Time：2019/7/2 15:00-16:00

Location：Mingde Main Building 1016

Topic：Acrobatic Regression in Detection and Attribution Analyses of Climate Change

Abstract:

Optimal fingerprinting, the standard approach to detection and attribution of climate change, is a linear regression where the observed climate variable is regressed on the signals of external forcings. The conclusions of detection and attribution analyses depend on the confidence intervals for the regression coefficients. In practice, such analyses are complicated by the facts that the response is spatiotemporally correlated and that the covariates are estimated instead of observed. A weight matrix is needed for efficiency improvement, but the optimal weight or the inverse of the variance matrix has to be estimated as well from independent climate simulations. The higher dimension of the variance matrix brings extra challenges. The standard weighted least squares and weighted total least squares all depend on a good estimate of the variance matrix. Under these complications, the performance of the confidence intervals for the regression coefficients has not been carefully studied. Our investigations reveal that the coverage rates of the confidence intervals can be much lower than their nominal levels. Consequently, misleading conclusions could have been obtained in existing detection and attribution analyses of climate change based on optimal fingerprinting. We propose a simple calibration approach that widens the confidence intervals to match the nominal coverage rate. Our simulation study suggests that when the signals are strong, the conclusions in existing studies would not be affected; when the signals are weak, however, the conclusions in existing studies may need to be revisited because of the widened confidence intervals. We applied the method to detection and attribution analyses of changes in annual mean temperatures in different scale of spatial regions. The confidence intervals from our approach lead to different conclusions than those from the existing approaches in analyses at the subcontinental scale.

Resume：

Dr. Jun Yan is a Professor of Statistics at the University of Connecticut. He received his Ph.D. in Statistics in 2003 from University of Wisconsin - Madison. Before he joined UConn in 2007, he was an Assistant Professor in the Department of Statistics and Actuarial Science, the University of Iowa for four years. His research interests include dynamic survival models, clustered data analysis, spatial statistics, extremes, statistical computing, big data analytics, and applications in public health, economics, and environmental sciences. He is committed to making advanced statistical methods easily accessible to practitioners through open source, quality controlled statistical software packages. He is a Fellow of the American Statistical Association.