1 | Scope model by provisionally selecting a plausible number of classes based on available literature and the structure based on plausible clinical patterns. | Examine linearity of the shape of standardised residual plots for each of the classes in a model with no random effects. |
2 | Refine the model from step 1 to confirm the optimal number of classes, typically testing K=1–7 classes. | Lowest Bayesian information criteria value. |
3 | Refine optimal model structure from fixed through to unrestricted random effects of the model using the favoured K derived in step 2. | |
4 | Run model adequacy assessments as described in online supplementary table S3 including posterior probability of assignments (APPA), odds of correct classification (OCC) and relative entropy. |
APPA: average of maximum probabilities should be greater than 70% for all classes. OCC values greater than 5.0. Relative entropy values greater than 0.5.
|
5 | Investigate graphical presentation |
Plot mean trajectories across time for each class in a single graph. Plot mean trajectories with 95% predictive intervals for each class (one class per graph). Plot individual class ‘spaghetti plots’ across time for a random sample.
|
6 | Run additional tools to assess discrimination including Degrees of separation (DoS) and Elsensohn’s envelope of residuals |
|
7 | Assess for clinical characterisation and plausibility. |
Tabulation of characteristics by latent classes. Are the trajectory patterns clinically meaningful? Perhaps, consider classes with a minimum percentage of the population. Are the trajectory patterns clinically plausible? Concordance of class characteristics with those for other well-established variables.
|
8 | Conduct sensitivity analyses, for example, testing models without complete data at all time points. | General assessment of patterns of trajectories compared with main model. |