Predicting Drinking Water Quality Using Low-Cost Sensing: Water Aesthetics and Chemistry Strips

At-home water quality test kits and aesthetic observations can report the quality of drinking water at the tap through low-cost and convenient approaches. These data are collected at low resolution and report high rates of error, however, and are typically used only anecdotally to indicate problems with the quality of drinking water. This research tests the statistical rigor of using low-resolution and low-quality data and explores the use of these data in regression modeling to indicate water quality problems. This research uses a participatory science (citizen science) approach to develop data and uses statistical analysis to test the relationships of building and occupant characteristics, aesthetics, and chemistry (reagent) strips with chemicals present in drinking water. Through an online portal, participants report household and demographic characteristics and the results of chemistry strips. Participants also mail tap water samples to a laboratory for analysis. This research tests the associations of predictor variables, including building and occupant characteristics, aesthetics, and chemistry strips with 20 chemicals present in tap water. Results demonstrated that chemistry strip parameters and aesthetics report significant associations with chemical contaminants. This research also tests the relationships of predictor variables with response variables through backward stepwise linear regression. Results identified that chemistry strip parameters, including alkalinity, hardness, lead, and fluoride, are significantly related to chemicals in drinking water. Other significant predictor variables include well water and metallic taste. The combined use of aesthetic variables, chemistry strip variables, and building characteristics led to models with r2 values in the range of 0.6–0.7 for some chemical constituents. The performance of regression models are compared with the accuracy of chemistry strips for four chemicals. Using low-resolution low-quality data within a regression model substantially improves the mean absolute error and the root-mean square error reported by chemistry strips. Although chemistry strips are subject to high error, chemistry strip data are significantly associated with chemical concentrations in water and may be useful in early warning systems and prioritizing households for further water quality testing.