The fit of the survival models was compared separately for each of the phenotypes using information criteria (Akaike’s Information Criterion, AIC, the standard Baysian Information Criterion, BIC, and an “sample-size adjusted” BIC which corrects for sample size in a milder way than the standard BIC). Note that the BIC has the strongest tendency to favor parsimonious models compared to AIC and saBIC. The survival models differed significantly with respect to model parsimony. For instance, constraining the hazard to be proportional and constant reduces the number of estimated parameters from 126 (M1C1) to 24 (M2C1) and 17 (M3C1), respectively. Lubke and Neale (2006, 2008) have shown that when comparing mixture models more parsimonious models can be preferred incorrectly by the BIC in case of smaller sample sizes, although Nylund et al. (2007) show that the BIC performs well in general mixture models. No specific simulation results are available for survival mixture models with training data.
