Participants
The sample for the present study comprised 334 adolescents from Project AHEAD, a national longitudinal study of adolescent development in the United States. The sample was almost evenly divided between girls (
n = 164) and boys (
n = 169), with one participant identifying as gender queer (removed from the analysis to avoid conflation with cisgender youth’ experiences). Regarding ethnicity/race, 50% of the sample was non-Hispanic White (
n = 169), while 19% were African American (
n = 65), 15% were Latino/a/x (
n = 51), 9% were multiracial (
n = 31), 4% were Asian American (
n = 12), 1% identified as another ethnicity (
n = 5), and >1% were Native American (
n = 1). Mothers’ level of formal education was used as a proxy for social class (Harding,
2006), with 22% of the sample having mothers with a high school education or less (
n = 77), 46% with some college (
n = 159), and 31% with a 4-year degree or higher (
n = 105).
In October 2019, parents of adolescents were recruited using a third-party research service, Bovitz®, which retains a nationally representative panel of research participants. A stratified random sample of this panel was drawn using national quotas for gender, racial/ethnic identity, parent education, and geographic region. Inclusion criteria were that adolescents had to be between 14 and 17 years of age, and be in the 9th, 10th, or 11th grades at their schools. Just under 1,000 parents were contacted through the service’s online survey platform. A description of the study was provided that allowed parents to consent to their children’s participation. Parents were then asked to provide the survey to their adolescent child. In total, 570 adolescents assented and completed the survey at Time 1 (T1) in October 2019. Follow-up surveys were administered every six months thereafter (April 2020, October 2020, April 2021), for a total of four waves. At these ensuing waves, an email invitation was sent to all participants that included a link to the survey. Upon opening the invitation, parental consent and adolescent assent were obtained. Assenting adolescents were directed to the survey, which asked about their experiences and attitudes with academics, interpersonal relationships, and mental health. To ensure validity of responses, attention checks were implemented at each wave and all responses were back validated with prior waves to ensure consistency of identifying data (e.g., birthdates). Surveys took approximately 30 minutes to complete, and adolescents were compensated with a $20 Amazon e-gift card at each wave for participating. All procedures were approved by the Brigham Young University IRB.
A total of 570 adolescents began the study at Wave 1. For the purposes of this analysis, only those who participated in at least three of the four waves, for a total of 334 (59%). T-tests and chi-square analyses were conducted to examine patterns of attrition. Those who dropped out of the study were not meaningfully different than those who participated in 3 or more waves on most of the study variables, including the socio-demographic controls. The only exception was with homophobic victimization. Those who did not participate in at least 3 assessments reported slightly higher levels of homophobic victimization (
M = 1.28,
SD = 0.61) compared to those who were retained for 3 waves or more (
M = 1.20,
SD = 0.52). This may have been due to increased participant stress due to frequent victimization. However, this difference produced a minimal effect size (Cohen’s
d = 0.15). Conventional interpretations of effect size consider anything lower than
d = 0.20 to be negligible in practice (i.e., “merely statistical”; Fritz et al.,
2012), so the analyses were retained as designed. Power analysis showed that a sample size of 300 is adequate to detect relatively small effects that are common in related research (
r = 0.20; Faul et al.,
2007). In multilevel models, the size of the highest-order level (in this case individuals) is the most important limiting factor of a study’s power (Snijders,
2005).
Measures
Masculine-Typed Behaviors
At all four waves, masculine-typed behaviors were measured using the Adolescent Masculinity in Relationships Scale (Chu et al.,
2005) as adapted by Rogers and colleagues (2017) to reflect the degree of endorsement of broadly recognized masculine gender roles within one’s relationships, including emotionally restrictive behavior, physical toughness and aggression. Specifically, participants indicated their level of agreement to eleven items on a Likert scale of 1 (
strongly disagree) to 5 (
strongly agree). Items were averaged such that higher scores indicated more traditionally masculine social behaviors. Example items include “I cannot respect a friend who backs down from a fight” and “If I tell my friends my worries, I will look weak.” See Appendix A for a complete list of items included in this measure. This scale demonstrated adequate internal consistency at all waves (W1 a = 0.82; W2 a = 0.82; W3 a = 0.84; W4 a = 0.84), and in prior studies has shown construct validity as a unidimensional assessment of endorsement of traditionally masculine behavior (Rogers et al.,
2017).
