A1411
Title: Regression-based robust simultaneous prediction intervals
Authors: Christia Tangin - University of the Philippines (Diliman) School of Statistics (Philippines) [presenting]
Michael Daniel Lucagbo - University of the Philippines Diliman (Philippines)
Abstract: The multidimensional nature of some complex medical conditions requires observations on multiple analytes, ensuring accurate assessment of patient health status. Multivariate reference regions (MRR) capture the complex dynamics of these biomarkers while accounting for their correlations. In the laboratory medicine literature, the customary approach to construct MRRs makes use of ellipsoidal regions. However, such regions have a major drawback: their inability to detect component-wise extreme values. This drawback of ellipsoidal MRRs makes them notoriously difficult to interpret. To mitigate this limitation, rectangular MRRs are proposed. Moreover, the expected baseline levels of different analytes in healthy individuals may vary according to specific patient characteristics (i.e., covariates). Information on the covariate values associated with each patient should be incorporated into the computation of the MRR, as these covariates may influence the location, scale, and shape of the resulting probabilistic region. Regression-based simultaneous prediction intervals using biweight estimation are developed to be used as rectangular MRRs. These regions exhibit robust properties, such as insensitivity to the presence of outliers in the dataset. Under the assumption of symmetry, a nonparametric solution is used to estimate the common prediction factor that accounts for the cross-correlations among the analytes.
