stigma in the context in which it is most likely to appear: social networks.
Goffman (1963) claimed that stigmatized individuals are socially isolated, either because they are shunned or because they avoid social interactions in expectation of being shunned. Consequently, they connect with other stigmatized or marginalized individuals who share their situation. Thus, we should expect children with incarcerated parents to have fewer friends and be more marginal in social networks. We would also expect them to disproportionately befriend other stigmatized adolescents, like those who also have an incarcerated parent or who have experienced other stigmatic forms of family disruption. I test both possibilities, assessing the proportion of their friends who have also experienced parental incarceration and the proportion who come from two-parent households. I also examine whether adolescents with incarcerated parents withdraw from social networks or are shunned by examining how many peers they name as friends versus how many peers name them as friends. Finally, I consider how well connected these adolescents are to the rest of their in-school peers.
Both channels—stigma and behavioral differences—generate the same hypotheses: adolescents experiencing parental incarceration will be more socially isolated and have fewer, more-delinquent, and more-disadvantaged friends than other adolescents. Because stigma is much more difficult to operationalize and measure than behavioral problems, I primarily focus on controlling for behavioral characteristics in the following analyses. In the model that controls most fully for behavioral differences, I assume that stigma is likely the primary driver of remaining differences in social network characteristics between children who have experienced paternal incarceration and those who have not.
Data, Measures, and Analytic Approach
Add Health is a U.S. nationally representative survey that has followed more than 15,000 adolescents from grades 7–12 through adulthood. Respondents were initially surveyed in the 1994–1995 school year when they were 11–19 years old, with follow-up interviews conducted in 1996, 2001–2002, and 2008. Add Health used stratified sampling to select 80 high schools—defined as containing grade 11 and more than 30 students—representative with respect to region, urbanicity, size, type, and ethnic mix. For the 52 high schools not spanning grades 7–12, one feeder middle school was selected, yielding a total of 132 schools in 80 U.S. communities. In-school surveys were administered to all students present on the day of the initial survey (N = 90,118). Approximately one-quarter of students in these schools (N = 20,745) were selected for the longitudinal in-home study sample. More than 75 % (N = 15,701) of original longitudinal sample members participated in the 2008 survey (Harris 2013).
The Wave I in-school survey asked respondents to nominate up to five male and five female friends in order of closeness.1 From these nominations, Add Health created within-school global and egocentric network structure measures, several of which I use as dependent variables. Because most nominated friends also completed in-school surveys, friend characteristics can be calculated from friends’ self-reports rather than respondents’ perceptions of friends, which has been a limitation of previous studies that lack complete network data (e.g., Warr and Stafford 1991; Zimmerman and Messner 2010).
Parents, usually the resident mother, also completed interviews at Wave I for 85 % of longitudinal sample members (N = 17,670). The fourth survey wave in 2008 collected retrospective data on parental incarceration from respondents. Because parental incarceration data are available only for respondents who participated in Wave IV, the analytic sample is restricted to longitudinal sample members who provided valid responses to questions about biological father incarceration history in Wave IV and who could be matched to their Wave I in-school questionnaires (N = 11,356). Table S1 in Online Resource 1 shows a detailed schematic of sample restrictions.
Add Health is unique in containing detailed information about both adolescent networks and parental incarceration from a nationally representative sample. The survey’s school-based design enables reliable calculation of social location measures in a complete network, as well as accurate estimation of friends’ self-reported characteristics. The school-based sampling frame may underobserve high-risk teenagers who dropped out or have a high rate of absenteeism, however. Paternal incarceration effects may thus be underestimated with this school-based design, given that children of incarcerated parents are more likely to drop out of school than other adolescents (Cho 2011).
