A Prognostat tool was developed to predict survival for palliative care patients. The Cox proportional hazards model was fitted to the survival outcomes incorporating several clinical characteristics together with some demographic variables. The significance of Prognostat components were assessed using Wald tests for the corresponding (log)-hazard ratios. A backwards elimination process was used to formulate a final survival prediction model. Survival nomograms were generated using Harrell’s Design  library for R which incorporates Cox model analyses to compute weights that are applied to the Prognostat components to determine an overall Prognostat Score.

It is common to encounter situations where the choice of an estimator or test is predicated on the outcome of some prior statistical test. When this is the case, the properties of the resulting estimator (or second test) are usually affected in quite complex ways. For example, estimators that are unbiased under standard assumptions may become biased when a preliminary test precedes their selection and application.

"Preliminary test inference" has been explored widely since at least 1944. However, the implications of this important literature are frequently overlooked by practitioners. Recently, there has been a renewed interest in this problem.

In this talk I give a brief overview of some of the issues associated with inference after a preliminary test, and illustrate these issues with some simple examples that have been selected to be accessible to students.

Abstract: Persistent homology is a relatively new effort by algebraic topologists, in a sense it represents the idea that classical homology, due to Poincare over 100 year ago, when married with modern computers should produce novel ways of describing the shape of large, high-dimensional data sets.   This idea has been gaining popularity recently.  I will describe some experiments that William Sethares and I have done, applying persistent homology to musical data.  It seems persistent homology can be useful for discovering some hard-to-see non-linear aspects of data, as well as some rather subtle metrical aspects of data.    There are many foundational questions that are still open in this rather young field, some of which will be mentioned.

The empirical likelihood is a modern non-parametric method of inference with many applications in statistics. It combines modern computing power with the classical asymptotic approach to yield a powerful and versatile inference tool, and it continues to find new applications and surprising new theoretical results.

Empirical likelihood ratio confidence regions are known to suffer from an undercoverage problem in that their observed coverage probabilities tend to be lower than the desired nominal levels. In this talk, I will give a brief introduction to the general idea of the empirical likelihood and review the undercoverage problem described in the previous talk. Then, I will discuss two well-known high order methods for dealing with this problem and talk about my own recent work on a geometric solution for this problem.

I shall talk about a way of adjusting parametric failure-time distributions which allows for much greater flexibility in the shape of the hazard-rate function, including the possibility of a bathtub shape.  The adjustment involves the introduction of one extra parameter, so that the power-law adjusted forms of the common two-parameter distributions such as the Weibull, lognormal, gamma etc. are three-parameter distributions.

To motivate the talk (and to make it last 50 minutes!) I shall trace a path that I followed over perhaps the last ten years of my career to produce this work.  This may give some idea to graduate students of how research can progress.  The results and applications may be far from what one originally anticipated!

Abstract

Mixed models are useful for modelling the dependence among response variables
inherent in repeated measures studies. Classical goodness-of-fit and model checking
methods for such models can be vague due to the complexity of their error structures.
Formal assessment tools available in the literature range from graphical checks of the
distributional assumptions to likelihood based and cross validation tests.
The talk focuses on measures used to check and compare mixed models regardless of
their distributional assumptions. Such methods include concordance relation measures,
penalized likelihood criteria and Bayesian based model checking. Their application is
illustrated with clustered dataset examples containing multiple random effects.