We approached the analysis in two steps. First, we used LCA to identify latent classes of weekly polydrug use, based on patterns of drug type and route of administration. Second, we used logistic regression to identify demographic characteristics, HIV risk behaviors, and health outcomes that were associated with class membership. In this case, we used a combination of drug type (heroin, methamphetamine, prescription drugs, alcohol, and marijuana) and route of administration (injection, smoking, and swallowing) to define drug use profiles. In order to identify classes based on habitual (vs. episodic) use, we recoded each drug by specific route of administration into a binary variable (1 = used weekly or more frequently, 0 = used less than weekly or never). For example, heroin-injected, heroin-snorted, heroin-smoked were three separate variables coded as “used weekly or more” versus “less than weekly/never.” We then reviewed the distribution of drugs and selected drugs reported by at least 15% of the entire sample for inclusion in the LCA. Based on this standard, seven drug/administration-route combinations were included in the LCA: heroin injection, methamphetamine injection, methamphetamine smoking, methamphetamine snorting, prescription drug swallowing, binge drinking, and marijuana smoking. The prevalence of each drug assessed for inclusion in the model is depicted in appendix 1. Drugs not meeting inclusion criteria included: heroin smoke or snort; cocaine smoke, snort, or injection; simultaneous heroin & cocaine injection; simultaneous methamphetamine & cocaine injection; simultaneous methamphetamine & heroin injection; ketamine injection; oxycontin swallow, snort or injection; and prescription drug smoke, snort or injection.
We then examined models with between 2 and 5 classes. Fit statistics for each model are illustrated in Table 1. Smaller values of Akaike information criteria (AIC) and Bayesian information criteria (BIC) and higher values of entropy indicate better fit. A non-significant bootstrap likelihood-ratio test (LRT) P-value indicates that more classes does not improve the analysis (Gibson, 1959 ; Hagenaars & McCutcheon, 2002 ; McCuthcheon, 1987 ). Thus, we selected a two-class solution based on the goodness-of-fit indices.
After selecting the best fitting model, we used logistic regression to assess factors associated with class membership. Bivariate analyses were first conducted to determine demographic, behavioral, or health status indicators associated with class membership. Factors associated with class membership at the P < .20 level in bivariate analyses were considered for inclusion into a logistic regression model, using a manual backward stepwise approach. Variables achieving significance at the P < .05 level were retained in the final model. Models were checked for meaningful interactions and none were found to be statistically significant. Variables that produced a 10% or greater change between the crude and adjusted odds ratios were considered confounders and were retained in the final model regardless of their significance. All analyses were performed using SAS PROC LCA (Lanza, Dziak, Huang, Wagner, & Collins, 2012 ).