Table 5.

Features that should alert investigators to a potentially poor fitting latent class analysis models and their corresponding trouble-shooting solutions to improve model fit.

Features suggesting poor fitting models	Trouble-shooting
Failure to obtain multiple replications of maximum likelihood	- Increase the number of random starts - Check the scale of the continuous predictor variables are appropriately transformed and uniformly scaled - Check distribution of variables and seek extreme outliers - If models fail to replicate the maximum likelihood consider rejecting the model
Minor perturbation of indicators leading to large changes in the model fit statistics and/or VLMR values	- Check correlation between indicators - Check correlation between indicators within each class - Check if the data transformation/imputation of the continuous indicators has led to extreme scaling of important variables (see Figure 3)
A two class model comprising of a class with less than 15% of the sample or Models comprising three or more classes contain a class(es) with less than 10% of the sample	- Check to ensure that a single indicator is not the pre-dominant determinant of the classes - If a single variable determines the class:  - Check the scale of the continuous predictor variables are appropriately transformed and scaled  - Consider rejecting the model  - Validate the findings in an independent cohort
Models with low entropy	- Assess the quality of the indicators:  - Examine the entropy of individual indicators. The variables may be of insufficient quality to separate the classes  - Consider adding novel, higher quality, indicators to the model

Note: the presented solutions may be helpful in rectifying poor model fit, when interpreting these features, however, it must always be considered that a given population may not have underlying latent classes.