WATER RESOURCES RESEARCH, VOL. 29, NO. 6, PAGES 1745-1752, JUNE 1993 L Moment Diagrams Should Replace Product Moment Diagrams RICHARD M. VOGEL aND NEIL M. FENNESSEY Department of Civil and Environmental Engineering, Taftt University, Medford, Massachusetts Its well known that product moment rato estimators of the coefficient of variation C,, skewness ‘y and kurtosis exhibit substantial bias and variance forthe small (n= 100) sampies normally Zhcountered im nyarologte applications. Consequentiy. L moment yao estimators, lermed coe Gient of variation 7», L skewness 7), and L kurtosis 7 are n0W advocated because they are neatly Unbiased for all underlying distnbutions, The advantages of L moment ratio estimators over product moment ratio estimators are not limited 10 small samples. Monte Carlo experiments reveal that product moment estimators of C, and vate also remarkably biased for extremely large samples (n= 100) from highly skewed distributions” A cave study Using large samples n= SO) af average daly Streamflow in Massachusetts reveals thet conventional moment diagrams based on estimates of product moments C,, 7. and «reveal almost no information about the distributional properties of dally Seauilow, wheteas E suvnent diagans based om estinators of 7, 74, and vy etbled us 10 iscriminate among alternate distributional hypotheses. The oldest and most widely understood technique for fitting frequency distributions to observed data is known as, the method of moments. Method of moments estimates of a distriution’s parameters are ohtained hy equating the sao ple moments with theoretical moments, resulting in a system of nonlinear equations which are often easily solved. The primary advantages of method of moments estimators in- ‘clude the ease with which they are computed, their concep- tual simplicity, and the hope that they reproduce the theo- retical moments. Traditionally, the method of moments is applied by equating the theoretical product moments = E(x) (la) Var (X] = E{(X ~ »)7] (6) FLUX - 1)" ] Hw i er FLUX») aa 20 (Qe) Copyright 1993 by the American Geophysical Union, Paper number 93WROO4I (0045-139799393 WR-00381805.00 ea Dur ' where 4, 07, y, and x denote the theoretical mean, variance, skewness, and kurtosis, respectively, and m, s?, G, and k denote the sample mean, variance, skewness. and kurtosis, respectively. It is well known that small sample (n = 100) estimates of s* and G exhibit remarkable bias and variance (Fischer, 1929; Hazen, 1930; Wallis et al., 1974]. Wallis etal. [1974] showed that the sampling properties (bias and vari ance) of these product moment estimators are distribution dependent; hence subsequent efforts to develop unbiased estimators of G, for example, have met with limited success. Efforts to unbias G usually lead to skew estimators with increased variance because such unbiased estimators usually take the form aG and Var (aG) = a? Var (G) with a > 1 because Gis always downward biased. For a recent review of methods for unbiasing sample skewness see Vogel and MeMarun (1991) ‘The remarkable small-sample bias associated with the estimator G, documented by Wallis cr al. [1974], is only the tip of the iceberg. Monte Carlo experiments reported here reveal that both G and the coefficient of variation C, = s/m exhibit even more remarkable bias for highly skewed popu- ations (y > 2) and both small and large sample sizes (n < 5000). These experiments are performed in an effort 10 ‘understand the distributional properties of daily streamflow sequences which often contain thousands of daily flows and ‘exhibit extraordinary skewness. We found product moments, to be of littic value for discriminating among potential candidate distributions of daily streamflows. Another limita- tion of product moment estimators is that they are hounded. For example, Kirby [1974] shows that these bounds depend ‘only upon sample size so that C, is bounded by the interval, (0, (n= 1") and [G| = (1 ~ 2y/(n ~ YE Yet still, product moment estimators are in widespread use, n part due fo their availabilty on most calcufators and in statistical software packages and textbooks, and in part dduc to their familiarity and commonly understood interpre tation. The situation today is quite different from the situa: tion in 1974 when Walle 7 al [1978] questioned the value af product moment estimators. Today a simple and attractive1746 alternative exists: the method of L moments introduced by Hosking (1990]. The method of L moments is an exact analogue to the method of