Drinking water utilities face complex decisions when balancing new and changing regulatory requirements with competing finished water quality objectives. Tools are needed to help utilities better understand treatment plant performance in light of changing regulations. These tools must assess the influent water quality and treatment process performance. This thesis presents results to this end, including characterizing water quality, simulating input uncertainty, and modeling conventional treatment processes. All analyses were conducted using the United States (U.S.) Environmental Protection Agency's Information Collection Rule database.

Select water quality variables in the U.S. were characterized in terms of their variability. Spatial variability was examined using a local polynomial method that allowed for general patterns in drinking water quality variables to emerge. Influent water quality variables examined include total organic carbon (TOC), alkalinity, bromide, turbidity, and total specific ultraviolet absorbance. Finished water TOC, as well as total trihalomethanes and five haloacetic acids concentrations after the maximum detention time in the distribution system, were also chosen to be examined.

Input uncertainty of water quality was quantified. A K-nearest neighbor (K-NN) bootstrap technique was developed to generate ensembles of influent water quality conditioned on a "feature vector" that included annual average concentration and location. The approach was applied to simulate monthly ensembles of TOC, alkalinity, and bromide. The simulations provided a rich variability, captured the historical observations well, and were viewed in light of recent available data.

Statistical models were developed using traditional linear (parametric) and relatively new local polynomial (nonparametric) regression methods. Models were implemented to predict the removal of TOC from raw water by conventional surface water treatment and to track the behavior of pH and alkalinity. All models were evaluated in terms of their fit and predictive capability, and for all variables explored, the nonparametric models outperformed their parametric counterparts. Finally, input uncertainty was incorporated into the TOC model to see output scenarios and the probability of exceeding a given limit.