Size but also the number ofNIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author ManuscriptJ Res Adolesc. Author manuscript; available in PMC 2015 June 01.Gordon et al.Pageimputations, which are typically small enough to deviate from the normal distribution (Li, Raghunathan, Rubin, 1991; Rubin, 1996; StataCorp, 2011). In requesting these calculations, we used the vce(cluster) option so that within each replicate data set robust standard errors were calculated to adjust for the clustering of multiple time points within each participant (Johnson Young, 2011; Rubin, 1996; Wooldridge, 2009) prior to combining estimates. We then used a multinomial logit model to test whether the odds of particular combinations of serious delinquency remained significantly higher for active gang members, even after adjusting for covariates. We again calculated these models within each of the 25 replicate data sets, adjusting for clustering of multiple time points within participants and combining the results with Rubin’s rules. The multinomial logit model is similar to a logit (logistic regression) model but allows for more than two outcome categories (Long, 1997; Long Freese, 2003). In the logit model, the probability of success (the category coded one on the outcome variable) and the probability of a failure (the category coded zero on the outcome variable) are complements, and the log of the odds is the outcome, where the odds is the ratio of these two probabilities. Logit coefficients can be exponentiated to interpret results as odds ratios. Implicitly, the failure probability is the reference outcome category for this interpretation. For example, if a Leupeptin (hemisulfate) site dichotomous indicator of active gang membership was the outcome and a dichotomous indicator of Black race was a predictor variable then an odds ratio statistically larger than one would indicate that Black purchase BRDU versus non-Black youth had higher odds of being gang members than being non-gang members. In the multinomial logit model, odds ratios can be calculated for each pair of outcome categories. One outcome category must be explicitly selected as reference for model estimation. Odds ratios for each of the outcome categories versus the reference outcome category can be read from the default output. The remaining odds ratios for other pairs of outcome categories can be calculated by re-estimating the model with another reference category or by using post-estimation commands. We used post-estimation commands in Stata to recover odds ratios for all of the 28 possible contrasts among the eight configurations of serious delinquency defined above. Because of the many individual tests, we first used an omnibus test of the null hypothesis that all of the contrasts were zero versus the alternative that at least one of the contrasts differed significantly from zero (Long, 1997; Long Freese, 2003). This F test has 14 numerator degrees of freedom accounting for the 14 coefficients for the two indicators of the three gang-status categories (never in a gang, ever in a gang but not in the reference period, ever in a gang including in the reference period) across the seven default equations for the eight outcome categories. We used similar models and calculations to examine our second question regarding risk and protective factors for serious delinquency and gang membership. To determine whether covariates were differentially associated with each configuration of serious delinquency among young men who were a.Size but also the number ofNIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author ManuscriptJ Res Adolesc. Author manuscript; available in PMC 2015 June 01.Gordon et al.Pageimputations, which are typically small enough to deviate from the normal distribution (Li, Raghunathan, Rubin, 1991; Rubin, 1996; StataCorp, 2011). In requesting these calculations, we used the vce(cluster) option so that within each replicate data set robust standard errors were calculated to adjust for the clustering of multiple time points within each participant (Johnson Young, 2011; Rubin, 1996; Wooldridge, 2009) prior to combining estimates. We then used a multinomial logit model to test whether the odds of particular combinations of serious delinquency remained significantly higher for active gang members, even after adjusting for covariates. We again calculated these models within each of the 25 replicate data sets, adjusting for clustering of multiple time points within participants and combining the results with Rubin’s rules. The multinomial logit model is similar to a logit (logistic regression) model but allows for more than two outcome categories (Long, 1997; Long Freese, 2003). In the logit model, the probability of success (the category coded one on the outcome variable) and the probability of a failure (the category coded zero on the outcome variable) are complements, and the log of the odds is the outcome, where the odds is the ratio of these two probabilities. Logit coefficients can be exponentiated to interpret results as odds ratios. Implicitly, the failure probability is the reference outcome category for this interpretation. For example, if a dichotomous indicator of active gang membership was the outcome and a dichotomous indicator of Black race was a predictor variable then an odds ratio statistically larger than one would indicate that Black versus non-Black youth had higher odds of being gang members than being non-gang members. In the multinomial logit model, odds ratios can be calculated for each pair of outcome categories. One outcome category must be explicitly selected as reference for model estimation. Odds ratios for each of the outcome categories versus the reference outcome category can be read from the default output. The remaining odds ratios for other pairs of outcome categories can be calculated by re-estimating the model with another reference category or by using post-estimation commands. We used post-estimation commands in Stata to recover odds ratios for all of the 28 possible contrasts among the eight configurations of serious delinquency defined above. Because of the many individual tests, we first used an omnibus test of the null hypothesis that all of the contrasts were zero versus the alternative that at least one of the contrasts differed significantly from zero (Long, 1997; Long Freese, 2003). This F test has 14 numerator degrees of freedom accounting for the 14 coefficients for the two indicators of the three gang-status categories (never in a gang, ever in a gang but not in the reference period, ever in a gang including in the reference period) across the seven default equations for the eight outcome categories. We used similar models and calculations to examine our second question regarding risk and protective factors for serious delinquency and gang membership. To determine whether covariates were differentially associated with each configuration of serious delinquency among young men who were a.