moments. However, L moment estimators are nearly unbiased for all sample sizes and all distributions; hence they should always be preferred in situations wheu uubiased muuients aie required, sucl a9 ia the construction of moment-ratio diagrams. Hosking [1992] also shows that L moments are preferred to product mo ‘ments in the context of evaluating the power of alternative hypothesis tests for the normal distribution. Since the theary ‘of L moments is relatively new [Hosking, 1990] we outline it briefly inthe following section. The remainder of the paper is devoted to comparisons between L moment and product ‘moment ratio diagrams and to Monte Carlo experiments which reveal the remarkable bias associated with product moment ratio estimators. ‘Tue Tueory oF L Moments L moments and probability weighted moments (PWMs) are analogous to ordinary moments in that their purpose is to summarize theoretical probability distributions and ob- served samples. Similar to ordinary product moments, L moments can be also be used for parameter estimation, interval estimation, and hypothesis testing. Although the theory and application of L moments parallel those for conventional moments, L. moments have several important advantages. Since sample estimators of £ moments are always linear combinations of the ranked observations, they fae subject to less bias than ordinary product 1 This is because ordinary product moment estimators such as 5? ‘and G require squaring and cubing the observations, respec: tively, which causes them to give greater weight to the ‘observations far from the mean, resulting in substantial bias and variance. Hosking [1990] and Stedinger et al. (1993] provide a summary of the theory and application of L moments. Greenwood et al. [1979] summarize the theory of PWMs. Perhaps the simplest approach to describing L moments is by first defining probability weighted moments because L ‘moments are linear functions of PWMSs [Greenwood et dl., 1979; Hosking, 1990]. PWMs may be defined by B= EXP ON) where Fy(x) is the cumulative distribution function of x. When r = 0, By is the mean streamflow y defined earlier in (1a). Hence a sample estimate of the first PWM, which we term bo, is given by m in (2a). All higher-order PWMSs are simply linear combinations of the order statistics Xj) = ++ = Xqy Landwehr et al. {19791 recommends the use of biased estimates of PWMs and L moments, since such estimators ‘often produce quantile estimates with lower root-mean- square error than unbiased alternatives. Nevertheless, unbi- ased estimators are often preferred in goodness of fit evalu- ations such as L moment diagrams. Unbiased estimators are preferred because they have less bias for estimating ry and ay aind ey ate invariaut if the data are multiplied by a constant, which is not the case for the biased estimators. Unbiased cample estimates of the PWMs, for any distribu- tion can be computed from eo ‘Vos. AND Feanessey: L Mower Disokans 4a) a a) SS [ia- am -i- 0) ba (a2 De-2) Jeo “wo [rae -s- ym -j-2) n Bl mtn = Dla 2m 3) fe “a where xy) represents the ordered streamflows with x.) being the largest observation and x,.) the smallest. The PWM estimators in (4) can be more” generally described using ) For any distribution, the first four L moments are easily computed from the PWMs using A= Bo (6a) 281 Bo (66) Ay = 6B2— 681 + Bo (60) A= 2085 ~ 3082+ 12B1~ (oa) ‘The first four unbiased L moment sample estimators are obtained by substituting the PWM sample estimators 6, from (4) into the L moment equations in (6). Equations (6a)-(6d) are special cases of the general recursion Secor) dee ‘A computer program is available for implementing the method of L moments (Hosking, 19916]. In the following sections we briefly define L moment ratios, discuss their relationship to conventional moments, and introduce £ ‘moment diagrams L Moment RAvios AND THE IvtereRetaTion oF L MOMENTS Analogous to the product moment rativs, evellicient of | variation C, = of, skewness y, and kurtosis x, Hosking [1990] defines the J. moment ratios = L woeficient of variation (8a)oats. ano Fesnessey: L Moment DiAckaMs aol oF Z| a 2 5 2 a L-Skewness, % lationship between L kurtosis and L skewness for the exponen- til, uniform, normal, Gumbel, Pearson type 3(P3), three porameter lognormal (LN3), generalized Pareto (GPA), Generalized Extreme Value (GEY), and the lower bound of the five-parameter Wakeby (WAS) aistributons. L skewness (86) r= 15 kurtosis (Be) % where A,,7= 1, +++ ,4 are the first four L moments and +5, ty, and 7 are