Doctor of Philosophy / Department of Statistics / Shie-Shien Yang / When the primary research interest is in the marginal dependence between the response and the covariates, logistic GEE (Generalized Estimating Equation) models are often used to analyze clustered binary data. Relative to ordinary logistic regression, very little work has been done to assess the lack of fit of a logistic GEE model. A new method addressing the LOF of a logistic GEE model was proposed. Simulation results indicate the proposed method performs better than or as well as other currently available LOF methods for logistic GEE models. A SAS macro was developed to implement the proposed method. Nonlinear regression models are widely used in medical science. Before the models can be fit and parameters interpreted, researchers need to decide which design points in a prespecified design space should be included in the experiment. Careful choices at this stage will lead to efficient usage of limited resources. We proposed a cost efficient Bayesian optimal design method for nonlinear combinations of parameters in a nonlinear model with quantitative predictors. An R package was developed to implement the proposed method.

This thesis develops methodologies for the construction of various types of optimal designs with applications in maximum likelihood estimation and factorial structure design. The methodologies are applied to some real data sets throughout the thesis. We start with a broad review of optimal design theory including various types of optimal designs along with some fundamental concepts. We then consider a class of optimization problems and determine the optimality conditions. An important tool is the directional derivative of a criterion function. We study extensively the properties of the directional derivatives. In order to determine the optimal designs, we consider a class of multiplicative algorithms indexed by a function, which satisfies certain conditions. The most important and popular design criterion in applications is D-optimality. We construct such designs for various regression models and develop some useful strategies for better convergence of the algorithms. The remaining thesis is devoted to some important applications of optimal design theory. We first consider the problem of determining maximum likelihood estimates of the cell probabilities under the hypothesis of marginal homogeneity in a square contingency table. We formulate the Lagrangian function and remove the Lagrange parameters by substitution. We then transform the problem to one of maximizing some functions of the cell probabilities simultaneously. We apply this problem to some real data sets, namely, a US Migration data, and a data on grading of unaided distance vision. We solve another estimation problem to determine the maximum likelihood estimation of the parameters of the latent variable models such as Bradley-Terry model where the data come from a paired comparisons experiment. We approach this problem by considering the observed frequency having a binomial distribution and then replacing the binomial parameters in terms of optimal design weights. We apply this problem to a data set from American League Baseball Teams. Finally, we construct some optimal structure designs for comparing test treatments with a control. We introduce different structure designs and establish their properties using the incidence and characteristic matrices. We also develop methods of obtaining optimal R-type structure designs and show how such designs are trace, A- and MV-optimal. / October 2016

Mixed integer programming (MIP) problems are highly parameterized, and finding parameter settings that achieve high performance for specific types of MIP instances is challenging. This paper presents a method to find the information about how CPLEX solver parameter settings perform for the different classes of mixed integer linear programs by using designed experiments and statistical models. Fitting a model through design of experiments helps in finding the optimal region across all combinations of parameter settings. The study involves recognizing the best parameter settings that results in the best performance for a specific class of instances. Choosing good setting has a large effect in minimizing the solution time and optimality gap.

PopED is a software developed by the Pharmacometrics Research Group at the Department of Pharmaceutical Biosiences, Uppsala University written mainly in MATLAB. It uses pharmacometric population models to describe the pharmacokinetics and pharmacodynamics of a drug and then estimates an optimal design of a trial for that drug. With optimization calculations in average taking a very long time, it was desirable to increase the calculation speed of the software by parallelizing the serial calculation script. The goal of this project was to investigate different methods of parallelization and implement the method which seemed the best for the circumstances.The parallelization was implemented in C/C++ by using Open MPI and tested on the UPPMAX Kalkyl High-Performance Computation Cluster. Some alterations were made in the original MATLAB script to adapt PopED to the new parallel code. The methods which where parallelized included the Random Search and the Line Search algorithms. The testing showed a significant performance increase, with effectiveness per active core rangingfrom 55% to 89% depending on model and number of evaluated designs.

