TABLE 3.

Comparison of PK modeling and simulation approaches in increasing order of complexity from top to bottom

Approach	Between-subject variability	Accuracy of predictions	Comments
Naïve pooling	Ignored (i.e., assumed to be zero or very low)	Only mean profiles can be predicted	Can be adequate to simulate mean concentration profiles, if variability is low; yields biased predictions if variability is moderate or large; cannot simulate between-subject variability
Standard two-stage	Often overestimated	Predicted concn range may be too broad	Can be adequate to simulate mean concentration profiles, if variability is low; requires serial sampling, which may be problematic for mouse PK studies
Population modeling (approximate log-likelihood)	Bias can be large for sparse data	Can simulate variability, but may be considerably biased	Can simulate mean concentration profiles and between-subject variability but may yield biased results for sparse data
Population modeling (exact log-likelihood)	Often most suitable choice	Often most reasonable choice	Can simulate mean concentration profiles and between-subject variability with no (or less) bias; can handle complex PK models with multiple dependent variables (e.g., PK, PD, and resistance)
Population modeling (advanced three-stage methods)	Very powerful, can leverage prior information via a Bayesian approach	Can account for uncertainty as well as for between-subject variability	Powerful, but more complex; requires more expertise and modeling time (e.g., for sensitivity analyses)