the L coefficient of variation (L-C,), L skewness, and L kurtosis, respectively. The first L moment A is equal to the mean streamflow ya, hence itis @ measure of location. Hosking [1990] shows that A2, 13, and 14 can be ‘thought of as measures of a distributions scale, skewness, and kurtosis, respectively, analogous to the ordinary mo: ments o, 7, and x, respectively. Moment Ratio Diackams Conventional product moment ratio diagrams compare the sample estimates of the product moment ratios C, = sim, G, and k with their thearetical comnterparts ej, . al x for range of assumed distributions. MfcCuen {1985} provides an introduction to product moment ratio diagrams. Similarly, L moment ratio diagrams introduced by Hosking [1990] com- pare sample estimates of the dimensionless ratios 7, 7), and ‘with their population counterparts for a range of assumed distributions. significant advantage of moment ratio dia- [grams is that one ean compare the fit of several distributions Using a single graphical instrument. For example, Figure | displays the theoretical relationships between L kurtosis r, and L skewness 7) for the normal, Gumbel, exponential, ‘uniform, three-parameter lognormal (LN3), Pearson type 3 (P3), generalized Pareto (GPA), generalized extreme value (GEV), and the lower bound of the five-parameter Wakeby (WAS) distnbutions, Here one observes the tremendous flexibility of the three-parameter distributions (LN3, P3, GPA, and GEY) in comparison to the two-parameter alter- natives. The five-parameter Wakeby distribution is the most flexible distribution considered since it defines a two- ‘dimensional region in Figure 1 rather than a curve or a point as isthe case for the three- and two-parameter distributions, respectively (see Kotz and Johnson [1988] for a summary of a7 the Wakeby distribution). The theoretical relationships be- ‘ween 74 and 75 depicted in Pigure 1 were obtained from the polynomial approximations developed by Hosking [19914] nd summarized in the work by Stedinger et al. (1993) Stedinger et al. (1993) also describe the theoretical proper- ties of these distributions. Hosking (1990] introduced L moment diagrams for the purpose of selecting a suitable pdf for modeling hydrologic variables at many sites. Examples of L moment diagrams can be found in the works by Wallis (1988, 1989]. Hosking [1990], Hosking und Wallis [1987a, 1991], and Voxel etal [1993a, 6]. For example, Hosking and Wallis (1987a} found L moment diagrams useful for selecting the GEV distribution over the gamma distribution for modeling annual ‘maximum hourly rainfall data. Similarly. Wallis 11988, Fis ure 3] found an Z. moment diagram useftl for rejecting Jain and Singhs (1987) conclusion that annual maximum flood: flows at 44 sites were well approximated by a Gumbel distribution and for suggesting a GEV distribution instead Vogel eral. (1993a} used L. moment diagrams to show that floodflows at 383 sites in southwestern United States were equally well approximated by the LP3, LN3, and GEV distributions. Vogel etal. [1993] used L moment diagrams to show that floodfiows were well approximated by a GEV distribution in the regions of Australia which receive most rainfall during winter months and by a GPA distribution in the regions of Australia which receive most rainfall during summer months. ‘E moment diagrams are clearly useful for evaluating which distribution(s), among a suite of possible models, provides a satisfactory approximation to the distribution of « particular hydrologic variable in a region, A COMPARISON OF L MOMENT AND Propuct Moses Ramo Disckasts In this section we evaluate the use of product-moment ratio diagrams for selecting a suitable probability density function (pdi) for modeling the frequency and magnitude of average daily streamflow in Massachusetts. Selection of a suitable pdf for average daily streamflow is important in ‘studies which seek to develop regional hydrologic models for describing daily streamflow duration curves (see Fennessey ‘and Voge! (1990] for a recent review of the literature on regional flow duration curves). Daily streamflow duration curves are used routinely for solving a wide range of water resource engineering problems including, but not limited to. hydropower feasibility analyses, wasteload allocation, river and reservoir sedimentation, water quality management, and the determination of instream flow requirements. Stream. flow duration curves for ungaged basins are often abtained using regional hydrologic procedures which require an