PLEASE NOTE THAT THIS COURSE HAS ALREADY BEEN COMPLETED.

The Centre hosted this short course for statisticians and epidemiologists in July 2014.

Presenter: Sander Greenland
Professor of Epidemiology, UCLA School of Public Health
Professor of Statistics, UCLA College of Letters and Science

Professor Sander Greenland is one of the most prolific and influential authors on epidemiological methods of the past 2-3 decades, including a co-author (with K Rothman) of the widely prescribed textbook ‘Modern Epidemiology’ and a first author of 177 articles in epidemiology and biostatistics journals.

Date: 24-25 July 2014 (Thursday-Friday)

Supported by: WHO Collaborating Centre for Reference and Research on Influenza, and the Centre for Epidemiology and Biostatistics, Melbourne School of Population and Global Health, The University of Melbourne

Organised by: Sheena Sullivan and Julie Simpson

Location: Peter Doherty Institute for Infection and Immunity, Corner Grattan and Elizabeth Streets, Melbourne, Victoria

Background: Bayesian methods continue to become more popular in statistical modelling, but are not covered in most basic teaching. This lag may in part be due to common misconceptions (encouraged by most expositions) that Bayesian methods are conceptually distinct from frequentist methods and require special software. In fact, Bayesian methods are examples of penalized ("shrinkage") estimation and thus are perfectly acceptable frequentist methods; conversely, common frequentist methods are special types of Bayesian methods in which prior distributions are noninformative (so penalties are either zero or infinite). This short course will explain and illustrate the relationship between the two perspectives with real examples and will show how penalization allows one to deal with a number of common problems that render ordinary statistical methods misleading for epidemiological research.

Thursday 24 July
On day 1, Bayesian and penalized regression methods were introduced, as an alternative to standard frequentist approaches, for analysing data from observational studies in health and social sciences. How to include prior information or suitable shrinkage, without requiring specialist software (such as WinBUGs), was demonstrated, with SAS and Stata coding provided in the computer practicals.

Friday 25 July
On day 2, the methods were extended for more general regression modelling, including hierarchical (multilevel) and bias modelling. These methods provide an alternative to the parsimony-oriented approaches of standard regression analyses. In particular, they replace arbitrary variable-selection criteria by penalized estimation, which has many desirable frequentist properties and which facilitates realistic use of vague but important prior information. The methods facilitate handling problems of sparse data, multiple comparisons, weight stabilization, and sensitivity analysis with multiple bias sources.

Target audience: The course was designed for researchers with previous formal training in epidemiology and multivariable regression methods. During computing sessions the participants were provided with examples of computer code, solutions and assistance from tutors in Stata and SAS only.

For further information about this course please contact Sheena Sullivan by email: whoflu@influenzacentre.org or fax: (03) 9342 9329

Reading

Greenland, S. (2006). Bayesian perspectives for epidemiologic research. I. Foundations and basic methods. Int J Epidemiol, 35: 765-778. comment and reply. Related Stata file

Greenland, S. (2007). Bayesian perspectives for epidemiologic research. II. Regression analysis. Int J Epidemiol, 36: 195-202. Related Stata file

Greenland, S. (2009). Bayesian perspectives for epidemiologic research. III. Bias analysis via missing-data methods. Int J Epidemiol, 38: 1662- 1673. Corrigenda

Sullivan, S., and Greenland, S. (2013). Bayesian regression in SAS software. Int J Epidemiol, 42, 308-317. Letter. Supplementary files

Cole, S., Chu, H., and Greenland, S. (2014) Maximum likelihood, profile likelihood, and penalized likelihood: a primer. Am J Epidemiol, 179: 252-60. Related SAS file

Greenland, S. (2007). Prior data for non-normal priors. Statist Med, 2007, 26: 3578–90

Greenland, S. (2008). Invited Commentary: Variable Selection versus Shrinkage in the Control of Multiple Confounders. Am J Epidemiol. 167: 523-529.

COURSE MATERIALS

Labs
Lab 1 for SAS and Stata
Lab 2 for SAS / Stata
Lab 3 for SAS / Stata

ADO files for Stata users (You will need to right click and save these files to the directory C:\ado\plus\p) pllf.zip