Nutrient pollution in rivers is a common problem. It can provoke algae blooms which are related to increased fish mortality. To restore the water status, the regulator recently has promulgated more restrictive regulations. In Flanders for instance, the government has introduced several manure decrees (MDs) to restrict nutrient Pollution. Environmental regulations are commonly expressed in terms of threshold levels. This provides a binary response to the decision maker. To handle such data. we propose the use of marginalised generalised linear mixed models. They provide valid inference on trends in the exceedance frequency. The spatio-temporal dependence of the river monitoring network is incorporated by the use of a latent variable. The temporal dependence is assumed to be AR(l) and the spatial dependence is derived from the river topology. The mean model contains a term for the trend and corrects for seasonal variation. The model formulation allows in assessment oil the level of individual sampling locations and oil a more regional scale. The methodology is applied to a case study oil the river Yzer (Flanders). It assesses the impact of the MDs on the violation probability of the nitrate standard. A trend change is detected after the introduction of the second MD. Copyright (C) 2008 John Wiley & Sons, Ltd.