Table 4.

Summary of key steps and recommendations when interpreting the final model.

	Step	Description	Recommendation	Presentation
Interpreting Final Model	Convergence	A form of internal validation where the maximum likelihood for each model is generated using random starts.	- Multiple random starts are recommended (minimum 50) to replicate the maximum likelihood at least 20 times. - Increase number of starts with models with increased complexity.(55) - If likelihoods are not replicated evaluate data structure and type. Consider rejecting model if maximum likelihoods are not rejected.	- Confirm that maximum likelihood was replicated at least 20 times for all models in the analysis - Presenting the maximum likelihood is optional as the AIC and BIC are generated using this value
	Classification	Probabilities generated by the model are used to classify each observation to a class.	- Probabilities cut-offs to assign class should be determined a priori. - If the entropy of the model is low with poor class separation, the uncertainty of class membership should be incorporated in the analysis.	- Present the probability distribution of the classes in the optimal model that best describes the population (final model).
	Salsa Effect	This refers to the coercion of classes to fit a population that may not have latent classes.	- Examine the distributions of the indicator variables to see if they suggest a single population has been spread out along a continuum.	- Not applicable.
	Outcome Measures	To demonstrate that the identified classes are of value, certain key variables are shown to differ between the classes.	- A priori key discriminant outcome measures should be described in the analysis plan. - Investigators should be blinded from these outcome measures when determining the best fitting model.	- An a priori analysis plan should describe the likely metric that will be used to determine differences between the latent classes and gauge their clinical utility.