How many different substances are considered in your study and what is your sample size? In principle, it would probably be most informative to consider each substance as a separate latent class indicator if possible.

I would use sex, age, and similar variables as covariates rather than latent class indicators. Sex and age cannot be caused by a latent class variable. Therefore, it seems more logical and more theoretically sound to use them as predictors of class membership (covariates) rather than indicators.

Aleksandar P. Medarevic

Your approach to conducting a Latent Class Analysis (LCA) on poisoning-related deaths is well-justified, as it can help uncover hidden patterns and subgroups based on the presence of substances. Regarding substance grouping, fentanyl definitely deserves to be treated as a separate category due to its high toxicity and the fact that it frequently causes death even without being combined with other substances. Z-drugs, although they have a similar mechanism of action to benzodiazepines, often differ in patterns of use. If they are sufficiently represented in your sample (e.g., over 10%), it may make sense to analyze them separately; otherwise, grouping them with benzodiazepines could be justified. The 5% threshold for including substances is a practical guideline to avoid diluting the model with rare variables, and if most of your substances fall within the 10% to 20% range, that is generally acceptable.

When it comes to sex, age, and F diagnoses, the most common approach is to first define the classes based solely on patterns of substance presence, and then include these demographic and clinical variables as covariates in a subsequent analysis (latent class regression). However, if you believe that, for example, age has a direct influence on poisoning patterns, you could include it as a concomitant variable, although this requires more careful interpretation. Before finalizing your groupings, it may be helpful to conduct an exploratory cluster analysis based solely on the substances to identify natural grouping patterns. Additionally, you can refer to classifications from toxicological or clinical literature, which group substances based on pharmacological action. Finally, it’s important that the identified classes have clinical and public health relevance—consulting with experts in toxicology, psychiatry, or forensic medicine can be helpful for interpreting your results.

Dragan Ugrinov , Christian Geiser , thanks both.

Regarding, Christian´s question. I have 300+ substances. The cohort is pretty big. 6000+ people.

That's a really nice sample size. Nonetheless, having 300+ indicators is still a lot for an LCA, even with several thousand cases. Could you limit yourself to a smaller set of substances that are theoretically/substantively most relevant?

@Christian Geiser, , that is the main issue. how to classified them. for instance, all opioids in one group or fentanyl, bup+met, all other in the third.

Could you restrict the analysis to those substances that are most frequently used in the given sample? For example, exclude those substances that are used by less than 5% or 1% of individuals? Obviously, those are arbitrary cutoffs, but I guess you'll have to reduce the number of indicators somehow to make your analysis feasible.