Ich occurred 6 months after screening) as baseline and we use some measures from baseline as covariates. We refer to the first time point where we begin examining gang membership and serious delinquency as the start of our focal decade. The ages of the boys varied at baseline and ranged between ages 5 and 9 for the youngest cohort and between ages 12 and 16 for the oldest (see again Table 1). The focal decade began in 1991 to 1992 for the youngest cohort and 1990 to 1992 for the oldest cohort, when boys were approximately three to four years older than baseline. Our focal decade ended for the youngest cohort in 2000 to 2002 when boys were ages 18 to 22 and for the oldest cohort in 1999 to 2002 when they were ages 24 to 28. At mid-decade, when the crack cocaine epidemic was peaking, the boys in the youngest cohort were ages 12 to 16 and the boys in the oldest cohort were ages 19 to 23. We discuss below how we modeled developmental ages, historical time periods, and cohort. To address missing data, we implemented multiple imputations with the mi suite in Stata 12 to create 25 replicate data sets with chained imputation and to combine estimates with Rubin’s rules (Johnson Young, 2011; Rubin, 1996). We required participation in at least half of the 10 waves of our focal decade; 79 boys who participated in four or fewer waves were excluded, leaving 930 young men (the GS-5816MedChemExpress GS-5816 excluded and included boys did not differ significantly on baseline characteristics, including their own early antisocial order PNPP behavior and delinquent peers; results available from the authors). We appended together the 10 waves of data from our focal decade for each participant; thus, each of the 930 youth had 10 records in the data file producing 9,300 records (or “personperiods”) in total. Each time-constant variable (e.g., race and antisocial behavior at baseline) was repeated on each record. In other words, within each participant, the 10 records had the same value for these time-constant variables. Each time-varying variable (e.g., study wave, calendar year, and youth’s age as well as gang membership status and serious delinquency during the reference period for the current wave) was indicated on the appropriate record. For example, each participant’s first record reflected his reports at the first of the 10 focal waves; and, each participant’s 10 records could have different values for these time-varying variables, reflecting his responses at each of the 10 study waves. As discussed further below, we used robust standard errors to adjust for the clustering of multiple time periods within participants. The sample sizes reported in the tables are smaller than 930 youth and 9,300 person-periods because we excluded those who never reported serious delinquency (i.e., serious violence, serious theft, and drug selling; see below) across the 10 study waves (in order to sharpen theNIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author ManuscriptJ Res Adolesc. Author manuscript; available in PMC 2015 June 01.Gordon et al.Pagecomparison with gang members) and we excluded waves when youth reported being incarcerated (because during these times their deviant behaviors were institutionally restricted). Because both wave-specific serious delinquency and incarceration status were imputed variables, the sample sizes varied across our 25 replicate data sets. Across these replicates, our final samples included 628 to 646 young men (274 to 284 from the youngest cohort and 351 to 363 fr.Ich occurred 6 months after screening) as baseline and we use some measures from baseline as covariates. We refer to the first time point where we begin examining gang membership and serious delinquency as the start of our focal decade. The ages of the boys varied at baseline and ranged between ages 5 and 9 for the youngest cohort and between ages 12 and 16 for the oldest (see again Table 1). The focal decade began in 1991 to 1992 for the youngest cohort and 1990 to 1992 for the oldest cohort, when boys were approximately three to four years older than baseline. Our focal decade ended for the youngest cohort in 2000 to 2002 when boys were ages 18 to 22 and for the oldest cohort in 1999 to 2002 when they were ages 24 to 28. At mid-decade, when the crack cocaine epidemic was peaking, the boys in the youngest cohort were ages 12 to 16 and the boys in the oldest cohort were ages 19 to 23. We discuss below how we modeled developmental ages, historical time periods, and cohort. To address missing data, we implemented multiple imputations with the mi suite in Stata 12 to create 25 replicate data sets with chained imputation and to combine estimates with Rubin’s rules (Johnson Young, 2011; Rubin, 1996). We required participation in at least half of the 10 waves of our focal decade; 79 boys who participated in four or fewer waves were excluded, leaving 930 young men (the excluded and included boys did not differ significantly on baseline characteristics, including their own early antisocial behavior and delinquent peers; results available from the authors). We appended together the 10 waves of data from our focal decade for each participant; thus, each of the 930 youth had 10 records in the data file producing 9,300 records (or “personperiods”) in total. Each time-constant variable (e.g., race and antisocial behavior at baseline) was repeated on each record. In other words, within each participant, the 10 records had the same value for these time-constant variables. Each time-varying variable (e.g., study wave, calendar year, and youth’s age as well as gang membership status and serious delinquency during the reference period for the current wave) was indicated on the appropriate record. For example, each participant’s first record reflected his reports at the first of the 10 focal waves; and, each participant’s 10 records could have different values for these time-varying variables, reflecting his responses at each of the 10 study waves. As discussed further below, we used robust standard errors to adjust for the clustering of multiple time periods within participants. The sample sizes reported in the tables are smaller than 930 youth and 9,300 person-periods because we excluded those who never reported serious delinquency (i.e., serious violence, serious theft, and drug selling; see below) across the 10 study waves (in order to sharpen theNIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author ManuscriptJ Res Adolesc. Author manuscript; available in PMC 2015 June 01.Gordon et al.Pagecomparison with gang members) and we excluded waves when youth reported being incarcerated (because during these times their deviant behaviors were institutionally restricted). Because both wave-specific serious delinquency and incarceration status were imputed variables, the sample sizes varied across our 25 replicate data sets. Across these replicates, our final samples included 628 to 646 young men (274 to 284 from the youngest cohort and 351 to 363 fr.