Abstract

Optimal design allows for estimating parameters of statistical models according to important optimality criteria, e. g., minimizing standard errors of estimators. Thus, optimal designs may considerably reduce the number of experimental units, such as respondents or items in empirical studies. For a long time, optimal design has not received much attention within psychology, but meanwhile interest for this subject is rapidly increasing as such designs are needed, e. g., in large scale assessment, adaptive testing or automatic item generation. In this presentation, first, fundamental principles of optimal design are introduced using well-known linear models, e. g. analysis of variance or simple regression. The rationale of adaptive, Bayesian, and minimax designs needed for nonlinear models will then be outlined. Such designs are presented for Item Response Theory (IRT) models, e.g., 1Pl and 2PL model or linear logistic model. Finally, two R packages for deriving Bayesian and minimax designs based on recently developed algorithms will briefly be demonstrated.