Atrazine concentrations in near-surface aquifers: A censored regression approach

In 1991, the U.S. Geological Survey (USGS) conducted a study to investigate the occurrence of atrazine (2-chloro-4-ethylamino-6-isopropylamino-s-triazine) and other agricultural chemicals in near-surface aquifers in the midcontinental USA. Because about 83% of the atrazine concentrations from the USGS study were censored, standard statistical estimation procedures could not be used. To determine factors that affect atrazine concentrations in groundwater while accommodating the high degree of data censoring, Tobit models were used (normal homoscedastic, normal heteroscedastic, lognormal homoscedastic, and lognormal heteroscedastic). Empirical results suggest that the lognormal heteroscedastic Tobit model is the model of choice for this type of study. This model determined the following factors to have the strongest effect on atrazine concentrations in groundwater : percent of pasture within 3.2 km, percent of forest within 3.2 km (2 mi), mean open interval of the well, primary water use of a well, aquifer class (unconsolidated or bedrock), aquifer type (unconfined or confined), existence of a stream within 30 m (100 ft), existence of a stream within 30 m to 0.4 km (0.25 mi), and existence of a stream within 0.4 to 3.2 km. Examining the elasticities of the continuous explanatory factors provides further insight into their effects on atrazine concentrations in groundwater. This study documents a viable statistical method that can be used to accommodate the complicating presence of censored data, a feature that commonly occurs in environmental data.