Abstract
Linear parametric models are commonly assumed and used for unknown data-generating mechanisms. This study demonstrates the value of inferring statistics of meteorological and climatological time series by using a computer-intensive subsampling method that allows one to avoid time series analysis anchored in parametric models with imposed perceived physical assumptions. A first-order autoregressive model, typically adopted as the default model for correlated time series in climate studies, has been selected and altered with a nonlinear component to provide insight into possible errors in estimation due to nonlinearities in the real data-generating mechanism. The nonlinearity undetected by basic diagnostic procedures is shown to invalidate statistical inference based on the linear model, whereas the inference derived through subsampling remains valid. It is argued that subsampling and other resampling methods are preferable in complex dependent-data situations that are typical for atmospheric and climatic series when the real data-generating mechanism is unknown.
Corresponding author address: Alexander Gluhovsky, Department of Earth and Atmospheric Sciences, Purdue University, West Lafayette, IN 47907. Email: aglu@purdue.edu