as- sumption regarding the distributional properties of daily streamflow in a region. The following experiments document that in spite of the complexity of the physical processes which generate daily streamflow, it may be reasonable 10 approximate the distributional behavior of daily streamflow Using a simple three-parameter probability distribution, in Massachusetts. L moment ratio and product-moment ratio diagrams sre constructed using average daily streamflows for 23 of the U.S, Geological Survey’s streamflow gaging stations in Mas- sachusetts with the following attributes.8 Voce. AND FENNESSEY: L Mowewr Diaonas TABLE 1. Characteristics of Streamflow Records of 23 Massachusets Sites Produet Moment L Moment Ratios Ratios USGS Gage Years of Areg, Meena), ‘Number Rood ome ont Gna 1180500 St 5270 1078 O66 0583 0.48 5 1658 © 6691116 0590 O61 0273 53581 2 4139-912 O87 0822 0329 S21 ae M339 61 sms 06 0.266 478 469 2 Bios 87 ss4 OaI8 0.249 203 19 sim 786 ORS OT? 0.270 408298 Ho $467.7) 0528 0421 0228 29139 % 2S 49593. OL30s O34 Gn 67 230369 0.868 0.389 0.192 3233 373518004627 0.536 0361 as 19792 50 B00 RSS O60S O88 0.359 68S 91d 2 i620 SHO 0454 0.277 Soe 630 Moos Ose OIL Om 67 M64 St 94001906 OIL OSA 0.356 BRR S42 58-4090 96.56 0590 S40 0.386 62 813 do 2312070379 O4y6 02m sa 43 4 20 SOS O46 08 479 is 6 2102000590 048s 08S Sm a Si $400 96:8 OSS? OBI 0.304 sal 840 71 150.00 24490487. 0.367 0.187 S47 916 3 19306 397 0489 OI Ls RTS 2820 m 17 26 60 SIZ OSI 174 999 BIA = BL, 1. All of the rivers are perennial 2. No significant withdrawals, diversions, or artificial recharge areas are contained in the basins; hence we con- rider these streamflows to be estentially unregulated. 3. The streamflow records range in length from 11 to 77 Years. typical of a wide range of regional samples encoun- tered in practice. ‘The U.S. Geological Survey Station numbers, record lengths, drainage areas, L moment ratio, and product- ‘moment ratio estimates are summarized in Table |. Further information regarding these 23 sites, including selected streamflow statistics, other drainage basin characteristics ‘and a location map ie included in studies by Vogel and Kroll 1990, 1992] and Fennessey and Vogel (1990). L Moment Ratio Diagrams for Daily Streamftows in Massachusetts Figure 2 (top) is an L moment diagram which uses solid circles to illustrate the relation between sample estimates of L kurtosis and sample estimates of L skewness, computed frou the complete records of daily streauilow at the 23 basins in Massachusetts. For comparison, the curves reveal the theoretical relationships between L kurtosis and L skewness for the LN3, GEV, P3, GPA, and the lower bound for the WAS distribution. Sample estimates of L skewness and L kurtosis are obtained using the unbiased L moment estimators given in (4), (6), and (8). Even though estimates of the L moments 4,,r = 1, 2, 3, and 4 are unbiased, this does ‘not imply that the sample ratio estimators L-C,, L skew: ress, and L kurtosis are unbiased. fall the three-parameter models tested in Figure 2 (top), the generalized Paroto (GPA) model provides the best ap- proximation to the observed daily streamflows. When one ‘compares the overall behavior of L moment diagrams for these distributions in Figure 1 to the results in Figure 2 (Wp), ‘one observes the remarkable ability of L_ moment diagrams for distinguishing among relatively similar distributional hypotheses. We conclude from Figure 2 (top) that the GPA distribution provides a better overall approximation to the distribution of daily streamflows in Massachusetts than the LN3, GEV, P3, Gumbel, normal, uniform, or exponential models. Since the GPA distribution is a special case of the WAS distribution, one would expect the WAS distribution to provide the best ft among all models compared in Figure 2, at the expense of having to estimate two additional param- tere Figure 2 (bottom) is an L moment diagram which illus- trates sample estimates of L-C, 73. versus sample estimates, of L skewness 7, forthe same sites as in Figure 2 (top). For comparison, we plot the theoretical relationship (using a solid line) between and 1 corresponding 0 the (wo- parameter GPA distribution. Hosking and Wallis [19876], ‘Stedinger etal. 1993}, and Vogel et al. {19936} deseribe the theoretical properties of the two- and three-parameter GPA distribution. The (wo parameter GPA distribution is equiv alent to the three-parameter version with the lower bound &, set to zero. The two-parameter GPA distribution provides a ‘good approximation