Updating urban design floods for changes in central tendency and variability using regression

Abstract

Ordinary least squares (OLS) regression offers a decision-oriented approach for modeling trends in annual peak flows. We introduce a two-stage OLS approach for nonstationary flood frequency analysis that (i) models changes in their central tendency (median) in response to environmental perturbations with one regression and then (ii) examines changes in the coefficient of variation (Cv) by running a second regression on Anscombe-transformed residuals from the first regression. Monte Carlo simulations show that this approach yields 100-year flood estimates with mean squared errors comparable to estimates made with an advanced generalized linear model-based method. Also, this second-stage regression often produces approximately normal residuals, which permits statistical inferences on Cv trends. Case studies illustrate the dramatic impact that decreasing and increasing Cv trends can have on 100-year floods. Findings motivate the incorporation of trends in variability in infrastructure design along with further research examining asymmetric changes in urban flood variability.