In this paper we study the loss of Fisher information in approximating a continous distribution by a multinominal distribution coming from a partition of the sample space into a finite number of intervals. We describe and characterize the Fisher information as a function of the partition chosen especially for location parameters. For a small number of intervals the consequences of the choice is demonstrated by instructive examples. For increasing number of individuals we give the asymptotically optimal partition. (author's abstract) / Series: Forschungsberichte / Institut für Statistik

Consider the multiresponse polynomial regression model with one control variable and arbitrary covariance matrix among responses. The present results complement solutions by Krafft and Schaefer (1992) and Imhof (2000), who obtained the n-point D-optimal designs for the multiresponse regression model with one linear and one quadratic. We will show that the D-optimal design is invariant under linear transformation of the control variable. Moreover, the most cases of the exact D-optimal designs on [-1,1] for responses consisting of linear and quadratic polynomials only are derived. The efficiency of the exact D-optimal designs for the univariate quadratic model to that for the above model are also discussed. Some conjectures based on intensively numerical results are also included.

The D-optimal design problems in polynomial regression models with a one-dimensional control variable and k-dimensional response variable Y=(Y_1,...,Y_k) where there are some common unknown parameters are discussed. The approximate D-optimal designs are shown to be independent of the covariance structure between the k responses when the degrees of the k responses are of the same order. Then, the exact n-point D-optimal designs are also discussed. Krafft and Schaefer (1992) and Imhof (2000) are useful in obtaining our results. We extend the proof of symmetric cases for k>= 2.

In this work, we investigate c-optimal design for polynomial regression model without intercept. Huang and Chen (1996) showed that the c-optimal design for the dth degree polynomial with intercept is still the optimal design for the no-intercept model for estimating certain individual coe cients over [−1, 1]. We found the c-optimal designs explicitly for estimating other individual coe cients over [−1, 1], which have not been obtained earlier. For the no-intercept model, it is shown that the support points are scale invariant over [−b, b]. Finally some special cases are discussed for estimating the coe cients of the 2nd degree polynomial without intercept by Elfving theorem over nonsymmetric interval [a, b].

No description available.

This thesis first describes the reverse method of transformer design. An existing magnetic model for full-core shell-type transformers, based on circuit theory, is summarised. A magneto-static finite element model is introduced and two sample transformers are analysed. The magnetic model based on finite element analysis is shown to be more accurate than the model based on circuit theory. Partial-core resonant transformers are then introduced and their characteristics are explained using an equivalent circuit model. A method of measuring the winding inductances under resonant operation is developed and used to investigate the characteristics of two different tuning methods. A finite element model of the partial-core resonant transformer is developed by adopting the model for full-core shell-type transformers. The model results accurately match the measured inductance variation characteristics of three sample transformers and predict the onset of core saturation in both axial-offset and centre-gap arrangements. A new design of partial-core resonant transformer is arrived at, having an alternative core and winding layout, as well as multiple winding taps. The finite element model is extended to accommodate the new design and a framework of analysis tools is developed. A general design methodology for partial-core resonant transformers with fixed inductance is developed. A multiple design method is applied to obtain an optimal design for a given set of specifications and restrictions. The design methodology is then extended to devices with variable inductance. Three design examples of partial-core resonant transformers with variable inductance are presented. In the first two design examples, existing devices are replaced. The new transformer designs are significantly lighter and the saturation effects are removed. The third design example is a kitset for high-voltage testing, with the capability to test any hydro-generator stator in New Zealand. The kitset is built and tested in the laboratory, demonstrating design capability. Other significant test results, for which no models have yet been developed, are also presented. Heating effects in the core are reduced by adopting an alternative core construction method, where the laminations are stacked radially, rather than in the usual parallel direction. The new kitset is yet to be used in the field.