to the relationship between L-C, and L. skewness for this region. Comparison of Product Moment Ratio and L Moment Ratio Diagrams for Daily Streamfiows in Massachusetts The immediate question which arises in the mind of anyone who has not experienced the value of L moment diagrams is, Would the use of conventional product momentVoc. AND FENNESSEY: J. MOMENT DIAGRAMS ratio diagrams lead to the same conclusions as the use of L moment ratio diagrams?. In Figures 3 and 4 we compare the use of Z moment ratio diagrams with conventional product, ‘moment ratio diagrams, using the same Massachusetts sites as in Table 1 and Figure 2. In all cases, the open circles denote the use of the complete period of record. and the solid circles denote the use of the complete period of record without the largest observation. One would expect the sample estimates of both L moment ratios and conventional product moment ratios to be relatively insensitive to the largest observation, since we are dealing with samples wit thousands, and at some sites, tens of thousands of daily streamflow observations. As expected, the L moment ratio diagrams depicted in Figures 3 (op) and 4 (top) are relatively unaffected by dropping the largest observation. However. the conventional product moment ratio diagrams illustrated in Figures 3 (botiom) and 4 (bottom) are significantly pacted by dropping the largest observation. Dropping the largest observation should tend to reduce the estimated kurtosis, skewness, C,., L kurtosis, L skewness, and L-C,, since it is the large’ observations which tend to cause skewness (in a positively skewed population). This is exactly the effect we observe in the L moment diagrams in Figures 3 Gop) and 4 (top); all the paints are shifted down and to the left when the largest observation is removed. Yet the con- ventional product moments behave much more unpredi (0.407 Gheerved Valuse of Dally Streamflow, et oan] rere gow] 3 5 t 2 aaa L-Skewness, 7, om © Ober Ya jan ooernes © Saw B wh 089: 2 s °4§50° 085 040 Ons 08d 085 O80 L-Skewness, 1, Fix, 2. moment dram describing the snp an hoe ical relationships between (op) L kurtosis 74 and L skewness 7 and (ottom) Z coefficient of variation 7; and L skewness 7 for average dally rtreaonlows at the 23 ster in Massachueote, simmarized in Table | 19 a0 © tiered bay Sonate os] ° fiSoat wanimum Vase” re — Generalised Pareto, GPA e ° ° i ° ic Te aso 7 ous i S30 08s G4 o4s 05 OBS G0 : 7 8 L-Skewness, 7 g *[oaeminy aman > ] Becher uly Surantow “ 5 S 20 2 = Bus 3 8 8 j % 0 Skewness, G Fie, 3. (Toe) moment eto diagram comparing the samole and GPA theoretical relationship Between L-C, 7 and L skewness and (bottom) product moment ratio diagram comparing, the ‘Ample and GPA theoretical relationship between coeficient of Vatation Cand skewness G, using average daly streamiows at the 28 sites in Massachusetts, summarized in Table | ably. In some instances dropping the largest observation leads 10 remarkable reductions in kurtosis and skewness. (note kurtosis is plotted using a logarithmic axis), yet in ‘other instances the sample estimates are not noticeably affected, Hence itis not surprising that the sample product ‘moment ratios are in poor agreement with the theoretical relationships for the GPA distribution illustrated in Figures 3 (bottom) and 4 (bottom). Apparently, estimated sample values of C,,.y, and x are highly biased even when one is 1.73). Note that even though the skewness of the GPA distribution is undefined for most values of C, considered here, L moments are still ‘meaningful; another important advantage of L moments over product moments. The results are reported in Tigures S and 6. Results Figure 5 summarizes Bias(C,) and rmse(C,) as a fraction of the true value C, versus sample size n and population C,, tor LNZ and GPA observations. Similarly, Figure 6 summa: izes Bias(G) and rmse(G) as a fraction of the true value 7, versus sample size m and population 7, for LN2 ubseivar tions. What is most revealing about these results is that the bias of both C, and G is extraordinary for samples with values of C, in excess of 5, even for sample sizes in excessVoce. anb PENNessey: L Mowent DiacaMs 1751 x= Bia] auselel Sample Size, n Sample Size, n Sample Size, n ‘The bias and rme of product moment estimates of Cy (ata fraction of the true value of C) are illustrated 1s a function of sample size m and the population value of C for two-parameter lognormal (LN2) samples and ‘wo-parameter generalized Pareto (GPA) samples. Each point in the