Pre-conference courses  

Sunday, 23 August 2009

Missing data are ubiquitous in medical research, and raise particular issues as the validity of any analysis depends on inherently untestable assumptions. The aim of this course is to familiarize participants with these issues, and the statistical methods for missing data. This will include a discussion of recent and potential future developments.

Specifically the course will:

1. Introduce the issues missing data, review common jargon and argue for a principled, systematic approach
2. Review the implications for ITT and 'on-treatment' analysis of clinical trials
3. Introduce multiple imputation and mixed models for the analysis of partially observed data discuss their relative merits
4. Introduce inverse probability weighting and doubly robust estimation for missing data, and contrast these methods with model based approaches
5. Discuss methods for sensitivity analysis
   

The course will have a mix of lectures and practical (paper-based) exercises.

 

Course 2 (full day):
Clinical Trial Methodology
Stephen Senn (University og Glasgow, U.K.)

This course will be based on the author's book by the same name and will take a critical look at various aspects of clinical trial design and analysis, in particular as practiced in the pharmaceutical industry. Following introductory sections on causality in clinical trials and different philosophies of statistics, the topics of allocation, analysis and the use of covariate information will be examined in more depth. As much as anything, the object will be to provoke discussion and re-evaluation of issues that are too often taken for granted.

The goal of this short course is to provide an introduction to statistical methods for analyzing longitudinal data. The main emphasis is on the practical rather than the theoretical aspects of longitudinal analysis. The course begins with a review of established methods for analyzing longitudinal data when the response of interest is continuous. A general introduction to linear mixed effects models for continuous responses is presented. When the response of interest is categorical (e.g., binary or count data), a number of extensions of generalized linear models to longitudinal data have been proposed. We present a broad overview of two main types of models: "marginal models" and "generalized linear mixed models". While both classes of models account for the within-subject correlation among the repeated measures, they differ in approach. Moreover, these two classes of models have regression coefficients with quite distinct interpretations and address somewhat different questions regarding longitudinal change in the response. In this course we highlight the main distinctions between these two types of models and discuss the types of scientific questions addressed by each.

Microarrays are a powerful technology which biological researchers use to profile the expression patterns of genes, tens of thousands of genes at a time. How to properly analyze and interpret the enormous amounts of data this technology generates remains somewhat of a challenge but significant progress is been made. Generally, a multi-faceted approach is likely to be the most effective at extracting reliable information while overcoming the danger of overfitting due to overparametrization. Thus, for a standard well-designed comparative microarray experiment, a fairly typical prescription for determining a gene expression signature could include (1) an individual gene analysis to identify differentially expressed genes using a method that borrows strength across genes to increase efficiency (2) an analysis of gene sets to identify affected biological processes and pathways (3) an ensemble classification procedure to identify similarities and/or dissimilarities amongst the samples and the genes associated with any dissimilarities (4) a procedure to integrate concomitant data to assess concurrence of findings. This course will introduce the issues underlying microarray data analysis and review this multi-faceted approach.