Evaluating regression techniques for sodis: Weighted vs. ordinary least squares for predicting Escherichia coli inactivation rate

Abstract

This study evaluates regression models for predicting Escherichia coli inactivation during Solar Water Disinfection (SODIS), using key process parameters-UV intensity, water temperature, and turbidity-as predictor variables. The regression dataset was obtained through 5 months of SODIS experiments. Due to heteroskedasticity and outliers in the data, Ordinary Least Squares (OLS) and Weighted Least Squares (WLS) regression techniques were compared. Model performance was evaluated using a Compromise Programming Index (CPI), integrating key model performance metrics. The WLS model, with UV intensity and temperature as predictors, provided the most accurate E. coli inactivation predictions (Adjusted R-square: 0.893 RMSE: 0.369, MAE: 0.282; PRESS: 6.424). However, WLS did not significantly reduce heteroskedasticity or outliers and may increase multicollinearity. These models can assist water and public health practitioners in making informed decisions, optimizing SODIS, and reducing the risks of misapplication of SODIS.