(v) Model evaluation.

Models were selected based on feasibility and precision of parameter estimates and goodness-of-fit plots, including the plots of (i) observed concentration versus population predicted concentration, (ii) observed concentration versus individual predicted concentration, (iii) conditional weighted residual (CWRES) versus population predicted concentration, and (iv) CWRES versus time. For a good model, all data points would scatter evenly around the identity line in the former two plots and around the zero line in the latter two plots. Nested models were compared based on the difference in OFV (ΔOFV). ΔOFV was assumed to have a χ² distribution, with the degree of freedom being the difference in the number of parameters between the two nested models. On this basis, the addition of one parameter to the model was considered significantly improved model performance if OFV decreased more than 6.63, corresponding to a P value of <0.01. For nonnested models, AIC was used. The model with a smaller AIC was considered better.

Model predictive performance was evaluated using prediction-corrected visual predictive check of 1,000 simulations. Model predictive performance was acceptable if the observed 5th, 50th, and 95th percentiles fall within the 95% confidence interval of the corresponding simulated percentiles.