Key Independent Variable: Father Incarcerated in Three Years Before Wave I
Respondents indicated in Wave IV whether any of their parents had ever been incarcerated, how many times each was incarcerated, and their own age when each was first incarcerated and last released. Because parental incarceration history is collected through retrospective respondent reports, rather than directly from parents or corrections records, there is likely to be underreporting—or misreporting of timing—for parental incarceration spells that occurred before the respondent’s birth or early in her life. For such incarcerations, respondents can presumably report only what family members have told them because they are unlikely to recall these events themselves. Partially for this reason, these analyses focus on paternal incarceration that respondents reported experiencing during adolescence—a period during which their recollections are likely to be more accurate. I use respondents’ age at first paternal incarceration and last release to create an indicator variable identifying respondents whose biological father was incarcerated at some point in the three years prior to and including the baseline survey (N = 467), which collected complete in-school network data. I also focus on paternal incarceration in this three-year window on the assumption that more-recent paternal incarceration is likely to be more salient for both the respondent and her peers. I also consider paternal incarceration at any time prior to Wave I and at various other ages. These findings are discussed briefly following the main results.2
To assess whether and how adolescents’ network size and social location differ by paternal incarceration history, I examine total number of friends nominated by the respondent, number of friend nominations received from other students in her school, whether the respondent is a social isolate, respondent’s centrality within her school social network, and her network reach in three steps within school.3
Whether a respondent is an isolate—that is, no friends are nominated, and no nominations are received—is the purest test of social isolation. However, examining her friend group size relative to that of her peers can also help illuminate the extent of social isolation she faces within her school. Examining how many friends a respondent nominates versus number of nominations received provides insight as to whether she isolates herself from peers or is excluded by in-school peers. Bonacich centrality captures respondent’s prominence in her whole school social network: it is essentially a measure of how many friends a respondent has, weighted by those friends’ popularity (Bonacich 1987; Carolina Population Center 2001).4 I standardize Bonacich centrality scores to ease interpretation of coefficients. Network reach in three steps captures each respondent’s extended social network size by counting how many students she is connected to in three steps (i.e., friends of friends’ friends), providing an indication of the extent to which the respondent and her friends are socially isolated within their school.5
I also examine differences in the following friend characteristics: proportion of friends with an incarcerated parent, proportion from two-parent households, mean GPA, and mean level of delinquent behavior. These measures can be calculated only for those friends who completed the in-school survey.6 See Table 3 in the appendix for detailed variable coding descriptions.
Sample size varies across outcome measures depending on whether the response rate in the respondent’s school was high enough to create reliable social location measures,7 whether the respondent nominated any friends in her school or sister school, and how many of her friends completed the in-school survey and provided valid responses on relevant items (e.g., GPA, household composition).8 The sample size for each model is reported in Table 2.
Complete social network data for all schools are collected only at Wave I, preventing comparison of network outcomes before and after paternal incarceration.9 Therefore, I employ two strategies to account for confounding factors and test the sensitivity of the relationship between paternal incarceration and adolescent social network characteristics: (1) regression with controls, and (2) comparison group restriction to a subsample likely to receive “treatment” (i.e., recent paternal incarceration).10 In all models, I use random or fixed effects to account for clustered observations and correlated error structure within schools.11
Because network data were collected in Wave I, Add Health does not contain pretreatment observations of respondents. In the absence of pretreatment data, I first estimate a regression model with school random effects that controls only for demographic characteristics that could not have been affected by paternal incarceration: race, age, and gender.
The next regression model adds posttreatment controls that may have been affected by paternal incarceration but nevertheless represent factors that could confound the relationship between paternal incarceration and network and friend characteristics. By controlling for variables that may be either confounders or mediators, this second model is a more conservative estimate than the first, which includes only exogenous covariates. These models are thus intended as upper and lower bounds on the causal association.
Because they experience greater housing instability (Geller et al. 2009; Tasca et al. 2011), children of incarcerated parents may change schools more often than other children and may therefore have less time to establish friendships. To account for residential mobility, I control for the student’s self-reported years in attendance at her current school. Additionally, I control for respondent’s GPA and standardized delinquency index score to account for the possibility that network differences between children of incarcerated fathers and their peers are attributable to behavioral differences between these two groups. I also control for parental education; but because not every respondent lives with a mother and father figure, I interact mother and father figures’ education with dummy variables indicating whether the respondent had a mother or father figure present in her household at Wave I. I multiply impute missing values for years in current school, GPA, delinquency index score, and mother and father figure education (when present in the household). The appendix contains detailed variable coding descriptions.
