Spatiotemporal Modeling of Distributional Climate Data Using Semiparametric Quantile Functional Regression

• Activity Number: 224 - New Insights from Analyzing Functional Data in Biomedical Research
• Type: Invited
• Date/Time: Wednesday, August 11, 2021 : 10:00 AM to 11:50 AM
• Sponsor: ENAR
• Author(s): Jeffrey S. Morris* and Hojin Yang and Veerabhadran Baladandayuthapani and Birgir Hrafnkelsson
• Companies: University of Pennsylvania Perelman School of Medicine and Pusan National University and University of Michigan and University of Iceland
• Keywords: functional data analysis; quantile regression; spatiotemporal modeling; climate change; kernel methods

Abstract

It is important to characterize climate change spatially and temporally, and mean changes are not sufficient. We present a spatiotemporal quantile functional regression model that represents the distribution of daily average temperatures for a given year as functional data that is spatially correlated across sites and temporally correlated across years via spatiotemporal kernel methods, while adjusting for important covariates including latitude, elevation, and distance from the coastline. This model produces distributional estimates for any location on the map via spatial kriging, smoothly characterizes how these distributions have changed over time at each location via temporal smoothing, and the entire distribution is smoothed via functional data analysis. From the modeling results, we can construct smooth trends and pointwise and joint credible bands for any distributional summary, including the mean, variance, skewness, kurtosis, any quantile, or CDF value such as proportion of time above freezing for each location. We apply this model to summer temperature values from 255 sites in Iceland between 1949-2017 to characterize systematic climate shifts across the country.