On the Statistical Estimation of Asymmetrical Relationship Between Two Climate Variables

Two simple methods commonly used to detect asymmetry in climate research, composite analysis, and asymmetric linear regression, are discussed and compared using mathematical derivation and synthetic data. Asymmetric regression is shown to provide unbiased estimates only when the respective mean of positive and negative events is removed from both independent and dependent variables (i.e., non-zero y-intercepts). Composite analysis always provides biased results and strongly underestimates the asymmetry, albeit less so for very larger thresholds, which cannot be used with limited observational data. Hence, the unbiased asymmetric regression should be used, even though uncertainties can be large for small samples. Differences in estimated asymmetry are illustrated for the sea surface temperature and winter sea level pressure signals associated with El Nino and La Nina.