The Influence of Dating Relationships on Friendship Networks, Identity Development, and Delinquency

Abstract

Prior research has documented general associations between dating and delinquency, but little is known about the specific ways in which heterosexual experiences influence levels of delinquency involvement and substance use. In the current study, we hypothesize that an adolescent's level of effort and involvement in heterosexual relationships play a significant role in forming the types of friendship networks and views of self that influence the likelihood of delinquency involvement and substance use. Analyses based on a longitudinal sample of adolescent youth (n = 1,090) show that high levels of dating effort and involvement with multiple partners significantly increases unstructured and delinquent peer contacts, and influences self‐views as troublemaker. These broader peer contexts and related self‐views, in turn, mediate the path between dating relationships, self‐reported delinquency, and substance use. Findings also document moderation effects: among those youths who have developed a troublemaker identity and who associate with delinquent peers, dating heightens the risk for delinquent involvement. In contrast, among those individuals who have largely rejected the troublemaker identity and who do not associate with delinquent friends, dating relationships may confer a neutral or even protective benefit. The analyses further explore the role of gender and the delinquency of the romantic partner.

Keywords:

• dating
• delinquency
• peers
• social identity

Notes

1. While our theoretical discussion focuses on the development of heterosexual relationships as an influence on the life‐course of crime and substance use, it could be argued that extensive involvement in dating and juvenile offending are rather manifestations of the same underlying propensity toward risky activities (Gottfredson & Hirschi 1990; Jessor & Jessor, 1977). Accordingly, all of the wave 1 survey items were factor analyzed using a principle iterated extraction method and oblique rotation (Hatcher, 1994). Oblique rotation is preferred here as the factors are assumed to be correlated (Reise, Waller, & Comrey, 2000). Factors with Eigenvalues below one were not extracted. The rotated factor pattern identified two major latent dimensions that explained over 99 percent of the total variance. All delinquency and substance use items and most of the identity items loaded strongly (.48 to .67) on the first dimension, which explains about 64 percent of the total variance. These same items loaded near zero on the second factor. In contrast, all items in the dating effort scale and the dating partner item loaded strongly (.44 to .65) on the second dimension, and close to zero on the first. This second dimension explained an additional 36 percent of the total variance. Only the “partier” identity item loaded on both dimensions in relatively equal proportions. These supplemental analyses support our notion that high levels of dating involvement are conceptually distinct from the other risk‐related behaviors and attitudes we examine in the models described below. Nevertheless, these two factors are positively correlated (r = .28, p < .001 [with the partier item]; r = .27; p < .001 [without the partier item]), consistent with our overall argument that extensive dating involvement heightens the risk for crime and substance use.

2. Estimates for the “rate of change” (dating involvement × time) are explored in models for substance use, peer associations and social identity, however, as the effect size for a rate of change may be relatively small, BIC is used as a guide here for judging whether the a cross‐product represents a general improvement in the model fit (Dempster et al., 1977). Age effects are also explored (dating involvement × age) as the relatively older respondents (i.e., those with adult legal status) may be further along in their dating experience and thus relate differently to involvement in delinquency and substance use.

3. The sampling frame was divided into 18 strata by grade, race/ethnicity, and sex. When students who were initially selected dropped out of the study, the sample was expanded by selecting the “next” unselected student from the same stratum. Sampling weights were calculated based on the inverse probability of selection.

4. Multivariate data formats are converted in person‐period formats by repeating each person‐level id for all variables, across all longitudinal data waves (see Singer, 1998 for more detailed instructions).

5. Logistic regression shows no significant relationship between the 20 (1.8 percent of the analytic sample) deleted cases that failed to answer the delinquency and substance use items and all other measures employed in the current study. Mean imputation is used to fill in the missing observations on the independent variables. Analyses for romantic partner substance use and delinquency employ a slightly smaller sample (1058) which excludes respondents who do not report dating anyone over the course of the study.

6. See Rowe,Vazsonyi, & Figueredo (1997) for a related, albeit differently focused, measure of attitudes and behaviors toward effort and involvement in heterosexual relationships.

7. Several of our predictors have an overdispersed distribution where the variance exceeds the mean by a factor greater than 2.The natural log of these continuous variables is estimated in the multilevel analyses, however, the raw scores are presented in the descriptive results (Table 1). Logging the raw scores normalizes the distribution and reduces the influence that extreme outlying values have on the mean. This transformation also allows for an elasticity interpretation (Woolridge, 2000) for the effects of the logged continuous variables on the log of delinquency; that is a (p) percent increase in (y) for each 1 percent increase in (x). Non‐logged predictors can be exponentiated to recover the percent increase in (y) per unit increase in (x). Further, logging continuous variables in multilevel models helps to ameliorate problems of non‐convergence that arise due to unequal scaling across the variables (Singer and Willett, 2003).

8. The principle iterated extraction method and oblique rotation (SAS “promax”) is preferred here as factors are assumed to be correlated. The troublemaker factor reported Eigenvalues above 1 across all waves.

9. The reporting parent is the biological mother of the adolescent respondent for over eighty percent of the sample. The next largest parent category is the biological father who makes up approximately eight percent of the parent sample.

10. Within‐persons t‐tests are non‐significant for self‐reported delinquency (w2 – w1: t = .43, w3 – w2: t = .97, w3 – w1: t = 1.30, all p>.10) Within‐persons t‐tests confirm significant positive increases in mean scores for self‐reported substance use (w2 – w1: t = 12.10, w3 – w2: t = 13.83, w3 – w1: t = 22.45, all p < .001).

11. In all multivariate analyses, self‐reported delinquency and substance use is matched with its “friend” analog measure. For example, self‐reported substance use (not delinquency) is controlled for in the model that estimates friends' substance use. And in the models for romantic partner delinquency and substance use, matched self‐reported and friend measures are included. Unstructured socializing and troublemaking self‐views are also controlled for in all full model estimates. This modeling strategy recognizes the mutually influential role of peers, identity, and past behavior, while providing a statistically conservative test for the effects of dating involvement in relationship to these factors.

12. For the purpose of conserving journal space, only the effects for the focus variables (dating involvement) are presented in Table 2. In these models, the effects of the control variables on peer associations and the troublemaker identity are similar to results for delinquency and substance use shown in Table 3 (results available by request).

13. The random variance components for intercept and time indicate that, net of controls, there is significant unexplained variability in the initial levels of delinquency and substance use, as well as change in these scores over time. The models' chi‐square indicates that the multilevel estimates are a significant improvement over pooled‐OLS. The decrease in BIC from models 1 to models 2 indicates an improvement in the goodness‐of‐fit. The models' rsquare statistics indicate the total proportion of variance explained between and within‐person levels of delinquency and substance use.

14. Targeted‐centering (DeMaris, 2004) is used to test for significance across levels of the moderator. This method allows SAS to compute the significance test and avoids the otherwise tedious hand computation of covariance algebra that is normally required to evaluate the effects of x on y, across levels of z.