Identification and Estimation of Location and Dispersion Effects in Screening Experiments

• Activity Number: 78
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Kwame Kankam*+ and Jim Rosenberger
• Companies: Statistics Department, Pennsylvania State University and Penn State University
• Keywords: Log-linear dispersion model ; Variance Effects

Abstract

In screening experiments with heteroscedasticity, it is of interest to identify and estimate a subset of a potentially large number of factors that might have an effect on the location and the dispersion of the response. Several well-known approaches first identify and then fit a location model and the residuals from the initial location fit are used to identify the active dispersion effects. These approaches may lead to spurious identification of dispersion effects if the location model is incorrect. Penalized likelihood methods are widely used in the high dimensional setting to identify and estimate location parameters. We propose a method which simultaneously identifies and estimates both the active location and active dispersion effects by penalizing the L1 norm of the mean effects and the dispersion effects. Our method makes use of the fact that the problem consists of two parts, both of which are penalized generalized linear models .Thus, we apply the coordinate descent algorithm for penalized GLMs proposed by Friedman et al (2010). We select the best model using AIC and BIC. We compare our method to several methods that have been proposed in the literature.