Title: Evaluation of outlier detection algorithms in linear regression for temperature validation
Authors: Mei-Hsien Lee - University of Taipei (Taiwan) [presenting]
Yu-Chung Wei - Feng Chia University (Taiwan)
Abstract: Outlier detection is an important part of data quality control. Meteorological data verification has a significant impact on the future accurate construction of forecasting systems and other related industries. Erroneous observations are detected through the comparison with estimated references. Linear regression model is generally adopted to construct the relationship between observations and references. And then the potential false data points regarded as outliers are identified. Outlier detection methods via linear regression model are evaluated in temperature data from severs instrument stations. For frequentist approaches, studentized residuals are the easy way to detect the erroneous observations. DFFITS and Cooks distance are inappropriate because the data points resulted from extreme climatic rather than false observations are detected. Moreover, Bayesian predictive discordancy test and extreme posterior probabilities of random error with conjugate prior involved station specific information make identify outliers genuinely from the former meteorological dataset. An easy understanding of the statisticians on the application of meteorological verification is provided, as well as a reference for the selection of appropriate statistical models for the calibration of meteorological data in the meteorological field.