In this work minimally supported D-optimal designs for response surface models with spatially correlated errors are studied. The spatially correlated errors describe the correlation between two measurements depending on their distance d through the covariance function C(d)=exp(-rd). In one dimensional design space, the minimally supported D-optimal designs for polynomial models with spatially correlated errors include two end points and are symmetric to the center of the design region. Exact solutions for simple linear and quadratic regression models are presented. For models with third or higher order, numerical solutions are given. While in two dimensional design space, the minimally supported D-optimal designs are invariant under translation¡Brotation and reflection. Numerical results show that a regular triangle on the experimental region of a circle is a minimally supported D-optimal design for the first-order response surface model.

A mixture experiment is an experiment in which the q-ingredients {x_i,i=1,2,...,q} are nonnegative and ubject to the simplex restriction £Ux_i=1 on the (q-1)-dimensional probability simplex S^{q-1}. It is usually assumed that the observations are uncorrelated, although in many applications the observations are correlated. We study the difference between the ordinary least square estimator and the Gauss Markov estimator under correlated observations. It is shown that for certain models and a special covariance structure for the mixture experiments, the unknown parameter vector for the ordinary least square estimators and the Gauss Markov estimators are the same. Moreover, we also show that the corresponding optimal designs may be obtained from previous D- and A-optimal designs for uncorrelated observations. The models studied here includ Scheff'e models, log contrast models, models containing homogeneous functions, and models containing inverse terms.

Optimal design problems aim at selecting design points optimally with respect to certain statistical criteria. The research of this thesis focuses on optimal design problems with respect to A-, D- and E-optimal criteria, which minimize the trace, determinant and largest eigenvalue of the information matrix, respectively. Semide nite programming (SDP) is concerned with optimizing a linear objective function subject to a linear matrix being positive semide nite. Two powerful MATLAB add-ons, SeDuMi and CVX, have been developed to solve SDP problems e ciently. In this paper, we show in detail how to formulate A- and E-optimal design problems as SDP problems and solve them by SeDuMi and CVX. This technique can be used to construct approximate A-optimal and E-optimal designs for all linear and non-linear models with discrete design spaces. The results can also provide guidance to nd optimal designs on continuous design spaces. For one variable polynomial regression models, we solve the A- and E- optimal designs on the continuous design space by using a two-stage procedure. In the rst stage we nd the optimal moments by casting it as an SDP problem and in the second stage we extract the optimal designs from the optimal moments obtained from the rst stage. Unlike E- and A-optimal design problems, the objective function of D-optimal design problem is nonlinear. So D-optimal design problems cannot be reformulated as an SDP. However, it can be cast as a convex problem and solved by an interior point method. In this thesis we give details on how to use the interior point method to solve D-optimal design problems. Finally several numerical examples for A-, D-, and E-optimal designs along with the MATLAB codes are presented. / Graduate

This thesis focuses on the construction of optimum designs for the estimation of the variance components in multistage variability models. Variance components are the model parameters that represent the different sources of variability that affect the response of a system. A general and highly detailed way to define the linear mixed effects model is proposed. The extension considers the explicit definition of all the elements needed to construct a model. One key aspect of this formulation is that the random part is stated as a functional that individually determines the form of the design matrices for each random regressor, which gives significant flexibility. Further, the model is strictly divided into the treatment structure and the variability structure. This allows separate definitions of each structure but using the single rationale of combining, with little restrictions, simple design arrangements called factor layouts. To provide flexibility for considering different models, methodology to find and select optimum designs for variance components is presented using MLE and REML estimators and an alternative method known as the dispersion-mean model. Different forms of information matrices for variance components were obtained. This was mainly done for the cases when the information matrix is a function of the ratios of variances. Closed form expressions for balanced designs for random models with 3-stage variability structure, in crossed and nested layouts were found. The nested case was obtained when the information matrix is a function of the variance components. A general expression for the information matrix for the ratios using REML is presented. An approach to using unbalanced models, which requires the use of general formulae, is discussed. Additionally, D-optimality and A-optimality criteria of design optimality are restated for the case of variance components, and a specific version of pseudo-Bayesian criteria is introduced. Algorithms to construct optimum designs for the variance components based on the aforementioned methodologies were defined. These algorithms have been implemented in the R language. The results are communicated using a simple, but highly informative, graphical approach not seen before in this context. The proposed plots convey enough details for the experimenter to make an informed decision about the design to use in practice. An industrial internship allowed some the results herein to be put into practice, although no new research outcomes originated. Nonetheless, this is evidence of the potential for applications. Equally valuable is the experience of providing statistical advice and reporting conclusions to a non statistical audience.

