Statistical Estimation and Interpretation of Trends in Water Quality Time Series

Three approaches to trend analysis of water quality time series are discussed: (1) seasonal model, with a test for trend based on ranks of observations, with observations assumed to be m dependent; (2) transfer function noise model, in which covariate series may be included by means of transfer functions, with the remaining noise modeled as a seasonal autoregressive moving average process; and (3) component model, with the noise decomposed into series which describe trends, and irregular and seasonal variation. Models are studied with regards to their ability to include covariate series, possibility of interpretation of trends, treatment of seasonal variation and serial dependence, and robustness for outliers. We regard the component model being the most realistic and the most informative of the three approaches. Models are applied to series of monthly phosphorus concentration in the Ljungbyån River in Southern Sweden