Annual Report 2008.pdf - SAMSI

Bias Correction in Pharmaceutical Risk-Benefit Assessment
Non-randomized studies, such as prospective patient registries and retrospective administrative
claims analyzes, have great potential for addressing pharmaceutical risk - benefit trade offs. Due
to treatment selection and channeling bias and inclusion of multiple, diverse patient subpopulations,
traditional covariate adjustment methods based upon a single, smooth, global,
parametric model are far from adequate in removing bias from observed associations.
Systematic use of much more local, robust, non-parametric approaches is easily justified when
data are voluminous. I will describe and interrelate some relatively new, alternative methods that
I will also compare and contrast in an invited review paper that I am currently writing for
PharmacoEconomics. I will also discuss an emerging credibility crisis due to conflicts of interest
in the analysis and interpretation of non-randomized health care data and point to a possible
solution.
Gregory Parnell
United States Military Academy, Department of Systems Engineering
gregory.parnell@usma.edu
Decision Analysis Terrorism Risk Analysis
Using bioterrorism as an example, I will use decision analysis to perform a risk assessment. Then
I will show how the decision analysis model can be modified to analyze risk management
options to reduce the risk of bioterrorism.
Shyamal Peddada
NIEHS, Biostatistics Branch
peddada@niehs.nih.gov
Incorporating Historical Control Data When Comparing Tumor Incidence Rates
Laboratories at various pharmaceutical companies, as well as those at federal agencies such as
the National Toxicology Program (NTP), routinely conduct animal carcinogenicity studies to
evaluate carcinogenic effects of chemicals. Since such studies are routinely conducted by these
labs, over time they accumulate data on control animals from multiple studies, thus creating a
historical control database. In any given study, the goal is to detect a dose-related trend in tumor
rates. However, toxicologists are always careful in drawing conclusions about a chemical purely
based on the statistical significance(or lack there of) noted in the given study, since such
conclusions could have far reaching consequences regarding the chemical. For this reason the
toxicologists use other important information to arrive at a conclusion regarding the chemical.
One such type of information is provided by the historical control database. The current practice
is to use this database informally without using a formal decision rule. For example, if the
current control group tumor rate is below the lower limit of historical control range and the
tumor rate of the highest dose is within the historical control range then the statistical
significance noted in the current data is likely to be discounted. For years there has been interest
in developing a formal decision rule that would incorporate the historical control data while