This fuller model also adds school fixed effects because certain local or school-level characteristics (e.g., prevalence of parental incarceration or size and diversity of the student body) may confound the relationship between paternal incarceration and social network characteristics. Fixed-effects models yield the average difference in network characteristics within schools for children with and without a recently incarcerated father, thus controlling for all school-level variables correlated with paternal incarceration and adolescent networks. Because Add Health schools are located in different communities, the school fixed effects should also reduce bias due to unobserved community-level characteristics associated with incarceration and social networks.
In the final model, I restrict the comparison group to respondents whose father was last released from prison when they were age 5 or younger—that is, before they were enrolled in school and could therefore be subject to marginalization by school-based peers. Restricting comparison to individuals similar to the treatment group, and thus at risk of receiving treatment, can significantly reduce bias in the estimation of causal effects compared with relying on a general population comparison group with regression adjustment (LaLonde 1986; Western 2002). Porter and King (2015) followed this approach with Add Health data, selecting a strategic comparison group for children who experience paternal incarceration prior to Wave I to account for unobserved characteristics that may confound the effect of paternal incarceration in the absence of pre- and posttreatment observation of outcomes of interest.
Because I focus on paternal incarceration in the three years leading up to Wave I, I can use paternal incarceration in early life as the basis for my comparison group, which helps account for both familial and behavioral differences that may confound the relationship between paternal incarceration and social network characteristics. Fathers who were still criminally active in their children’s early years are likely more similar to fathers still criminally active in their children’s later childhood (three years leading up to Wave I) than fathers who ceased criminal activity before their child’s birth or continued it into their child’s young adulthood (after Wave I). Moreover, previous research has indicated that paternal incarceration is associated with behavioral problems in early childhood that are likely to persist through later childhood and adolescence (Geller et al. 2012; Wakefield and Wildeman 2011; Wildeman 2010). Therefore, restricting the comparison group to respondents who experienced paternal incarceration when they were age 5 or younger should reduce the chance that differences in social network outcomes are due wholly to unobserved differences in familial and behavioral characteristics between treatment and control groups not captured by my covariates. Thus, this model should control best for behavioral differences between the treatment and control groups. Consequently, findings from this model may provide the strongest test of the proposed stigma mechanism.12
In this model, I control for all covariates used in the second, fuller regression model. Because many respondents who recently experienced paternal incarceration do not have a comparison group member who experienced early paternal incarceration in their school, I use random effects rather than fixed effects to account for clustering within schools.
I present results from these three models, which tackle the selection bias problem via different means, in order to provide a range of reasonable estimates of the relationship between paternal incarceration and adolescent social network characteristics. Although I cannot present one model that strongly identifies the causal effect of paternal incarceration on network outcomes because of data constraints, consistency in the estimates across models suggests a likely causal relationship between paternal incarceration and adolescent social network outcomes.
Table 1 displays control variable characteristics for analytic sample members by paternal incarceration history. Just over 4 % of the analytic sample experienced paternal incarceration in the three years preceding Wave I.13 Adolescents with a recently incarcerated father have a mother figure in their household about as often as other adolescents, but they tend to be slightly younger, slightly newer to their schools, and more delinquent than students who have not experienced recent paternal incarceration. They are also more likely to be black, Hispanic, and female, with lower GPAs and less-educated mother and father figures than adolescents whose fathers have not been incarcerated in recent years. The restricted comparison group model is successful in generating a more-similar comparison group; respondents who last experienced paternal incarceration at or before age 5 are more similar to respondents recently exposed to paternal incarceration, particularly on behavioral characteristics.
Table 2 displays estimates of the relationship between recent paternal incarceration and social location and friend characteristics. (Tables S3–S5 in Online Resource 1 display control variable coefficients and standard errors.) The bivariate associations in column 1 show that the social networks of adolescents with recently incarcerated fathers differ significantly from other adolescents’ networks. On average, respondents who reported recent paternal incarceration are more socially isolated than their peers: they nominate 0.5 fewer friends (7 % fewer than average) and are named as a friend by 0.34 fewer students in their school (8 % fewer than average). Because total social isolation is uncommon (experienced by only 1.9 % of the full sample), the 1.4 percentage point higher prevalence of isolates among adolescents with recently incarcerated fathers is 74 % higher than the prevalence among adolescents without recently incarcerated fathers.14 Respondents with recently incarcerated fathers are also less connected to other students in their schools. Their centrality scores are one-quarter standard deviation lower, and they can reach nine fewer students in three steps than other adolescents. These differences are all consistent with the hypotheses and statistically significant.
