Spatial prediction in the presence of left-censoring |
| |
Affiliation: | 1. Department of Bioengineering, University of California, Riverside, CA 92507, USA;2. Division of Biosciences, The Ohio State University College of Dentistry, Columbus, OH 43210, USA;3. Division of Orthodontics, The Ohio State University College of Dentistry, Columbus, OH 43210, USA;4. Department of Physiology and Cell Biology, The Ohio State University, Columbus, OH 43210, USA;5. Department of Internal Medicine, The Ohio State University College of Medicine, Columbus, OH 43210, USA;6. Department of Biomedical Engineering, The Ohio State University, Columbus, OH 43210, USA;7. Division of Periodontics, The Ohio State University College of Dentistry, Columbus, OH 43210, USA;8. Hormel Institute, University of Minnesota, MN 55901, USA;9. Department of Orthopedics, The Ohio State University College of Medicine, Columbus, OH 43210, USA;10. Department of Physiology, University of Kentucky, Lexington, KY 40536, USA;11. Department of Veterinary and Animal Science, University of Massachusetts-Amherst, MA 01003, USA |
| |
Abstract: | Environmental (spatial) monitoring of different variables often involves left-censored observations falling below the minimum detection limit (MDL) of the instruments used to quantify them. Several methods to predict the variables at new locations given left-censored observations of a stationary spatial process are compared. The methods use versions of kriging predictors, being the best linear unbiased predictors minimizing the mean squared prediction errors. A semi-naive method that determines imputed values at censored locations in an iterative algorithm together with variogram estimation is proposed. It is compared with a computationally intensive method relying on Gaussian assumptions, as well as with two distribution-free methods that impute the MDL or MDL divided by two at the locations with censored values. Their predictive performance is compared in a simulation study for both Gaussian and non-Gaussian processes and discussed in relation to the complexity of the methods from a user’s perspective. The method relying on Gaussian assumptions performs, as expected, best not only for Gaussian processes, but also for other processes with symmetric marginal distributions. Some of the (semi-)naive methods also work well for these cases. For processes with skewed marginal distributions (semi-)naive methods work better. The main differences in predictive performance arise for small true values. For large true values no difference between methods is apparent. |
| |
Keywords: | Kriging Left-censoring Minimum detection limit Prediction Spatial process |
本文献已被 ScienceDirect 等数据库收录! |
|