EFFECTS OF OBSERVATION ERRORS IN LINEAR-REGRESSION AND BIN-AVERAGE ANALYSES

Effects of observation errors in linear regression and bin-averaged (B A) validation techniques are investigated using the example of marine wind speeds. It is shown that a conventional linear regression systema tically underestimates the slope of the regression line, and systemati cally overestimates the random model error. A BA analysis systematical ly underestimates extreme wind speeds, incorporates spurious nonlinear ity, and overestimates random model errors. Correction techniques are suggested for studies in which the observation error can be estimated. Using synthetic data the potential of the correction techniques is il lustrated, and it is shown that the above errors are generally not neg ligible for wind speed validation studies. Practical examples consider the random errors of anemometers and wind speed estimates from satell ites. These examples highlight the importance of the error corrections , and illustrate the difficulty of estimating observation errors. Fina lly, it is argued that the well-known symmetric slope regression shoul d not be used for the validation of forecast systems. Although the pre sent study deals with marine wind speeds, its results are expected to be valid for a wide range of validation studies.