Adolescents with recently incarcerated fathers also have friends who are less advantaged, less academically successful, and more delinquent, on average, than other adolescents’ friends; the proportion of their friends who experience parental incarceration is 6.5 percentage points higher, the proportion living in two-parent households is 8.6 percentage points lower, their friends’ mean GPA is 0.16 points lower, and their friends’ average delinquency scores are 0.13 standard deviations higher. Each difference is statistically significant.
Coefficients in the second column represent residual differences in social network outcomes after controlling for demographic differences in race, gender, and age, none of which could plausibly be affected by paternal incarceration. Adjusting for these compositional differences reduces between-group differences in all outcomes except friends’ delinquency. Only the difference in number of friendship nominations received, however, is no longer statistically significant, although the difference in isolate status probability becomes only marginally significant (p < .10).
The third column in Table 2 displays coefficients from the regression model that adds potentially confounding variables that may be endogenous to paternal incarceration (years in current school, GPA, delinquency, mother/father figure presence in household, and mother/father figure education) and school fixed effects. Coefficients in this column thus reflect within-school differences in social location and friend characteristics. Because the potentially confounding characteristics included in this model were measured posttreatment, estimates in column 3 may underestimate the size of the relationship between paternal incarceration and network outcomes.
After these additional controls and school fixed effects are included, coefficient magnitudes drop to 40 % to 80 % of their bivariate association size. Except for number of friend nominations received, however, the differences remain statistically significant, indicating that compositional differences between adolescents who did and did not experience recent paternal incarceration and differences in school context do not fully explain the differences observed in column 1. In this third model, net of controls and school fixed effects, adolescents who experienced recent paternal incarceration nominated 0.36 fewer friends, on average, than other adolescents—a magnitude roughly equivalent to the difference seen for respondents with three fewer years at their school than the average respondent (see Table S4 in Online Resource 1).15 They are also still 1.1 percentage points more likely to be isolates, which is roughly equivalent to the difference in isolation seen for respondents with five fewer years than average in their current school. Adolescents who recently experienced paternal incarceration now have centrality scores 0.15 standard deviations lower than their peers’, and their extended networks (in three steps) contain approximately six fewer students, which is 11 % lower than average network reach for all respondents. To help put these magnitudes in context, this difference in centrality is roughly equivalent to the difference seen for respondents with 2.5 fewer years at the school or a GPA one full point lower than average. The difference in extended network reach is similar to the difference associated with a 2 standard deviation increase in delinquency.
Significant differences in friend characteristics also remain with this fuller set of controls and school fixed effects. Compared with in-school peers, adolescents who experienced paternal incarceration in the last three years have a higher proportion of friends who have also experienced parental incarceration (4 percentage points higher than in-school peers, or 38 % higher than average) and a smaller share of friends from two-parent households (5 percentage points lower). These friends also have significantly lower GPAs (by 0.07 points) and are significantly more delinquent (0.09 standard deviations higher) than other adolescent’s friends. The difference in friends’ delinquency is approximately equal to what we see for a respondent whose GPA is a full point lower than the average respondent’s.
The last column in Table 2 displays network outcome differences when the comparison group is restricted to students who previously experienced paternal incarceration but not since age 5. The model used in this column includes the full set of covariates used in the previous column and random effects to account for student clustering within schools. As noted earlier, behavioral differences between the treatment and comparison groups should be minimized in this model: children who experienced paternal incarceration in early life should be at greater risk of receiving the treatment than adolescents who never experienced paternal incarceration and will have already been exposed to the behavioral effects of early childhood paternal incarceration. Stigma, conversely, should be greater for more recent paternal incarceration spells, particularly those occurring after the respondent was enrolled in school, when paternal incarceration could become known to school-based peers. Therefore, the estimates from this model should provide the strongest test of the stigma hypothesis.
