Read more about: Design of experiments

An experimental design is a plan that specifies which variables to study, at which levels measurements should be taken, and in what proportion experimental units should be allocated to the different combinations of variables. For example, in a clinical trial to find the best dose of a new drug, one has to decide on the number of doses to use, if any control variables should be included and how to allocate individuals to the different doses. Another example is pretesting of items for achievement tests, where the experimental design determines an allocation of items to examinees by matching item difficulties to examinee ability.

An optimal experimental design aims to achieve as high estimation precision as possible, given the statistical model. The optimality is derived with respect to a suitable criterion defined as a function of the model parameter estimators. Deriving optimal designs for nonlinear and generalized linear models is more challenging than for linear models, as these are generally dependent on the unknown true parameters. Approaches to overcome this issue are to consider locally optimal designs, optimal in expectation with respect to a prior distribution (Bayesian designs) and minimax designs constructed to be robust within a specified range of plausible parameter values.

List of recent and ongoing research topics at the department

• Optimal designs for generalized linear models (GLMs) with extensions to generalized linear mixed models (GLMMs) including random effects
• Development of algorithms to construct minimax designs to deal with the parameter dependence issue
• Optimal designs for dose response experiments for Emax models
• Models and designs for a class of experiments studying willingness to pay for non-market goods
• Calibration designs for estimation of item characteristics during pretesting prior to inclusion in operational tests
• Item selection for precise ability estimation in computerized adaptive tests and adaptive learning systems
• Optimal designs for estimation of treatment effects in social networks, while appropriately dealing with network dependencies and interference