A weighted prediction-based selection criterion for response surface designs

The goal of response surface designs is typically to make precise predictions. A commonly used prediction-based design selection criterion isitV-optimality, which seeks designs that minimize the average prediction variance over the entire experimental region. We propose an alternative criterion, which seeks designs that yield small prediction variances particularly in those parts of the experimental region where a response is expected to be interesting, important, or desirable. The new criterion is a weighted V-optimality criterion, which attaches higher weights to areas with such interesting outcomes. The weights in the new criterion are derived from a logistic regression model. We illustrate the value of the new criterion using an example from the automotive industry.