Robustness of Data Augmentation for the Analysis of Censored Spatial Data

• Activity Number: 403
• Type: Contributed
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Brooke Fridley*+
• Companies: University of Wisconsin, La Crosse
• Keywords: censored data ; spatial dependence ; data augmentation ; robustness

Abstract

The analysis of spatially dependent observations, some of which fall below a given detection level, occurs occasionally in environmental studies. A common method used to handle censored data is to replace all censored values with a function of their level of detection. This ad hoc method results in biased parameter estimates and hence incorrect prediction of contamination. Another method would be to integrate the censored observations out of the joint posterior distribution. The problem with this approach is that it often is impossible or very difficult to implement. One solution to eliminate the need to complete the high-dimensional integral is through the use of MCMC, referred to as data augmentation. As data augmentation is conditional on the specified model, investigation into the robustness of the procedure when the model is misspecified is needed. An incorrect model may lead to inaccurate augmentation of the censored data and hence incorrect prediction of contaminated regions. This talk will present results looking at the robustness of data augmentation to model misspecification in the analysis of censored spatial data.