<i>D</i>- and <i>A</i>-Optimal Screening Designs

In practice, optimal screening designs for arbitrary run sizes are traditionally generated using the D-criterion with factor settings fixed at ±1, even when considering continuous factors levels in [−1,1]. This article identifies cases of undesirable estimation variance properties such D-optimal and argues that generally A-optimal tend to push variances closer their minimum possible value. New insights about behavior criteria gained through a study respective coordinate-exchange formulas. The confirms existence comprised only ±1 both main effect interaction models blocked unblocked experiments. Scenarios also identified which manipulation coordinate between [−1,1] leads infinitely many each having different properties. For same conditions, A-criterion is shown have unique value improvement. We compare how Bayesian versions A- D-criteria balance minimization bias. Multiple examples considered various under non-Bayesian D-criteria.