Sample sizes are much smaller, and accordingly, standard errors are larger for the coefficients in this column. Still, we see significant differences in number of friends nominated, centrality, extended network size, friends’ academic performance, and friends’ delinquency. Except for friends’ mean GPA, these coefficients all fall somewhere between those estimated in the previous models. The difference in friends’ GPA is smaller than previous estimates but is still substantively similar: adolescents with recently incarcerated fathers have friends with mean GPAs 0.06 points lower than those of the friends of adolescents whose fathers have not been incarcerated since they were age 5 or younger.
Although the direction has not changed, differences in the probability of isolate status, proportion of friends with an incarcerated parent, and proportion of friends in two-parent households are roughly one-half as large as they were in the first two columns and are no longer significant. This result may indicate that the earlier observed differences in these characteristics are primarily driven by behavioral differences between adolescents who have recently experienced paternal incarceration and their peers, more so than by paternal incarceration stigma.
As in the previous models, difference in friend nominations received is not statistically significant, and the coefficient even becomes positive in the restricted comparison group model. This finding suggests that adolescents with recently incarcerated fathers are not necessarily shunned by their peers; instead, they may self-isolate to some degree, nominating fewer peers as friends.16 The fact that they have lower centrality scores, smaller extended networks, and less-advantaged, lower-achieving, and more-delinquent friends, however, suggests that they are befriending more-peripheral, less-advantageous peers within their schools.
I also estimated models accounting for household income and parental monitoring and closeness, finding substantively consistent results. Moreover, I tried a variety of specifications for the paternal incarceration variable—including incarceration at Wave I, incarceration any time between birth and Wave I, years of incarceration prior to Wave I, and a set of dummy variables representing age at paternal incarceration—consistently finding that paternal incarceration is associated with the same general patterns noted earlier.17 Additionally, I estimated Poisson regression models for the three count variables: total friends nominated, number of nominations received, and network reach in three steps. The results of these models are substantively similar to those presented earlier.
Because the relationship between recent paternal incarceration and network outcomes may be spuriously significant due to unobserved differences between children who experience paternal incarceration and those who do not, I conducted a falsification test in which I used paternal incarceration after Wave I to predict differences in social location and friend characteristics at Wave I using the full covariates and school fixed-effects model in column 3. With unobserved confounders, one would expect a significant association between network characteristics and post–Wave I paternal incarceration, even though post–Wave I incarceration could not have influenced Wave I network outcomes. Insignificant estimates in a falsification test thus help rule out unobserved confounders.
Results for seven of the nine dependent variables are insignificant, as one would hope. The only outcomes that differ significantly by paternal incarceration post–Wave I—proportion of friends living in two-parent households (p < .01) and number of friend nominations received (p < .05)—are outcomes that were not significant in the final restricted comparison group model, suggesting that they may primarily be due to behavioral differences. Moreover, number of friend nominations received was never significantly different beyond the bivariate associations. Thus, the falsification test results suggest that the differences in number of friends nominated, isolate status, centrality, extended network reach, proportion of friends with incarcerated parents, friends’ average GPA, and friends’ average delinquency observed earlier are not simply an artifact of unobserved differences in adolescents who experience paternal incarceration. Table S6 in Online Resource 1 displays these results.
I also tried restricting the comparison group to respondents who experience paternal incarceration after Wave I, following Porter and King (2015). Magnitudes differ, but results are substantively consistent with those displayed in Table 2. Appendix Table S7 (Online Resource 1) shows these results.
Because the effects of parental incarceration may differ by child gender (Roettger and Boardman 2012; Wildeman 2010) and because incarceration rates in the United States differ widely by race, I also examined variation in the paternal incarceration coefficients across four race and gender subgroups using the school fixed-effects and full covariate model from column 3. Figure 1 displays the coefficients from separate race and gender subgroup regressions using standardized versions of the dependent variables. (Table S8 of Online Resource 1 reports coefficients, standard errors, and sample sizes.)
In general, paternal incarceration coefficients are signed consistently across subgroups. Paternal incarceration is associated with nominating fewer friends, higher probability of isolate status, being less connected within one’s school, having fewer friends from two-parent households, and having friends who are less academically successful and more delinquent than other adolescents’ friends across race and gender subgroups. Models including gender and race interactions with paternal incarceration do not reveal any significant differences by gender. The only significant difference by race is in the relationship between recent paternal incarceration and friends’ average GPA; for black respondents, paternal incarceration is not associated with friends’ GPA.
