The sequential model-based optimal
design of experiments (e.g., A-, D-, and E-optimal design) is a well-known
technique for selecting experimental conditions that lead to informative
data for obtaining reliable parameter estimates and model predictions.
An important computational step for selecting new model-based experiments
is to compute the inverse of the Fisher information matrix (FIM) which may not be invertible. In this study, three different
methodologies for selecting new experiments are compared for situations
where the FIM is noninvertible. The first approach finds
and leaves out problematic parameters that make FIM noninvertible
and then designs experiments using a reduced FIM (LO
approach). The second approach uses a Moore-Penrose pseudoinverse
of the FIM in A-optimal design calculations (PI approach).
The third methodology is an ad hoc approach which
does not require optimization. In this MS approach, the modeler selects
settings at corners of the specified design space. Comparisons are
made using two linear regression models and a nonlinear dynamic model
for production of a pharmaceutical agent. Monte Carlo simulation results
show that experimental settings obtained by LO and PI approaches give
better parameter estimates on average than the MS approach, with the
LO approach giving the best estimates in 20 of 24 linear situations
studied. The LO approach also gives the best parameter estimates on
average for the nonlinear pharmaceutical model.