David Edwards

Non-linear normalization and background correction in one-channel cDNA microarray studies

MOTIVATIONnData from one-channel cDNA microarray studies may exhibit poor reproducibility due to spatial heterogeneity, non-linear array-to-array variation and problems in correcting for background. Uncorrected, these phenomena can give rise to misleading conclusions.nnnRESULTSnSpatial heterogeneity may be corrected using two-dimensional loess smoothing (Colantuoni et al., 2002). Non-linear between-array variation may be corrected using an iterative application of one-dimensional loess smoothing. A method for background correction using a smoothing function rather than simple subtraction is described. These techniques promote within-array spatial uniformity and between-array reproducibility. Their application is illustrated using data from a study of the effects of an insulin sensitizer, rosiglitazone, on gene expression in white adipose tissue in diabetic db/db mice. They may also be useful with data from two-channel cDNA microarrays and from oligonucleotide arrays.nnnAVAILABILITYnR functions for the methods described are available on request from the author.

A double-blind, placebo-controlled trial of tiagabine given three-times daily as add-on therapy for refractory partial seizures

Abstract In a multicentre, double-blind, parallel-group, placebo-controlled trial, a three-times daily regimen of tiagabine was evaluated as add-on therapy in 154 adult patients with refractory partial seizures. A total of 77 patients were randomised to treatment in each arm. Tiagabine HCl was titrated from an initial dose of 12–30 mg/day over 4 weeks. During the 12-week fixed-dose period, there was a significant reduction in the median 4-weekly seizure rate for all partial seizures and simple partial seizures ( P P P P

Continuous combined and sequential estradiol and norethindrone acetate treatment of postmenopausal women: Effect on plasma lipoproteins in a two-year placebo-controlled trial

OBJECTIVEnOur purpose was to examine the effects of postmenopausal estrogen therapy supplemented with progestogen on plasma lipoprotein levels.nnnSTUDY DESIGNnOne hundred thirteen women were randomized to receive either placebo or a combination of 17 beta-estradiol and norethindrone acetate administered continuously (Kliogest) or sequentially (Trisequens). Plasma lipoprotein levels were measured at baseline and after 2 years of treatment and compared by analysis of variance.nnnRESULTSnHormone therapy lowered plasma cholesterol levels (p < 0.001) and low-density lipoprotein cholesterol (Kiogest, p < 0.001; Trisequens, p < 0.01), whereas high-density lipoprotein cholesterol levels were unchanged (Trisequens) or reduced (Kliogest, p < 0.01), primarily because of a decrease in the high-density lipoprotein-2 subfraction (p < 0.05). Low-density lipoprotein/high-density lipoprotein cholesterol ratios remained unchanged.nnnCONCLUSIONSnAlthough hormonal replacement therapy with estradiol combined with norethindrone acetate eliminated the increase in high-density lipoprotein cholesterol levels observed with estrogen monotherapy, the reductions in low-density lipoprotein cholesterol concentrations still suggest reduced cardiovascular risk, according to the National Cholesterol Education Program and to recent observations indicating that risk is not necessarily inversely proportional to high-density lipoprotein cholesterol levels.

On model prespecification in confirmatory randomized studies.

Typically, the primary purpose of confirmatory randomized trials, such as drug trials sponsored by the pharmaceutical industry, is to determine whether there is a treatment effect, and if there is, to estimate the size of the effect. For such studies it is accepted practice to prespecify the statistical model to be used in the primary analysis. The reason for this is a concern that if the model were to be chosen on the basis of the data, the model most favourable to the sponsor might be chosen, with consequent inflation of the type I error. The purpose of this article is to show that, in a sense, this concern is needless. It is shown that if the model is chosen in a blinded fashion and randomization-based tests for no treatment effect are used, then the type I error is controlled. A similar technique to derive unbiased estimates of treatment effect is also described. This approach may be of value when there is uncertainty as to the correct model when the study is being planned.

Linkage analysis using loglinear models

Abstract Analysis of linkage between a set of genes or genetic markers is used to study their relative positions on the chromosomes. An exposition of linkage analysis is given in terms of loglinear models. It is shown that under a simplifying assumption (that of no interference) the independence graphs of the appropriate loglinear models consist of simple serial strings, exactly homologous with the chromosomes. Data from an experiment studying barley powdery mildew are analyzed by means of two loglinear model selection procedures, using asymptotic and exact tests respectively. Comparison of these and other approaches is given in a brief discussion.

Model Search: An Overview

Since standard software does not incorporate the full spectrum of model search techniques, more sophisticated methods are not often used in practice. This paper aims to give a rundown of what is possible in this area and indicates some ideas for applications not previously thought of.

Some Computational Aspects of Graphical Model Selection

Two alternative approaches to graphical model selection — stepwise edge elimination, and a so-called fast method proposed by Edwards and Havranek (1985, 1987) — are described and compared. Some emphasis is given to specific non — numerical computational aspects: in particular, an efficient algorithm for the dual representation problem is described. The model selection methods are applied to a contingency table concerning risk factors for coronary heart disease.

Directed Graphs and Their Models

Up to now the focus of this book has been on the undirected graphs and models. This chapter describes a variety of other types of independence graph and their associated models. In common for almost all of these is that some or all of the edges are drawn as arrows, indicating direction of influence or, sometimes, causal direction. The first section treats graphs with arrows only, the so-called DAGs: such graphs have a long history, starting in path analysis (Wright, 1921) and extending to categorical data (Goodman, 1973, and Wermuth and Lauritzen, 1983). The next section describes graphs of more recent origin that have both lines (i.e., undirectional edges) and arrows, the so-called chain graphs. These are appropriate when the variables can be grouped into blocks, so that the variables within a block are not ordered, but there is a clear ordering between the blocks. In the remaining sections we examine more briefly some other types of graphs. These include local independence graphs, which appear to be useful in analyzing stochastic processes; covariance graphs,in which marginal rather than conditional independences are represented; and reciprocal graphs, which capture the independence properties of simultaneous equation systems.

Model Selection and Criticism

In many applications of statistics, little prior knowledge or relevant theory is available, and so model choice becomes an entirely empirical, exploratory process. Three different approaches to model selection are described in the first three sections of this chapter. The first is a stepwise method, which starts from some initial model and successively adds or removes edges until some criterion is fulfilled. The second is a more global search technique proposed by Edwards and Havranek (1985, 1987), which seeks the simplest models consistent with the data. The third method is to select the model that optimizes one of the so-called information criteria (AIC or BIC). In Section 4 a brief comparison of the three approaches is made.