Biostatistics Joint Symposium

Directions

• Take I-71 South until Columbus.

• In Columbus take I-270 West.

• Then take 315 South, until you see the exit sign for Lane Avenue/OSU.

• Turn left onto Lane Avenue. The Holiday Inn is a few blocks east on the left.
  (Past Olentangy River Road, but before Tuttle Park Place)

• Parking is available for free in the Holiday Inn Garage

Schedule

This talk will discuss the methodologic importance of rigorous comparative trial design and analysis. Emphasis will be placed on design features that sometimes attract controversy, such as the use of placebos, randomization, and composition of the study cohort. Some contentious analytic methods will be discussed, such as those dealing with treatment non-adherence and the proper representation of evidence. The National Emphysema Treatment Trials is a large surgical trial that presents an interesting context in which to discuss these and other issues. The principles discussed will be extended to other trial designs and settings.

Clinical research encompasses a vast arena that may include pharmaceutical companies, medical device companies, academic research institutions and clinical care institutions. Consequently, the FDA, the NIH and other governing regulatory bodies may enforce compliance regulations for computerized systems used in research. This in turn may enforce compliance regulations in the areas of data management and analysis. As statisticians, we face a number of challenges in this evolving environment:
* Being aware of the regulations
* Knowing the questions to ask investigators and computing centers
* Minimizing the 'rigamarole' and maximizing the 'rigor'
* Participating in the development of the statistical methodology used in the regulations

The problem of missing data is ubiquitous in longitudinal studies. While ignoring or using overly simple methods to handle missing data often leads to invalid estimates and inference, and numerous missing data methods have been proposed in the recent biostatistical literature, a survey of current epidemiological literature indicates that proper statistical methods are rarely applied to deal with missing data. Thus, it is necessary to develop missing data methods with good statistical properties which can be readily implemented. We focus on "informative missingness" in longitudinal data, defined by the following "shared-parameters" model. In the analysis of longitudinal data, mixed effects models are often applied in which random effects are used to represent individual characteristics such as health awareness, and it is often plausible to assume that the missing pattern is related to these individual characteristics. We develop a method based on grouping together subjects with similar missing patterns, and show that it leads to estimators for the regression parameters with desirable robustness and efficiency properties. Our method can be applied to handle both monotone and intermittent missing data, and can be easily extended to semiparametric models where the progression of outcome over time is modeled nonparametrically. The results are illustrated using data from the Wisconsin Diabetes Registry Project, a longitudinal study tracking glycemic control.

The 2002 Biostatistics Joint symposium is funded by the Statistics Department and the Biostatistics Center of The Ohio State University