This talk will discuss the differences between the way that a user sees statistical graphics and the way a software developer sees statistical graphics. We will consider reasons for moving users towards the developers view and describe how the grid graphics add-on package for R allows users to make that transition.

The British Columbia Inter-university Research Data Centre provides access
to Statistics Canada's household surveys for faculty and students from SFU,
UBC and UVic. Researchers from agricultural sciences, business
administration, education, health and social sciences have projects at the
RDC using longitudinal complex-design survey microdata. This presentation
will discuss:
a) how to access Statistics Canada data
b) projects currently in progress at the RDC
c) what role can statisticians have in RDC projects

The Research Data Centre does not financially support any projects. However,
we can assist in connecting statisticians with a diverse range of applied
and policy-relevant research projects. The presentation will provide an
overview of the types of problems which statisticians could be involved with
when collaborating or consulting with a RDC project. This work could involve
correcting for response or sample bias, model development and diagnostics,
weighting of pooled records, and variance estimation within a complex survey
design.

I discuss the impact of three principles on the problem of choosing a good goodness-of-fit test. First: when testing statistical hypotheses alternatives of interest are neither indetectably nor grossly different from the null hypothesis. Second, good tests are designed to be sensitive to alternatives likely to arise in practice. Third, the purpose of limit theorems is to provide good approximate probability calculations of interest to statisticians.

I will use Bayesian priors on the alternative hypothesis to construct tests which maximize the expected power for a prior which depends on the sample size. Priors will be presented for which the optimal procedures are (approximately) such goodness-of-fit tests as the Cram'er-von Mises or the Anderson-Darling test.

We propose a discrete-time stochastic volatility model in which regime- switching serves three purposes. First, changes in regimes capture low frequency variations, which is their traditional role. Second, they specify intermediate frequency dynamics that are usually assigned to smooth autoregressive processes. Finally, high frequency switches gen- erate substantial outliers. Thus, a single mechanism captures three important features of the data that are typically addressed as distinct phenomena in the literature. Maximum likelihood estimation is de- veloped and shown to perform well in finite sample. We estimate on exchange rate data a version of the process with four parameters and more than a thousand states. The estimated model compares favor- ably to earlier specifications both in- and out-of-sample. Multifractal forecasts slightly improve on GARCH(1,1) at daily and weekly inter- vals, and provide considerable gains in accuracy at horizons of 10 to 50 days.

Keywords: Forecasting, long memory, Markov regime-switching, max-imum likelihood estimation, scaling, stochastic volatility, time defor-mation, volatility component, Vuong test.

The process of formulation and information encapsulated within social networks result in a form of "relational data". Relational data arise in many social science fields and graph models are a natural approach to representing the structure of these relations. This framework has many applications including, for example, the structure of social and communication networks and the behavior of certain epidemics. We consider statistical and stochastic models for such graphs that can be used to represent the structural characteristics of the networks. In our applications, the nodes usually represent people, and the edges represent a specified relationship between the people. A commonly used model formulation was introduced by Frank and Strauss (1986) and derived from developments in spatial statistics (Besag 1974). These models allow for the potentially complex dependencies within relational data structures.

To date, the use of graph models for networks has been limited by three interrelated factors: the complexity of realistic models, paucity of empirically relevant simulation studies, and a poor understanding of the properties of inferential methods. In this talk we discuss solutions to these limitations. We emphasize the important of likelihood-based inferential procedures and role of Markov Chain Monte Carlo (MCMC) algorithms for simulation and inference.

A primary ongoing issue is the identification of classes of realistic and parsimonious models. In this regard show the unsuitability of some commonly promoted Markov models classes because they can result in degenerate probability distributions. We also consider the suitability and inference for classes of "power law" models that have been proposed for certain random graphs.

The ideas are motivated and illustrated by the study of sexual relations networks with the objective of understanding the social determinants of HIV spread.

Hidden Markov models (HMMs) are a useful tool for capturing the behaviour of overdispersed, autocorrelated data. These models have been applied to many different problems, including speech recognition, precipitation modelling, and gene finding and profiling.

Typically, HMMs are applied to individual stochastic processes; HMMs for simultaneously modelling multiple processes have not been widely studied. In this context, random effects may be a natural choice for capturing differences between processes. In this talk, we develop the theory required for implementing and interpreting these models in a general setting, using the framework of generalized linear mixed models. We discuss parameter estimation and the properties of the resulting estimators, as well as hypothesis tests for the variance components. We then apply these models to two data sets, one relating to lesion counts in multiple sclerosis patients and the other to faecal coliform counts in sea water.

In conducting time series studies to investigate the relationship between air pollution and a health outcome, for example respiratory mortality, it is important to have a good measure of the level of pollution on any particular day. Often daily measurements are available from a number of monitoring sites across the study. Each of these monitors may measure different sets of pollutants, there may be periods of missing data, and all of the recorded measurements will be subject to error. Here, a (Bayesian) hierarchical model is used for the analysis of such data, addressing the issues described, and specifically, allows information from multiple sites on different pollutants to be combined. This allows an estimate of a 'smoothed', or underlying pollution level for each pollutant at each site to be obtained, incorporating any possible lag structure, along with a measure of uncertainty. These modelled levels of pollution can then be used in time series analyses examining the relationship with health outcome. The measure of uncertainty is particularly useful for accounting for the variation in the pollution level, whether informally, when interpreting the regression coefficients, or more formally via error-in-variables modelling. These methods are applied to levels of a number of pollutants, including PM10, CO, NO and SO2, measured at eight sites in London for the period 1993-96. Associations between the resulting modelled levels of pollution and daily mortality counts in London (from 1993-96) are then examined and compared with those obtained using the original pollution measurements. The sensitivity of relative risks and the width of their confidence intervals are examined with respect to model assumptions, with particular interest in the effect of periods of missing data.

This is a joint work with Jon Wakefield.