Invited speakers.

Slides

The speakers slides are here.

Abstracts

David Hand, "John Nelder: breadth and depth in statistics"

John Nelder made fundamental contributions across a wide range of statistical concepts and methods. In this brief talk, I introduce the man and his statistical achievements.

Peter McCullagh, "Survival models and the pilgrim process"

The pilgrim process is a two-parameter family of exchangeable survival process in the KHC class, named after J. Kalbfleisch, N. Hjort and D. Clayton. The process has several equivalent descriptions. One of these is a Chinese-restaurant-style sequential description, called the pilgrim process, in which the conditional distribution of $T_{n+1}$ given $T_1,\ldots, T_n$ has a close affinity with the Kaplan-Meier survivor function. Even in the presence of right censoring, the joint distribution is available in closed form, which means that the likelihood function can be computed directly without approximation. In the presence of inhomogeneities associated with covariates and unequal hazards, the likelihood is computable, and the parameter estimates can be computed by a simple iterative scheme, not dissimilar to that used for generalized linear models.

Roger Payne, "50 years of general balance"

John Nelder’s work in design was encapsulated in his two Royal Society papers in 1965. These led on to his 1968 paper on combination of information. They were also very influential in Graham Wilkinson’s ANOVA algorithm, which has been a key feature of another of John’s enthusiasms, namely Genstat. In this talk, I shall review the way in which the ideas associated with general balance and ANOVA developed over the years. Topics will include algorithms, efficiency factors, combination of information, simple combinability and variance components. A key objective will be to persuade you that the theory remains as relevant today, as when the papers were first published.

Stephen Senn, "‘Repligate’: reproducibility in statistical studies. What does it mean and in what sense does it matter?"

Various authors have suggested recently, that there is a crisis of reproducibility in science with subsequent studies failing to replicate the results of their predecessors. Many seem to be of the opinion that the cult of significance is particularly to blame and some have suggested that this could be solved by increasing the stringency of the standard for significance. Others have suggested using confidence intervals as a cure. Still others have suggested that Bayesian methods should be substituted instead. However, all these proposals are rather puzzling and raise a number of questions. What is a reasonable probability of a result being reproduced? How is it considered to have been reproduced? Is it only reproducibility of significance that matters or should non-significance also be reproducible? If the common level s of significance are to be made more stringent does this apply to subsequent studies also? Since of any two P-values, provided that nothing else is known, the probability that the first is lower than the second (excluding ties) is 1/2 and vice versa, how would this solve the problem? How big should any replicating study be? Should it be the same size as the original one and if so why? Since no two Bayesians are required to agree with each other in what sense could two Bayesian statements based on independent studies and prior information replicate each other and why should anyone care? On the other hand, if one talks about updating reproducibility, since prior distributions are exchangeable with the data to the degree defined by the model, what relevance does replication have to Bayesian updating?

Yudi Pawitan, "Variable selection and sparse models via random-effects in high-throughput data analyses"

High-throughput data are characterized by a large number of unordered variables measured on a relatively small sample size. Such data are now routinely collected in molecular studies. Sparsity consideration arises naturally as, for example, we expect only a fraction of genes to be involved in any particular biological process. Statistically, imposing sparsity constraint can be expected to lead to better prediction and interpretation. The random-effect approach –- developed and implemented within the h-likelihood framework -- leads to transparent model-based rather than algorithm-based methodologies. I will discuss the key areas covered under this framework, from the classical and structured variable selection problems in regression analyses to sparse version of the classical multivariate analysis techniques, such as the principal component or canonical correlation analyses.

Youngjo Lee, "H-likelihood approach to multiple tests"

Current multiple tests usually begin with summarizing statistics such as p-values or test statistics, ignoring a model selection to find the best fitting model for the basic responses. Furthermore, they are often derived based on independent assumptions among the observations. However, observations in genomics and neuro images are correlated and ignoring such a correlation structure can distort severely conclusions of the test. Moreover, most tests regard misspecifications in the signs of the effects as the power. Thus, it is desirable to modify the current error control to accommodate such a need. We derive an optimal extended likelihood tests under hidden Markov random field models, controlling directional false discovery rates. Real data examples for the gene expression data and neuroimage data show that finding the best-fitting model is crucial for efficiency of test. Proper modeling of correlation structure and model selection tools in the likelihood approach enhance the performance of the test. Reporting estimates of various error rates is useful for the validity of the test.

Rosemary Bailey,"Simple orthogonal block structures, nesting and marginality"

John Nelder introduced simple orthogonal block structures in one of his famous 1965 papers. They provide a compact description of many of the structures in common use in experiments, so much so that some people find it hard to understand a structure that cannot be expressed in this way. Terry Speed and I later generalized them to poset block structures. But there are still misunderstandings. If there are 5 blocks of 4 plots each, should the plot factor have 4 levels or 20? What is the difference between nesting and marginality? What is the difference between a factor, the effect of that factor (this effect may be called an interaction in some cases), and the smallest model which includes that factor whilst respecting marginality? John himself expressed strong views about people who ignored marginality in the model-fitting process. My take on this is that there are three different partial orders involved: I will try to explain the difference.

John Hinde, "GLMs: a transformative paradigm for statistical practice and education"

The ideas that went together to create generalized linear models had been around for some time, but the power of the 1972 paper by Nelder and Wedderburn was to pull all of these together into a coherent modelling framework. The paper itself is well worth a revisit, with a strong emphasis on applications and teaching. The intervening 40+ years have seen many extensions and generalisations to the original model. In this talk I hope to show how glms and some of these extensions have influenced both statistical practice and the general teaching of statistical modelling. For illustration, I will consider topics that have been of particular interest to me, such as overdispersion, random effects, and mixtures.