Student Seminar Series - September 27, 2007
University of Minnesota
School of Statistics
College of Liberal Arts



Principal Fitted Components on Small Samples


Kofi Placid Adragni


Tuesday, September 27, 2007
10 AM,
300 Ford Hall
Minneapolis, East Bank Campus

Refreshments at 9:30 PM
300 Ford Hall


Abstract


In various scientific fields, researchers are confronted with large datasets where there are many measured variables and few observations. With the presence of a measured outcome, possible regressions face a challenge due to the number of predictors exceeding the number of observations. Several methods are available to get round or deal with the challenge. Cook (2007) introduced the use of Principal Components (PC) and Principal Fitted Components (PFC) models in the inverse regression setting as an approach for dimension reduction. The development of the PFC model suggests a large sample (n>p) and also involves the use of basis functions. In the first part of this proposal, we explore various basis functions to be used in the PFC model. Using univariate PFC model, a method called Screening by Principal Fitted Components (SPFC) is derived to aid in selecting predictors related to the outcome. The relationship between the response variable and individual predictors can be complex, not necessarily linear. In the second part, we investigate a modeling scenario where the information on an outcome may accumulate when the number of predictors gets large. An extended PFC model is used. The maximum likelihood estimators of the parameters in the model are derived and the sufficient reduction of the predictors' subspace is obtained.