The speaker will, in particular, highlight some recent attempts to address the issue of improving productivity of "creative" work such as book writing, examination marking, and software engineering.

In previous works, the transition rate into the preclinical detectable phase is assumed to be constant over time for all ages. This does not lead to an increasing incidence of breast cancer with age but observations in several populations indicate that incidence increases at approximately the third power of age. This relationship is induced in the model by introducing a transition rate that is age dependent.

The presence of binary covariates requires some elements of a covariance matrix to be fixed. We develop a general approach for sampling such a constrained covariance matrix. The Bayesian inference in this context now demands the use of a Gibbs sampling algorithm.

[Joint work with Larry Weldon, Simon Fraser University]

The algorithm is based upon a generalized Hidden Markov model, an approach used by Burge and Karlin (1997) and others. This general approach combines several kinds of information to predict genetic structure. The two most important sources of information are models for coding and noncoding regions, and models for functional sites (translation initiation sites, translation termination sites, 5' splice sites, and 3' splice sites).

Log-linear models are used to estimate conditional hexamer probabilities, give C+G content, which are applied to calculate likelihoods for coding region and noncoding region models. A novel pattern recognition technique, called reference point logistic classification, is used to estimate functional site probabilities. The likelihood scores and functional site probabilities are combined with additional information to determine a score for each possible parse of a sequence. The parse maximizing the combined score is obtained via a dynamic programming algorithm.

Our answer to this question uses the Akaike entropy maximization criterion to find the relevance weights empirically. Although the problem of estimating (theta)₁ under these conditions has a long history, our estimator appears to be new. Unlike the MLE it is continuously differentiable. Unlike the Pitman estimator for this problem, but like the MLE, it has a simple form.

We will describe the derivation of our estimator, present some of its properties and compare it with some obvious competitors. Finally we present a number of open questions.

For completeness we will review the idea of the relevance weighted likelihood and give illustrative applications in addition to those described above that are central to the talk. In particular, we will show how to forecast the number of goals in NHL hockey as an illustrative digression.

The method performs well in comparison with 22 other classification methods on ten data sets from the European (ESPRIT) project StatLog. Examples include letter recognition and DNA sequence analysis.

I will present specific models from the statistical literature, discuss the computations for the models, give examples where I have used the models, and mention some open research problems.

References:

1. Joe, H. (1997). Multivariate Models and Dependence Concepts, Chapman & Hall.

2. Xu, J.J. (1996). Statistical Modelling and Inference for Multivariate and Longitudinal Discrete Response Data.

3. Molenberghs, G. and Lesaffre, E. (1994). Marginal modelling of correlated ordinal data using a multivariate Plackett distribution. J. Amer. Statist. Assoc., 89, 633-644.

Note: The speaker gives this talk in hope that someone will lead a follow-up journal club session based on

Glonek, G.F.V. (1996). A class of regression models for multivariate categorical responses. Biometrika v. 83, pp. 15-28.

We consider the ``modified cross-validation'' (also known as Leave-2l+1 out) and the ``blockwise cross-validation'' smoothing parameter selection techniques which were initially proposed by Wehrly and Hart (1988) and Hardle and Vieu (1992) respectively. These two smoothing parameter selection techniques take correlation into account and alleviate its effect on the regression function estimation.

We use a simulation study to evaluate the performance of our twosmoothing parameter selection techniques. We compare the results with a few commonly used parametric techniques. Our techniques are also applied to an air pollution data set where we estimate the underlying trend of daily and monthly ground ozone levels in southern Ontario.

Our aim is to do multivariate analysis for non-normal data, particularly data from the multivariate gamma distribution which is an example of a multivariate proper dispersion model, introduced by Jorgensen and Lauritzen (1998). This class provides a multivariate extension of the dispersion model density, following the spirit of the multivariate normal density.

We consider the saddlepoint approximation for small dispersion matrices, which, in turn, implies that the multivariate proper dispersion model is approximately multivariate normal for small dispersion matrices.

We want to mimic the basic technique of testing in multivariate normal, Hotelling's T-squared. Our version of the T-squared test applies asymptotically, for either small dispersion or large samples.

We also consider estimating the normalizing constant of the bivariate gamma by Monte Carlo simulation and we investigate the marginal density by using numerical integration. We also investigate the distribution of the T-squared statistic by Monte Carlo simulation.