Implications of choosing a correlation structure on model selection and parameter estimation.

Abstract

The purpose of this research was to evaluate implications of choosing a statistical or biological correlation structure on model selection and parameter estimation. were performed with and without biological (weight as a common covariate) or statistical (off diagonal element in omega matrix) correlation between clearance (CL) and volume of distribution (Vd). One-compartment model with IV bolus administration was used with 30% interindividual variability (%CV) on CL and Vd. The results were compared for model selection, parameter equivalence, bias, and imprecision. We found that estimation of fixed-effect parameters (CL, Vd) was robust and estimates of random-effect parameters were not influenced by inclusion or exclusion of statistical correlation irrespective of true correlation structure. However, CVCL and CVV were inflated (by 18 - 35%) when true biological correlation was ignored or accounted for by statistical correlation. It is important to note that in spite of the inflated estimates; these values represent the true variability in the simulated dataset, i.e., reflecting the random variance plus the variance associated with weight. Therefore, if statistical correlation was used in absence of true covariate information, the range of individual parameters in future simulations would be similar compared to a model that uses true biological correlation. A true correlation structure is unknown for real life examples; a statistical correlation is a suitable alternative.