Overall, the results demonstrate a strong and robust association between paternal incarceration and social marginalization among adolescents, even when common demographic differences and likely confounders are controlled for and comparisons are limited to the most similar comparison group. Children who experience paternal incarceration tend to report friend groups that are smaller, more socially isolated, and more disadvantageous than those of their peers. Moreover, the results from the restricted comparison group model suggest that most of these differences, apart from probability of isolate status and characteristics of friends’ households and parents, may be due more to stigma than behavioral differences between adolescents who experience paternal incarceration and those who do not.
Discussion and Conclusion
Parental incarceration is associated with a wide variety of both household- and child-level disadvantages (Wildeman and Western 2010). This study brings a network analysis approach to the little-examined social lives of children with incarcerated parents. I find that adolescents who experience paternal incarceration are embedded in less-advantaged, more-peripheral social relationships than their peers, indicating that children of incarcerated fathers experience social marginalization in their schools in addition to the resource deprivation and other disadvantages they face at home.
Previous qualitative studies have suggested that children of incarcerated parents are likely to experience social isolation, but if we characterize social isolation as meaningful differences in the number of one’s friendship ties and one’s probability of being an isolate, my findings provide limited support for this hypothesis. Differences in number of friendship nominations received are not meaningful, and although differences in the probability of being a social isolate are significant in most models, very few adolescents—with or without incarcerated fathers—are true isolates. Differences in the number of friends adolescents with recently incarcerated fathers nominate are consistent and significant, but it is unclear whether the magnitude (0.4–0.5 fewer friends) is large enough to be substantively meaningful.
These results suggest that children with incarcerated fathers are better described as socially marginalized than as socially isolated. The adolescents experiencing paternal incarceration in Add Health do have friends, although they might nominate slightly fewer. However, they are more peripheral within their schools—they are less central and can reach fewer other students in a few steps—and they befriend more-marginal peers: their friends get significantly lower grades, are significantly more delinquent, are more likely to come from single-parent households, and are more likely to experience parental incarceration than the average student’s friends. These results are robust across a variety of models and comparison groups, a falsification test, and race-gender subgroups.
The findings are consistent with the hypothesized effects of parental incarceration stigma, lending support to the general hypothesis that stigma contributes to poor outcomes for children of incarcerated parents. However, the finding that adolescents with recently incarcerated fathers are disadvantaged on every network measure except nominations received suggests that parental incarceration stigma may not necessarily cause adolescents to be shunned by their peers, but may instead lead them to withdraw from social groups and may shape the types of peers they befriend. Future research that focuses on isolating and manipulating the extent of stigma could help explain the role stigma plays in the lives of children of incarcerated parents and the extent to which it is experienced externally versus internally. Moreover, stigma may arise not just from parental incarceration itself but also from associated problems stemming from parental incarceration (e.g., behavioral issues, material hardship). Accordingly, future research that explores the root causes of stigmatization experienced by children of incarcerated parents would be valuable. Finally, testing whether the incarcerated parent’s gender affects extent of stigma would also be helpful, given that these results pertain only to paternal incarceration.
The finding that adolescents who experience paternal incarceration are in less-privileged social network positions and associate with more-disadvantageous friends—even after their own achievement and behavior are accounted for—may also help explain previous research linking parental incarceration to higher delinquency and lower academic achievement in adolescence. Given that academic achievement and delinquency are influenced by friends as well as social location in one’s school (Calvó-Armengol et al. 2009; Flashman 2014; Haynie 2002; Vitaro et al. 2000; Weerman and Smeenk 2005), these findings suggest that social networks may be a mechanism through which children of incarcerated parents become more delinquent and lower achieving. The strength of this mechanism could be tested in the future with data that observe the social networks of children with incarcerated parents in early adolescence and then observe academic achievement and delinquency in late adolescence.
Despite the consistency of the results under a variety of specifications, selection bias poses challenges for identifying the causal effects of incarceration on either adult or child outcomes. In particular, behavioral characteristics of fathers, such as criminal involvement and propensity for violence, are unobserved in Add Health. In the absence of data on these paternal characteristics, selection bias could be addressed by pre- and post-incarceration observations, but full network data are collected only in the first wave of Add Health, and no other data currently exist that measure children’s social network characteristics both before and after parental incarceration.
