We used multiple imputation algorithms to manage missing data for all potential confounding and exposure variables among all participants, generating 20 imputed datasets (Supplementary Tables S1 and S2 ).
To include illegal drug consumption profiles in our analyses, we conducted an LCA with the aim of categorizing an individual’s drug use during the previous 12 months. We included 14 recreational drugs—nitrates, phosphodiesterase-5 blockers and other erectile-dysfunction medication, natural or synthetic cannabinoids, amphetamines, methamphetamines, mephedrone or other synthetic stimulants (i.e., MDMA or “ecstasy”), GHB/GBL, ketamine, LSD, and cocaine—as observed indicators to identify classes for drug use. We ran the model from 1 to 10 latent classes and eventually chose the optimal number of latent classes after considering the following indicators: the lowest value of the adjusted Bayesian information criterion (aBIC), the consistent Akaike information criterion (CAIC), and the entropy index (values close to 0.80) and interpretability and clinical criteria [44 (link)].
We then performed a descriptive analysis stratifying by outcome (prevalence of depressive symptoms) and by prevalence of depressive symptoms and gender. We used measures of central tendency and dispersion for quantitative variables (median and interquartile range). For categorical variables, we calculated absolute frequencies and percentages. We also used the χ2 test and Mann–Whitney test to assess the association between each exposure independently and the outcome.
Finally, we fitted a multivariable logistic regression model. We used LASSO regression (Least Absolute Shrinkage and Selection Operator) as our variable selection model to avoid overfitting (Supplementary Table S3 ), considering the 20 imputed datasets. Continuous variables were included without further modifications. The odds ratio (OR) of continuous variables represents a change of 1 unit. We fixed gender as a potential confounding variable. We used Rubin’s rules to aggregate the results from the 20 imputed datasets [45 (link)]. The data were analyzed using R version 4.1.0 [46 ].
To include illegal drug consumption profiles in our analyses, we conducted an LCA with the aim of categorizing an individual’s drug use during the previous 12 months. We included 14 recreational drugs—nitrates, phosphodiesterase-5 blockers and other erectile-dysfunction medication, natural or synthetic cannabinoids, amphetamines, methamphetamines, mephedrone or other synthetic stimulants (i.e., MDMA or “ecstasy”), GHB/GBL, ketamine, LSD, and cocaine—as observed indicators to identify classes for drug use. We ran the model from 1 to 10 latent classes and eventually chose the optimal number of latent classes after considering the following indicators: the lowest value of the adjusted Bayesian information criterion (aBIC), the consistent Akaike information criterion (CAIC), and the entropy index (values close to 0.80) and interpretability and clinical criteria [44 (link)].
We then performed a descriptive analysis stratifying by outcome (prevalence of depressive symptoms) and by prevalence of depressive symptoms and gender. We used measures of central tendency and dispersion for quantitative variables (median and interquartile range). For categorical variables, we calculated absolute frequencies and percentages. We also used the χ2 test and Mann–Whitney test to assess the association between each exposure independently and the outcome.
Finally, we fitted a multivariable logistic regression model. We used LASSO regression (Least Absolute Shrinkage and Selection Operator) as our variable selection model to avoid overfitting (
Free full text: Click here
