Substance Use Profiles Identification via LPA

To identify substance use profiles, we used LPA, where all six indicators of substance use (i.e., cigarette, e-cigarette, snus, alcohol, cannabis, and ‘other illicit drug’ use) were included as continuous variables. Following recent guidelines on applying LPA,¹² we ran an iterative process to identify the best profile solutions and their correlates. 1) Profile enumeration phase. We tested four different variance-covariance structures for profiles (i.e., invariant diagonal, varying diagonal, invariant non-diagonal, and varying non-diagonal) for each set of models ranging from 1 to 6 profile solutions. We verified the replication of the best log-likelihood value to avoid local maxima with three different sets of random starting points among each model. 2) Model evaluation. We used the Bayesian information criterion (BIC) and sample size adjusted BIC (SABIC) to assess model fit. We also evaluated log-likelihood-based indices, such as the adjusted Lo-Mendell-Rubin likelihood ratio test (LMR-LRT) and the bootstrapped likelihood ratio test value (BLRT). Statistically significant results indicate that the K profile model fits the data better than the K-1 profile model. However, in studies with large sample sizes, the fit indices may have significant values in all comparisons, even when practical significance is low.¹² 3) Contender model assessment. Once we identified the preferred model (i.e., the contender model), we analysed whether the profiles of the model were identified correctly by calculating the Average Posterior Probability (AvePP) and Odds of Correct Classification (OCC). AvePP closer to 1 and OCC >5 support adequate profile separation and precision. 4) Latent profiles correlates. Finally, we assessed correlates of substance use profiles from the contender model by including potential correlates as predictors of latent profiles in multinomial logistic regressions with the three-step approach in Mplus 8.5.²² We reported odds ratios (ORs) and ORs with adjustments for age, gender, socio-economic status, parental control, and parents' permissiveness with adolescents' alcohol use.
LPA analyses were run with Mplus 8.5.²³ We set the level of significance to p < 0.01. We used full maximum likelihood estimation with robust standard errors to estimate the latent profiles. Moreover, multiple imputations with ten imputation samples were conducted to handle missing data in all logistic regression analyses under the missing at random (MAR) assumption.