Regional regression models of percentile flows for the contiguous United States: Expert versus data-driven independent variable selection

Abstract

Study region918 basins in the contiguous United States. Study focusRegional regression models were developed to predict 13 percentile flows for groups of basins clustered based on physical and climatic characteristics. The research question investigated how the number and information content of independent variables affected model performance, and compared data-driven versus expert assessment approaches for variable selection. New hydrological insights for the regionA set of three variables selected based on an expert assessment of factors that influence percentile flows performed similarly to larger sets of variables selected using a data-driven method. Expert assessment variables included mean annual precipitation, potential evapotranspiration, and baseflow index. Larger sets of up to 37 variables contributed little, if any, additional predictive information. Variables used to describe the distribution of basin data (e.g. standard deviation) were not useful, and average values were sufficient to characterize physical and climatic basin conditions. Effectiveness of the expert assessment variables may be due to the high degree of multicollinearity (i.e. cross-correlation) among additional variables. A tool is provided in the Supplementary material to predict percentile flows based on the three expert assessment variables. Future work should develop new variables with a strong understanding of the processes related to percentile flows.