Table A4.

Description of Control Variables and Network Processes Accounted for in the Stochastic Actor-Based Models

Control Variables
Risky Behaviors
Marijuana use in past month	“During the past month, how many times have you smoked marijuana (pot, reefer, weed, blunts)?” We dichotomize the responses as follows: 0=not at all, 1=at least once.
Drinking in past month	“During the past month, how many times have you had beer, wine, wine coolers, or other liquor?” We dichotomize the responses as follows: 0=not at all, 1=at least once.
Delinquency in past year	“During the past 12 months, how many times have you …?” We dichotomize the responses (0=none or 1=at least once) and then sum across the full set of items (0 to 11 delinquent behaviors). “taken something worth less than $25 that didn’t belong to you” “taken something worth $25 or more that didn’t belong to you” “beat up someone or physically fought with someone because they made you angry (other than just playing around)” “purposely damaged or destroyed property that did not belong to you” “broken into or tried to break into a building just for fun or to look around” “thrown objects such as rocks or bottles at people to hurt or scare them” “run away from home” “skipped school or classes without an excuse” “carried a hidden weapon” “avoided paying for things such as movies, rides, food, or computer services” “taken something from a store that you did not pay for?”
Sensation-seeking behavior	“How often do you do the following things?” Original responses range from “never” (1) to “always” (5), and our measure is the mean of these responses (alpha=0.78). “do what feels good, regardless of the consequences” “do something dangerous because someone dared you to do it” “do crazy things just to see the effects on others”
School
Missed school 7+ days past year	“How many days were you absent from school last year?” Options were “None,” “1–2 days,” “3–6 days,” “7–15 days,” and “16 or more days.” We combined the upper two categories so that 0=6 days or less and 1=7 days or more, because most students (70% of observations) missed fewer than seven days.
School attachment Family	“How true is each of the following statements for you?” Options ranged from “Never true” (1) to “Always true” (5), and the final measure is the mean of these items (alpha=0.80). “I like school a lot.” “I try hard in school.” “Grades are very important to me.” “School bores me.” (reverse coded) “I don’t feel like I really belong at school.” (reverse coded) “I feel very close to at least one of my teachers. “I get along well with my teachers.” “I feel that teachers are picking on me.” (reverse coded)
Family
Family relations	Mean composite of three standardized measures: affective qualities, parent-child activities, inductive reasoning. Affective qualities are measured with the item, “During the past month, when you and your [parent] have spent time talking or doing things together, how often…?” Response options range from (1) “always or almost always” to (5) “never or almost never.” Items are repeated and rephrased to refer to each parent’s behaviors toward the child and the child’s behaviors toward each parent (four separate subscales total, each given one-quarter weight). Items include: “did [she/he/you] let [you/her/him] know [she/he/you] really [care/cares] about [you/her/him]” “did [she/he/you] act loving and affectionate towards [you/her/him]” “did [she/he/you] let [you/her/him] know that [she/he/you] [appreciate/appreciates] [you/her/him], [your/her/his] ideas, or the things [you/she/he] [do/does]” Parent-child activities is mean of items from the question, “During the past month, how often did you do these things with your mom or dad?” Response options range from (1) “every day” to (5) “once during the past month.” “work on homework or a school project together” “do something active together, like playing sports, bike riding, exercising, or going for a walk” “talk about what’s going on at school” “work on something together around the house” “discuss what you want to do in the future” “do some other fun activity that you both enjoy” Inductive reasoning is a mean composite. Response options for each item range from (1) “always” to (5) “never.” “My parents give me reasons for their decisions” “My parents ask me what I think before making a decision that affects me” “When I don’t understand why my parents make a rule for me, they explain the reason”
Student Demographics
Male	1=male, 0=female
White	1=white, 0=nonwhite
Free or reduced lunch	“What do you usually do for lunch on school days?” Two of the options included “I receive free lunch from school” and “I buy my lunch at school at a reduced price.” We constructed a single binary variable in which 0=no free or reduced-price lunch and 1=free or reduced-price lunch. Prior research has recommended free or reduced-price lunch only be used as an indicator of disadvantage when alternative measures are not available, because youth who are eligible for subsidized meals do not always apply for them, especially as they get older and it may be more stigmatizing (Entwislea and Astone,1994; Hauser 1994). Therefore, to minimize concerns with inconsistency in responses in later waves, we follow Osgood, Baals, and Ramirez (2018; see also Baals 2018) by relying on a cross-wave measure of the frequency of free or reduced-price lunch (using our binary indicator), based on empirical Bayes shrinkage estimates (Raudenbush and Bryk 2002; Morris 1983). This approach gives greater weight to students who reported receiving subsidized lunch at multiple waves and in later grades (when it is less common). The resulting measure represents the log odds for receiving subsidized lunch across waves, a time-stable proxy of socioeconomic disadvantage.
Network Processes
School Changes
Transition into higher-level school	Accounts for changes in overall rate of friendship choice due to transition from middle school into high school. Estimated with ego parameters but reported with structural parameters because applies to all respondents in a network at a given wave and does not vary between individuals within a network.
Smaller schools merge into larger	Accounts for changes in overall rate of friendship choice due to merging of multiple schools into single school. Estimated with ego parameters but reported with structural parameters because applies to all respondents in a network at a given wave and does not vary between individuals within a network.
Structural Parameters
Reciprocity	Tendency for friendship ties to be reciprocated
Transitive triplets	Nomination of friends of the respondent’s friends
Transitive reciprocated triplets	Interaction between transitive triplets and reciprocity
Indegree popularity (square root)	Tendency for students who are nominated as friends to continue receiving more nominations
Outdegree truncated (1)	Naming of at least one friend
In-in degre^(1/2) assortativity	Tendency for students who are frequently nominated to name others who are also frequently nominated
Outdegree (density)	Overall rate of friendship choice
Friendship rate parameters	Account for the number of opportunities for ties to change in the simulations between each observation. Estimates remain consistent across different model specifications, but are not of substantive interest in our study and are omitted from the results tables.

Notes. PROSPER. The structural parameters derive from a larger effort to assess and improve goodness-of-fit tests for a series of less complex models. This involved comparing goodness-of-fit tests for models with different sets of structural network parameters, including those of prior studies using these data and several parameters recommended by one of the SIENA developers. We determined which configuration produced, on average, the best goodness-of-fit statistics. Overall, different configurations seem to have relatively little effect on estimates of non-structural parameters.