Although isolating exogenous variation in parental incarceration is challenging with observational data, the robust associations between paternal incarceration and network outcomes under a variety of adjustments for observable factors and for different comparison groups lend confidence that the observed differences are not only an artifact of unobserved confounding. To the extent that potentially confounding father characteristics, such as impulsivity and propensity to violence, are correlated with children's own delinquency, they are controlled for in the last two main models. Moreover, I address selection bias concerns as best as I can with the final restricted comparison group model. Future research could improve the current estimates with more-detailed measurement of paternal behaviors and child characteristics prior to incarceration. Despite the challenges of causal identification, the results clearly show inequality in the distribution of social capital associated with paternal incarceration.
Most research on adolescent social networks has treated network characteristics as independent variables, rarely considering the factors that shape social networks in adolescence. The few studies that have examined determinants of adolescent social networks have focused on race, health, and neighborhood-level characteristics as predictors (Goodreau et al. 2009; Haas et al. 2010; Harding 2008; Strauss and Pollack 2003). My findings provide evidence that household-level characteristics also influence adolescents’ social networks. Additional research may reveal other household-level experiences and context that affect adolescent networks.
In connecting paternal incarceration to childhood social network disadvantage, these findings also provide further evidence of the social exclusion associated with mass incarceration. Prior research has established mass incarceration as an institution for social exclusion and isolation for adults entangled in the system (Murray 2007; Travis 2002; Uggen et al. 2006). Previous studies have provided evidence of social exclusion—that is, disconnection from important social institutions—crossing generational bounds to affect the children of America’s prisoners (Foster and Hagan 2007, 2015a), but this study is the first to use a networks perspective to provide rigorous quantitative evidence of the social marginalization experienced by children of incarcerated fathers.
Moreover, in identifying the social network disadvantages faced by children of incarcerated fathers, this study contributes to existing literature on the constriction and disadvantage of social networks among the poor in modern America (Desmond 2012; Sampson et al. 1997; Smith 2007). These findings suggest that the disadvantages associated with paternal incarceration permeate even adolescent social networks, embedding the children of prisoners in locationally and compositionally marginal social groups. Given that 24 % of black children, 11 % of Hispanic children, and 4 % of white children in America experience parental incarceration by the time they turn 17 (Sykes and Pettit 2014), these network disadvantages are likely to touch far more than the 4.1 % of affected Add Health respondents on which this study focuses. The long-term implications of these social network differences are not yet clear but should be examined in future work.
A substantial body of research suggests that the punishment inflicted by America’s prison system extends well beyond those convicted of crimes and the stated duration of their sentences. This study highlights yet another context in which the children of prisoners are penalized for their parents’ infractions. By highlighting the more-marginal, less-advantageous friend groups in which children of incarcerated parents are embedded, this study enriches our understanding of the broad variety of ways in which both inmates and their children are marginalized in American society, providing further evidence of the social exclusion and inequality associated with mass incarceration.
I am incredibly grateful to Bruce Western and Alexandra Killewald for their patient guidance and thoughtful feedback throughout the course of this project. I also thank Christopher Jencks, Christopher Wildeman, Dana Rotz, Steven Raphael, Becky Pettit, and participants in the Quantitative Sociology workshop, Inequality and Social Policy proseminar, and Justice and Inequality reading group at Harvard University for their helpful comments and suggestions on earlier drafts of this article. This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Special acknowledgment is due to Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Information on how to obtain the Add Health data files is available on the Add Health website (http://www.cpc.unc.edu/addhealth). No direct support was received from Grant P01-HD31921 for this analysis.
Approximately one-half of respondents nominated the maximum of five male or five female friends, and 39 % nominated five of each. Therefore, friend count may be truncated for a sizable share of respondents. Because respondents were asked to nominate friends in order of closeness, however, ties who could not be nominated should constitute less-influential and important friends. Moreover, previous research has suggested close friend groups have about five or six members (Cotterell 1996; Dunphy 1963).
I focus on paternal incarceration because few respondents experienced maternal incarceration prior to Wave I (N = 232), and previous research has suggested that maternal and paternal incarceration affect children differently (Lee et al. 2013; Wildeman and Turney 2014).
