Evaluation of statistical models for predicting Escherichia coli particle attachment in fluvial systems

Modeling surface water Escherichia coli fate and transport requires partitioning E. coli into particle-attached and unattached fractions. Attachment is often assumed to be a constant fraction or is estimated using simple linear models. The objectives of this study were to: (i) develop statistical models for predicting E. coli attachment and virulence marker presence in fluvial systems, and (ii) relate E. coli attachment to a variety of environmental parameters. Stream water samples (n = 60) were collected at four locations in a rural, mixed-use watershed between June and October 2012, with four storm events (>20 mm rainfall) being captured. The percentage of E. coli attached to particles (>5 μm) and the occurrences of virulence markers were modeled using water quality, particle concentration, particle size distribution, hydrology and land use factors as explanatory variables. Three types of statistical models appropriate for highly collinear, multidimensional data were compared: least angle shrinkage and selection operator (LASSO), classification and regression trees using the general, unbiased, interaction detection and estimation (GUIDE) algorithm, and multivariate adaptive regression splines (MARS). All models showed that E. coli particle attachment and the presence of E. coli virulence markers in the attached and unattached states were influenced by a combination of water quality, hydrology, land-use and particle properties. Model performance statistics indicate that MARS models outperform LASSO and GUIDE models for predicting E. coli particle attachment and virulence marker occurrence. Validating the MARS modeling approach in multiple watersheds may allow for the development of a parameterizing model to be included in watershed simulation models.