Peer Socialization of Aggressive Behavior in Elementary School: Using Propensity Scores to Strengthen Causal Inference

Peer relationships are one of the most proximal developmental contexts for youth, yet causal inferences regarding peer influence are often weak due to the confounding of peer selection (i.e., processes through which individuals choose to interact with other youth) and peer socialization (i.e., process through which an individual’s behavior is shaped by other youth). The goal of this presentation is to apply modern causal inference techniques (Schafer & Kang, 2008) to limit the confounding effects of third variables that may contribute to the selection process (e.g., gender, SES, popularity) by balancing confounders across groups in order to mimic randomization (i.e., randomization to peer groups; Rosenbaum & Rubin, 1983).

Participants are 581 3^rd grade students (46% girls; 52% minority) in 27 classrooms from the Classroom Peer Ecologies Project (CPEP).  In peer sociometric assessments conducted at each of three assessments, each participant named an unlimited number of friends (“who are your friends?”) and nominated an unlimited number of classmates as aggressive (“who starts fights?”). Peer-nominated aggression was calculated as the proportion of classmates who named a child as aggressive. We include 19 potential confounders measured at W1 (first two months of school year).  At W2 (two months later), we calculated the number of reciprocated friends scoring greater than 1 SD above the classroom mean on peer-nominated aggression (M=0.16, SD=0.20):  15% percent of youth had 1 aggressive friend and 7% had 2 or more aggressive friends.  Our outcome variable is W3 (end of school year) peer-nominated aggressive behavior.

            Propensity scores were estimated using multinomial logistic regression.  We computed weights (Hirano & Imbens, 2001) to mimic a randomized sample by evenly distributing confounders across number of aggressive friends.  Checks for balance on each of the 19 confounders revealed small effects (i.e., standardized mean differences between each group on each confounder were < .2; Cohen, 1988).  Next, we estimated the Average Causal Effect (ACE; Rosenbaum & Rubin, 1983) of friends’ aggressive behavior on W3 aggressive behavior using linear regression with our weighted sample.  The ACE was significant such that youth with 0 aggressive friends at W2 displayed lower levels of aggressive behavior at W3 than youth with 1 aggressive friend (B= -0.08, p < .01).  There was no difference in aggressive behavior for youth with 2 or more aggressive friends compared to youth with 1 aggressive friend (B=-0.03, p > .05).

            Final analyses will also include ACE estimates for 1^st and 5^th grades.  Discussion will focus on implications for prevention of aggressive behavior across elementary school.