<p>This thesis proposes a simplification of the model for dependent Bernoulli variables presented in Cox and Snell (1989). The simplified model, referred to as the simplified Cox model, is developed for identically distributed and dependent Bernoulli variables.</p><p>Properties of the model are presented, including expressions for the loglikelihood function and the Fisher information. The special case of a bivariate symmetric model is studied in detail. For this particular model, it is found that the number of design points in a locally D-optimal design is determined by the log-odds ratio between the variables. Under mutual independence, both a general expression for the restrictions of the parameters and an analytical expression for locally D-optimal designs are derived.</p><p>Focusing on the bivariate case, score tests and likelihood ratio tests are derived to test for independence. Numerical illustrations of these test statistics are presented in three examples. In connection to testing for independence, an E-optimal design for maximizing the local asymptotic power of the score test is proposed.</p><p>The simplified Cox model is applied to a dental data. Based on the estimates of the model, optimal designs are derived. The analysis shows that these optimal designs yield considerably more precise parameter estimates compared to the original design. The original design is also compared against the E-optimal design with respect to the power of the score test. For most alternative hypotheses the E-optimal design provides a larger power compared to the original design.</p>

This thesis proposes a simplification of the model for dependent Bernoulli variables presented in Cox and Snell (1989). The simplified model, referred to as the simplified Cox model, is developed for identically distributed and dependent Bernoulli variables. Properties of the model are presented, including expressions for the loglikelihood function and the Fisher information. The special case of a bivariate symmetric model is studied in detail. For this particular model, it is found that the number of design points in a locally D-optimal design is determined by the log-odds ratio between the variables. Under mutual independence, both a general expression for the restrictions of the parameters and an analytical expression for locally D-optimal designs are derived. Focusing on the bivariate case, score tests and likelihood ratio tests are derived to test for independence. Numerical illustrations of these test statistics are presented in three examples. In connection to testing for independence, an E-optimal design for maximizing the local asymptotic power of the score test is proposed. The simplified Cox model is applied to a dental data. Based on the estimates of the model, optimal designs are derived. The analysis shows that these optimal designs yield considerably more precise parameter estimates compared to the original design. The original design is also compared against the E-optimal design with respect to the power of the score test. For most alternative hypotheses the E-optimal design provides a larger power compared to the original design.

Model reliability is important when groundwater models are used for evaluation of environmental impact and water resource management. Model attributes such as geohydrologic units and parameter values need to be quantified in order to obtain reliable results. A primary objective of sampling design for groundwater models is to increase the reliability of modelling results by selecting effective measurement locations and times. It is advantageous to employ simulation models to guide measurement strategies already in early investigation stages. Normally, optimal design is only possible when model attributes are known prior to constructing a design. This is not a meaningful requirement as the model attributes are the final result of the analysis and are not known beforehand. Thus, robust design methods are required that are effective for ranges of parameter values, measurement error types and for alternative conceptual models. Parameter sensitivity is the fundamental model property that is used in this thesis to create effective designs. For conceptual model uncertainty, large-scale sensitivity analysis is used to devise networks that capture sufficient information to determine which model best describes the system with a minimum of measurement points. In fixed conceptual models, effective parameter- and error-robust designs are based on criteria that minimise the size of the parameter covariance matrix (D-optimality). Optimal designs do not necessarily have observations with the highest parameter sensitivities because D-optimality reduces parameter estimation errors by balancing high sensitivity and low correlation between parameters. Ignoring correlation in sparse designs may result in considerably inefficient designs. Different measurement error assumptions may also give widely different optimal designs. Early stage design often involves simple homogenous models for which the design effectiveness may be seriously offset by significant aquifer heterogeneity. Simple automatic and manual methods are possible for design generation. While none of these guarantee globally optimal designs, they do generate designs that are more effective than those normally used for measurement programs. Effective designs are seldom intuitively obvious, indicating that this methodology is quite useful. A general benefit of this type of analysis, in addition to the actual generation of designs, is insight into the relative importance of model attributes and their relation to different measurement strategies.

