2.2. Data analysis

LCA uses multiple indicators to divide a population into underlying classes (Collins and Lanza, 2010). Latent class membership probabilities represent the estimated prevalence of classes (here, alcohol use behavior patterns). The estimated probabilities of endorsing particular responses to the indicators, conditional on class membership, are used to interpret the classes. The number of classes was chosen considering both fit criteria and theoretical interpretability; model identification of each candidate model was checked using 100 sets of starting values.

Prevalence rates of the multinomial latent class outcome (i.e., alcohol use behavior patterns) across ages 18 to 65 were allowed to vary flexibly across age as a continuous dimension, without dividing age into fixed categories and without assuming a simple (e.g., linear) shape for the prevalence rate trends, using higher-order polynomial regression, made feasible by the large sample size of the NESARC-III. Age was centered at 40 years to reduce correlation between the polynomial terms. Measurement invariance across ages was assumed to maintain consistent interpretations of the alcohol use behavior patterns. Sampling weights were used in all models to ensure that the age trends were representative of the population of U.S. adults aged 18 to 65. Models were estimated using PROC LCA (Lanza et al., 2015) in SAS 9.4.