WATER RESOURCES BULLETIN VOL 27, NO.1 AMERICAN WATER RESOURCES ASSOCIATION FEBRUARY 1991 DISCUSSION' "Instantaneous Peak Flow Estimation Procedures for Newfoundland Streams," by U. S. Panu and D. A. Smith2 L. M. Lye and E. Moore3 Authors Panu and Smith presented a procedure for estimating instantaneous peak flows at ungauged sites for the Island of Newfoundland. The Island was divided into three homogeneous regions, and regres- sion equations relating peak flows to several water- shed characteristics were developed for various return periods for each region. They showed that the proce- dure is better than any other available procedure for the Island of Newfoundland. The procedure presented by Panu and Smith seemed useful at first glance; however, the procedure as presented has several prob- lems from both a physical and statistical point of view. The first problem is with the regression relation- ships between flood flows of various return periods and watershed characteristics. These were assumed o be of the multiplicative, nonlinear type: QT=KT* j a. i= 1 To facilitate the regression computations, the authors pointed out that it is usual practice to lin- earize Equation (1) by applying a logarithmic trans- formation, which leads to: LogQ=LogK+ a.Logp. While this is usual practice, the fitting of the multi- plicative model of Equation (1) using the logarithmic transformation of Equation (2) leads to an equation that has biased partial-regression coefficients and provides biased estimates of QT. In addition, goodness-of-fit statistics, such as R2 for the logarith- mic form in Equation (2) do not reflect the accuracy of predictions made with Equation (1) (McCuen and Snyder, 1986; McCuen et al., 1990; Koch and Smillie, 1986). For example, consider the regression equation developed for the North Region [Equation (4c) in Panu and Smith (1989)] for T = 100 years: Log(Q T = K* T + a Log(DA) ÷ bLog(MAR) where: + c Log(LAT) K*T = —30.2744, a = 0.9076, (1) b = 1.7432, and c = 14.6735. Table 1 shows that LOg(Q) is unbiased as the sum of the errors is zero. However, QT is biased, with a bias of —67.99 m3/s, which means that on average the model overpredicted by about 6.8 m3/s. Because the estimates of QT obtained from Equation (3) are biased, the R2 values given in Table 4 of Panu and (2) Smith (1989) do not reflect the true goodness-of-fit of Equation (1). Furthermore, as pointed out by McCuen et al. (1990), the fitted coefficients of Equation (3) reflect the sensitivity of LOg(Q) to Log(p1) and not the sensitivity of QT to j. The next problem is the use of the Mean Annual Runoff (MAR) parameter in the regression equations. 1Discussion No. 88066D of the Water Resources Bulletin. 2Paper No. 88066 of the Water Resources Bulletin 25(6):1151-1162. 3Respectively, Assistant Professor and Professor, Faculty of Engineering and Applied Science, Memorial University of Newfoundland, A1B 3X5, Canada. 125 WATER RESOURCES BULLETIN