Pfizer Rutgers Colloquium
Friday, May. 1, 2009
Fiber Optics Auditorium,
Bush Campus, Rugers University, Piscataway NJ08854
Program
9:30 AM – 10:00 AMRegistration
10:00 AM – 10:05 PMOpening remarks
10:05 AM – 11:05 AM Michael Gaffney, Ph.D. , Pfizer Inc.
Propensity Scores in Observational Studies.
11:05 AM – 12:05 PMJavier Cabrera, Rutgers University
Discovering subgroups with high differential
treatment effect in clinical trial data.
12:05 PM – 1:30 PMLight lunch
Important Note:
*Please register my emailing Marcy .
If you have any questions or need additional information please do not hesitate to contact Marcy.
Parking instructions and directions can be found at
Abstracts
Propensity Scores in Observational Studies.
Michael Gaffney, Ph.D. and Jack Mardekian, Ph.D., Pfizer Inc.
The use of observational databases to compare the safety and efficacy of different treatment modalities is becoming more frequent. The lack of randomization in observational studies likely leads to confounding in estimating the treatment effect. Propensity scores is a method that has been developed to address the problem of confounding in the analysis of observational studies. This talk will use an actual example to illustrate important features of propensity score analysis. Topics to be addressed are the use of propensity scores in a multivariate regression model and as a means of stratification and matching. Principals of developing a propensity score, evaluating the utility of a propensity score and the relationship to multivariate regression will also be addressed.
Discovering subgroups with high differential treatment effect in clinical trial data.
Javier Cabrera, Rutgers University
One of the most important questions in the drug industry is to be able to characterize patients who benefit the most from a drug compared to older or competitive drugs or placebo. In the simplest setting we have two treatments A and B and we would like to discover subgroups of patients who benefit from drug A more than from drug B. We measure this benefit with a statistic that computes the differential treatment effects. Standard data mining techniques have been applied for this purpose (CART, C4.5, Bump Hunting) but they optimize criteria in order to estimate a response mean or proportion but do not perform any treatment comparison. These data mining methods may provide an answer to our question indirectly in some cases and in other cases they do not. I will discuss a new methodology that provides a direct answer to our question using robust criteria that evaluates partitions over multiple populations. The result of our analysis is a list of subsets that represents the characteristics of patients that benefit from drug A more than from drug B and that are significantly more different in the subgroup than in the entire population.
Subgroup discovery usually requires combining data sets from multiple clinical trials since single datasets are usually too small for obtaining statistically significant differences in even smaller subgroups. Therefore we study the possibility of normalizing datasets before combining them.
We will illustrate our method by analyzing a group of clinical studies and characterizing the patients who respond better to one drug over competing treatments.
This work is in collaboration with Frank Caridi, Jose Alvir, Ha Nguyen, Dhammika Amaratunga,David Madigan and Ching-Ray Yu.Prof. William T. Friedewald