Abstract: The calcium (Ca) content distribution of soil samples collected in several districts of Belgium (Condroz region) is right skewed and long-tailed resulting in frequent large Ca measurements. Robust statistical procedures based on the normal distribution are inappropriate for this kind of data since these will identify a too large number of data as suspicious. Therefore, heavy tailed or Pareto-type distributions are used to model the Ca content in soil samples taking into account the possible relationship with covariates such as pH. For some districts, the largest Ca observations can even be considered as outliers with respect to the conditional Pareto-type model (given a particular pH level). An automatic procedure is proposed to identify such suspicious Ca measurements. In a preliminary step, the generalized Pareto distribution is fitted locally in order to get an idea about the dependence of the tail of the Ca distribution on the covariate pH. Next, the Burr distribution is extended to a regression model by taking its shape parameter as a function of the covariate pH, using the information obtained in the exploratory nonparametric analysis. Suspicious Ca measurements are now identified automatically on the basis of the Burr conditional quantile function.