Network location measures (nominations received, centrality, and reach in three steps) are calculated using ties only within the respondent’s school, excluding nominations sent to and received from students in the sister school (when applicable). Consequently, these measures are calculated separately at the middle and high school level for paired sister schools but jointly for combined middle and high schools. Findings are substantively consistent when combined middle and high schools are excluded from analyses.
Centrality tends to be highly correlated with network size measures because centrality is partly premised on network size. Correlations between standardized Bonacich centrality and network size measures are as follows: total friends nominated (.65), nominations received (.41), and reach in three steps (.67). These moderately high correlations reflect that these measures are conceptually related but distinct (Valente et al. 2008).
The Add Health restricted access network data file includes reach in three steps. Therefore, it is easily replicable and has already been used in other publications (e.g., Gallupe and Bouchard 2015; Hatzenbuehler et al. 2012; Mundt and Zakletskaia 2014). Moreover, three steps is an appealing distance because it should be large enough to extend beyond clique members to capture liaisons and adjacent cliques (Ennett and Bauman 1993) but small enough that peers at that distance could still be reasonably expected to have some influence over the respondent (Payne and Cornwell 2007).
Average friend characteristics are calculated using all nominations that could be linked to friends’ in-school surveys, including friends in the sister school. Most identifiable friend nominations (98 %) are sent to students in the respondent’s own school, however.
To ensure reliability, number of nominations received, Bonacich centrality, and reach in three steps are constructed for only those students in schools with response rates of 50 % or higher (121 schools) and for friendship nominations in which both the sender and the receiver are uniquely identifiable students in the same school.
For friend characteristics analyses, I limit the sample to respondents who nominated at least one identifiable friend (N = 10,146) in her school or sister school so that friends’ characteristics can be measured from in-school survey responses. Approximately 11 % of nominations were to friends not in the respondent’s school pair; another 8 % were to students whose names were not on the roster (Carolina Population Center 2001). Adolescents with recently incarcerated fathers were only marginally more likely to nominate individuals not on school rosters as friends.
Network data are collected again at Wave II for all respondents in 16 schools, but only nine respondents in these schools experienced paternal incarceration in the year between Waves I and II, preventing meaningful pre- and posttreatment comparison.
I also used within-school nearest neighbor propensity score matching to restrict comparison to more-similar respondents, but lack of pretreatment observation of respondent characteristics and behavior pose significant challenges to credible propensity score model specification. Results are similar and available upon request.
I also tried clustered standard errors and generalized estimating equation (GEE) models with exchangeable working correlation structures to account for error nonindependence within schools and found consistent results across models.
Because paternal incarceration is respondent-reported, the restricted comparison group may exclude respondents who fail to report paternal incarceration at or before age 5 because they do not recall or are unaware of it. However, because respondents are presumably more likely to know about early-life paternal incarceration that was consequential for the family or long in duration, respondents inappropriately excluded from the comparison group because of incomplete information are likely less similar to the treatment group—and more similar to respondents with never-incarcerated fathers—than respondents who are aware of early-life paternal incarceration. Consequently, this potential exclusion of relevant comparison group members likely means that I underestimate differences in social network outcomes between respondents who experienced paternal incarceration shortly before Wave I and those who did not experience it after age 5.
This is about twice as large as point-in-time estimates of parental incarceration in the late 1990s (Mumola 2000), which is reasonable given that I consider paternal incarceration over a three-year period.
For ease of interpretation, I report linear probability model regression coefficients for isolate status. Logistic regression results are consistent.
The estimated difference in number of friends nominated is likely to be biased toward 0 because of the five-person-per-gender limit on friend nominations.
I tested whether the discrepancy between number of friend nominations sent and number received could be explained by depression or time spent in household and/or paid labor, in case adolescents withdraw from friendships because of increased housework or financial demands in the absence of a parent. However, adding these potential mechanisms did not change results for these or any other outcomes.
In general, the magnitudes of differences in network characteristics are larger for more recent paternal incarceration and for longer duration of incarceration. The only exception is for the relationship between paternal incarceration and friends’ delinquency, which appears to be largest for respondents who experience paternal incarceration between ages 4 and 8.