Social Support
At all waves, participants reported their perceived degree of social support from friends using the friend’s subscale of the Multidimensional Scale of Perceived Social Support (Zimet et al.,
1988). Participants indicated their agreement with four items such as
“I can count on my friends when things go wrong” on a Likert scale of 1 (
strongly disagree) to 7 (
strongly agree). Responses were averaged such that higher scores reflected greater perceived social support (W1 a = 0.91; W2 a = 0.88; W3 a = 0.90; W4 a = 0.90).
Negative Peer Treatment
Negative peer treatment was assessed at all timepoints using four items from the Peer Interactions subscale of the Early Adolescent Role Strain Inventory (EARSI; Fenzel,
1989). Participants rated how often they experienced negative treatment by peers (e.g., “How often are other students mean to you?” and “How often do other students exclude you from activities?”) on a 5-point rating scale (1 =
Never, 5 =
Almost Always). Items were averaged to create mean scores (α = 0.85), with higher scores reflecting more experiences of negative treatment from peers. Construct and convergent validity for the EARSI and its subscales has been demonstrated previously (Fenzel,
1989) and reliability was good at all waves in the present sample (W1 a = 0.86; W2 a = 0.85; W3 a = 0.87; W4 a = 0.87).
Homophobic Name-calling
Homophobic name-calling was measured at all waves using a modified version of the Homophobic Content Target Scale (Poteat & Espelage,
2005). The five-item scale was assessed using a 5-point Likert scale (
1=never, 2 =
1 or 2 times, 3 =
3 or 4 times, 4 =
5 or 6 times, 5 =
7 or more times). Five items asked how often participants had been the victim of homophobic name-calling. An example item includes “How many times in the past week has a classmate called you [gay, lesbo, fag, etc.]?” Reliability for this measure was good (W1 a = 0.87; W2 a = 0.86; W3 a = 0.77; W4 a = 0.87).
Socio-demographic Variables
At wave 1, participants reported their gender (0 = girl, 1 = boy). They also reported their ethnic identity (African American, Asian American, Latinx/Hispanic, White, and Other). For analysis, dummy variables for individual ethnic groups were considered, but cell sizes for most minority groups were small and underpowered in later analyses. To avoid Type II error, ethnicity was recoded for ethnic minority status (0 = non-Hispanic white; 1 = non-white ethnic minority). Finally, they reported their mothers’ highest level of formal education (1 = Less than High School, 2 = High School or Equivalent, 3 = Some College or Vocational Degree, 4 = Four-year College Degree, 5 = Master’s Degree, 6 = Doctoral or Professional Degree).
Plan of Analysis
Intraclass correlations (ICCs) were estimated to determine the proportion of variance in masculine-typed behavior at the within and between-person levels. A multilevel modeling framework was used to estimate developmental trajectories in masculine-typed behaviors, with time being indicated by adolescents’ age (calculated by the date of their survey minus their date of birth). Multilevel models adapt elegantly to nested observations to produce within-person estimates of social processes (i.e., intercepts, slopes, or rates of change). They also allow for the estimation of individual differences in these within-person trends. Furthermore, in accounting for the nested nature of the data, between-person traits and characteristics are controlled by virtue of the design itself, further enabling the estimation of unique, time-specific intraindividual associations.
Primary analysis began by using a model building approach to first find the best fitting growth trajectories. First, a no-growth model centered at age 14, where the intercept (but no slope) was estimated. Following, a linear slope was introduced, indicated by adolescent age. Then, a quadratic term (age2) was added to the model. For example, the multilevel equation for a model retaining the linear slope would be expressed as:
Level 1 Model:
$${{{\mathrm{Masc}}}}_{{{{\mathrm{ij}}}}} = \beta _{0{{{\mathrm{i}}}}} + \beta _{1{{{\mathrm{i}}}}}\left( {{{{\mathrm{age}}}}} \right) + \varepsilon _{{{{\mathrm{ij}}}}}$$
Level 2 Model:
$$\beta _{0{{{\mathrm{i}}}}} = \gamma _0 + {{{\mathrm{U}}}}_0$$
$$\beta _{{{{\mathrm{ii}}}}} = \gamma _0 + {{{\mathrm{U}}}}_0$$
Interpreted, the masculinity score of an adolescent (i) at timepoint (j) was modeled at Level 1 as a function of an intercept, β0 (his/her cross-time average), a slope, β1 (the effect of his/her age), and residual within-person variance, ε. The intercept and slope were then modeled at Level 2 as a function of the sample average (γ00 and γ10, respectively), and residual between-person variance (U0i and U1i, respectively).