In this paper we discuss an algorithm for the construction of D-optimal experimental designs for the parameters in a regression model when the errors have a correlation structure. We show that design points can collapse under the presence of some covariance structures and a so called nugget can be employed in a natural way. We also show that the information of equidistant design on covariance parameter is increasing with the number of design points under exponential variogram, however these designs are not D-optimal. Also in higher dimensions the exponential structure without nugget leads to collapsing of the D-optimal design when also parameters of covariance structure are of interest. However, if only trend parameters are of interest, the designs covering uniformly the whole design space are very efficient. For illustration some numerical examples are also included. (author's abstract) / Series: Research Report Series / Department of Statistics and Mathematics

No description available.

本研究旨在建立廣義線性模式下之D-與A-最適設計(optimal designs)，並依不同處理結構(treatment structure)分成完全隨機設計(completely randomized design, CRD)與隨機集區設計(randomized block design, RBD)兩部分探討。 根據完全隨機設計所推導出之行列式的性質與理論結果，我們首先提出一個能快速大幅限縮尋找D-最適正合(exact)設計範圍的演算法。解析解的部分，則從將v個處理的變異數分為兩類出發，建立其D-最適近似(approximate)設計，並由此發現 (1) 各水準對應之樣本最適配置的上下界並非與水準間不同變異有關，而是與有多少處理之變異相同有關；(2) 即使是變異很大的處理，也必須分配觀察值，始能極大化行列式值。此意味著當v較大時，均分應不失為一有效率(efficient)的設計。至於正合設計，我們僅能得出某一處理特別大或特別小時的D-最適設計，並舉例說明求不出一般解的原因。 除此之外，我們亦求出當三個處理的變異數皆不同時之D-最適近似設計，以及v個處理皆不同時之A-最適近似設計。 至於最適隨機集區設計的建立，我們的重點放在v=2及v=3的情形，並假設集區樣本數(block size)為給定。當v=2時，各集區對應之行列式值不受其他集區的影響，故僅需依照完全隨機設計之所得，將各集區之行列式值分別最佳化，即可得出D-與A-最適設計。值得一提的是，若進一步假設各集區中兩處理變異的比例(>1)皆相同，且集區大小皆相同，則將各處理的「近似設計下最適總和」取最接近的整數，再均分給各集區，其結果未必為最適設計。當v=3時，即使只有2個集區，行列式也十分複雜，我們目前僅能證明當集區內各處理的變異相同時(不同集區之處理變異可不同)，均分給定之集區樣本數為D-最適設計。當集區內各處理的變異不全相同時，我們僅能先以2個集區為例，類比完全隨機設計的性質，舉例猜想當兩集區中處理之變異大小順序相同時，各處理最適樣本配置的多寡亦與變異大小呈反比。由於本研究對處理與集區兩者之效應假設為可加，因此可合理假設集區中處理之變異大小順序相同。 / The problem of finding D- and A-optimal designs for the zero- and one-way elimination of heterogeneity under generalized linear models is considered. Since GLM designs rely on the values of parameters to be estimated, our strategy is to employ the locally optimal designs. For the zero-way elimination model, a theorem-based algorithm is proposed to search for the D-optimal exact designs. A formula for the construction of D-optimal approximate design when values of unknown parameters are split into two, with respective sizes m and v-m, are derived. Analytic solutions provided to the exact counterpart, however, are restricted to the cases when m=1 and m=v-1. An example is given to explain the problem involved. On the other hand, the upper bound and lower bound of the optimal number of replicates per treatment are proved dependent on m, rather than the unknown parameters. These bounds imply that designs having as equal number of replications for each treatment as possible are efficient in D-optimality. In addition, a D-optimal approximate design when values of unknown parameters are divided into three groups is also obtained. A closed-form expression for an A-optimal approximate design for comparing arbitrary v treatments is given. For the one-way elimination model, our focus is on studying the D-optimal designs for v=2 and v=3 with each block size given. The D- and A-optimality for v=2 can be achieved by assigning units proportional to square root of the ratio of two variances, which is larger than 1, to the treatment with smaller variance in each block separately. For v=3, the structure of determinant is much more complicated even for two blocks, and we can only show that, when treatment variances are the same within a block, design having equal number of replicates as possible in each block is a D-optimal block design. Some numerical evidences conjecture that a design satisfying the condition that the number of replicates are inversely proportional to the treatment variances per block is better in terms of D-optimality, as long as the ordering of treatment variances are the same across blocks, which is reasonable for an additive model as we assume.