With the addition of each time polynomial (no-growth, linear slope, quadratic term), model fit was assessed using the -2 log likelihood (-2LL), the Akaike information criterion (AIC), the Bayesian information criterion (BIC) and the adjusted Bayesian information criterion (A-BIC). As these are comparative fit indices with no inherent metric or scaling, they are only useful for comparing increasingly complex, nested models (Field & Wright,
2011). Lower values indicated better fit to the data. A model was retained if it showed better fit than the previous, more parsimonious model.
Predictors of Between- and Within-Person Variance in Masculine-typed Behavior
After fitting the most appropriate growth model, individual differences (between-person variance) were next examined in trajectories of masculine-typed behaviors. Adolescent sex, ethnic/racial minority status, and mother’s formal education (an indicator of social class; Kim et al.,
2013) were included as time-invariant predictors of both the intercept and slope at Level 2. This model was expressed as:
Level 1 Model:
$${{{\mathrm{Masc}}}}_{{{{\mathrm{ij}}}}} = \beta _{0{{{\mathrm{i}}}}} + \beta _{1{{{\mathrm{i}}}}}\left( {{{{\mathrm{age}}}}} \right) + \varepsilon _{{{{\mathrm{ij}}}}}$$
Level 2 Model:
$$\begin{array}{l}\beta _{0{{{\mathrm{i}}}}} = \gamma _{00} + \gamma _{01}\left( {{{{\mathrm{sex}}}}} \right) + \gamma _{02}\left( {{{{\mathrm{minority}}}}\,{{{\mathrm{status}}}}} \right) \\\qquad+\, \gamma _{03}\left( {{{{\mathrm{mother}}}}\,{{{\mathrm{education}}}}} \right){{{\mathrm{ + U}}}}_{0{{{\mathrm{i}}}}}\end{array}$$
$$\begin{array}{l}\beta _{1{{{\mathrm{i}}}}} = \gamma _{10} + \gamma _{11}\left( {{{{\mathrm{sex}}}}} \right) + \gamma _{12}\left( {{{{\mathrm{minority}}}}\,{{{\mathrm{status}}}}} \right)\\ \qquad +\, \gamma _{13}\left( {{{{\mathrm{mother}}}}\,{{{\mathrm{education}}}}} \right){{{\mathrm{ + U}}}}_{1{{{\mathrm{i}}}}}\end{array}$$
Interpreted, the Level 2 equation now specified adolescents’ intercepts and slopes as a function of the sample average (γ00 and γ10 respectively); their gender (γ01 and γ11); their ethnic/racial minority status (γ02 and γ12); their social class (γ03 and γ13); and residual between-person variance (U0i AND U1i). That is, variability in the intercepts and slopes of adolescents’ masculinity over time were predicted by gender, ethnic/racial minority status, and social class.
In a final step, peer interaction variables were entered as predictors of cross-time trajectories and occasion-specific fluctuations in masculinity. Specifically, friend support, negative peer treatment, and homophobic name-calling were included as Level 1, time-varying predictors of masculinity. Then, the cross-time averages of these same variables were included as Level 2 predictors of the intercept and slope.
Level 1 Model:
$$\begin{array}{ll}{{{\mathrm{Masc}}}}_{{{{\mathrm{ij}}}}} = \beta _{0{{{\mathrm{i}}}}} + \beta _{1{{{\mathrm{i}}}}}\left( {{{{\mathrm{age}}}}} \right) + \beta _{2{{{\mathrm{i}}}}}\left( {{{{\mathrm{friend}}}}\,{{{\mathrm{sup}}}}.} \right) \\\qquad\qquad+\, \beta _{3{{{\mathrm{i}}}}}\left( {{{{\mathrm{negative}}}}\,{{{\mathrm{treat}}}}.} \right) + \beta _{1{{{\mathrm{i}}}}}\left( {{{{\mathrm{homophobic}}}}\,{{{\mathrm{vict}}}}.} \right) + \varepsilon _{{{{\mathrm{ij}}}}}\end{array}$$
Level 2 Model:
$$\begin{array}{l}\beta _{0{{{\mathrm{i}}}}} = \gamma _{00} + \gamma _{01}\left( {{{{\mathrm{sex}}}}} \right) + \gamma _{02}\left( {{{{\mathrm{minority}}}}\,{{{\mathrm{status}}}}} \right)\\\qquad +\, \gamma _{03}\left( {{{{\mathrm{social}}}}\,{{{\mathrm{class}}}}} \right) + \gamma _{04}(\overline {friend\,{\it{sup}}} ) + \gamma _{05}(\overline {negative\,treat} )\\\qquad +\, \gamma _{06}(\overline {{\it{hom}}ophobic\,vict} ) + U_{0{{{\mathrm{i}}}}}\end{array}$$
$$\begin{array}{l}\beta _{1{{{\mathrm{i}}}}} = \gamma _{10} + \gamma _{11}\left( {{{{\mathrm{sex}}}}} \right) + \gamma _{12}\left( {{{{\mathrm{minority}}}}\,{{{\mathrm{status}}}}} \right)\\ \qquad+\, \gamma _{13}\left( {{{{\mathrm{social}}}}\,{{{\mathrm{class}}}}} \right) + \gamma _{14}(\overline {friend\,{\it{sup}}} ) + \gamma _{15}(\overline {negative\,treat} )\\\qquad\, + \gamma _{16}(\overline {{\it{hom}}ophobic\,vict} ) + U_{{{{\mathrm{0i}}}}}\end{array}$$
This model built on the models in the prior steps, such that the Level 1 equation now specified the masculinity score of an adolescent (i) at a specific timepoint (j) as a function of their intercept or cross-time average (β0), a slope or effect of their age (β1), their occasion-specific reports of friend support (β2), negative peer treatment (β3), homophobic name-calling (β4), and a within-person residual, ε. Then, at Level 2 the intercept and slope were each expressed as a function of a cross-time average (γ00 and γ10, respectively), the adolescents’ sex (γ01; γ11), ethnic/racial minority status (γ02; γ12), social class (γ03; γ13), and their own cross-time averages of friend support (γ04; γ14), negative peer treatment (γ05; γ15) and homophobic name-calling (γ06; γ16). For example, variability in a participant’s masculinity at a specific wave was predicted to fluctuate alongside their peer experiences; friend support, negative peer treatment, and homophobic name-calling at that same wave (Level 1). In addition, individual differences in masculinity trajectories across time were predicted by individual differences in the cross-time averages of these same peer experiences (Level 2).
To assist in interpretation of the resulting coefficients, the Level 1 predictors were group-mean centered, and the Level 2 predictors were grand-mean centered. For example, a significant effect of negative peer treatment at level 1 would indicate that on occasions in which adolescents experienced more negative peer treatment than their typical, cross-time average, they reported more elevated levels of masculine-typed behavior at that same time point. A significant effect of negative peer treatment at level 2 would indicate a contextual effect, such that adolescents with higher cross-time averages of negative peer treatment report a higher intercept or slope in masculinity, relative to the rest of the sample.
Altogether, this approach disaggregated time-specific and cross-time effects of peer interactions on masculinity scores. Specifically, it allowed for the estimation of (a) average within-person trajectories of masculine-typed behaviors from ages 14–17, (b) individual differences in these trajectories based on gender, ethnic/racial minority status, and social class, and finally (c) whether peer interactions could predict both cross-time trajectories in masculine-typed behavior, as well as time-specific fluctuations in the same. As a follow-up step, gender was included as a Level 2 moderator of these processes to examine if peer interactions were associated with masculinity differently for boys and girls. Analyses were conducted in Mplus v8.5 using full information maximum likelihood to handle cases with missing data (FIML; Enders,
2022).