figure is based on 10,000 replicate experiments, 19f 1000. Apparently, as the population values of both Ca increase, even large sample estimates of product moment Fatios contain almost no information about either the coeff- cient of variation or the skewness of the samples. Further- ‘more, the root-mean-square error (rmse) of both estimators CC, and G are generally dominated by bias, and since this bias disappears very slowly with sample size, the rmse of both C, and G also disappears very slowly for highly skewed samples. It is also clear from Figure 5 that the sampling properties of estimates of C, are quite different forthe LN2 ‘and GPA populations, hence efforts to derive unbiased estimators of C, would require distributional assumptions. Since our earlier objective was to employ moment ral diagrams for selecting a suitable distribution for daily lows, use of unbiased product moment estimators would be trou: blesome because a distributional assumption is required to develop unbiased product moment estimators (a chicken and gg problem) “The bias associated with both estimators C, and G results| from the fact that these product moment estimators requi squaring and cubing of the differences between each obs vation and the sample mean. As was shown in Figures 3 and 4, the sum of the square. cube. and fourth power of the difference between the observations and their sample mean is dominated by the few extraordinarily large observations typical of most highly skewed daily flow records. Hence one expects the bias and rmse of G to be much greater than the bias and rmse of C,, as is observed in Figures 5 and 6. Similarly, one expects the bias and rmse associated with Product moment estimates of kurtosis to be still greater: however, we elected not to compute it since our results for C, and » provide sufficient evidence to discontinue the use of product moment estimators for highly skewed samples, 7- 2G) 10 oe. om. os: og: oa oo RMSE[G] 7 We Distrfutlon ‘S10 io i000 Sample Size, n Fig. 6, The bias and rmse of product moment estimates of skewness G (as fraction ofthe true value, 9) ae illstated as 2 funetion of sample size m and the population value + for two- parameter lognormal (LN2) samples, Each point in the igre ic based on 10,000 replicate experiments.ns Concwusions ‘A comparison of conventional product moment ratio and L moment ratio diagrams in Figures 3 and 4 illustrates how remarkably sensitive sample estimates of the coefficient of variation C,, skewness G, and kurtosis k are to the large observations often associated with long records of daily streamflow. A Monte Carlo experiment documented (see Figures 5 snd 6) the significant downward bias associated “with large-sample estimates of C, and G for lognormal and peneralized Pareto samples. Combining the results of onr large-sample Monte Carlo experiments in Figures 5 and 6, ‘with the extensive small-sample computer experiments per- formed by Wallis et al. {1974}, we conclude that product ‘moment estimates of C, and y should be replaced by L ‘moment estimators for most goodness of fit applications 1m hydrology including both smal-sample, low-C, applications see Walla er al, 1974] and both small» and Tasge-sample, high-C, applications. ‘There are two general situations in which relatively unbi- ‘sed product moment estimators arc available, When one transforms the observations by taking logarithms, product moment estimators will behave similarly to L moment esti- mators. Unfortunately, transforming the data will not usu- ally assist us in determining the distributional properties of the original data except in special cases such as the LP3 and LN3 disuibutions. Fustherimote, streanullow cant always be transformed by taking logarithms, since many flow se- quences contain zeros. Another situation in which relatively unbiased product moment estimators are available is in large-sample, low-C,, applications; however, such situations are not usually encountered in hydrology. ‘We conclude that for the purpose of constructing moment diagrams in hydrology. L moment diagrams are always preferred over product moment diagrams because the L ‘moment ratios L-C,, L skewness, and £ kurtosis are neatly unbiased for all combinations of sample size and C,, and for all populations (Hosking, 1990) “Acknowledgments. The authors gratefully cknowtedge the f- ‘nancial support ofthe U.S. Geological Survey who sponsored this esearch under the terms of contract 14-08-0001-G2071 titled The ‘Application of Streamfow Duration Curves in Water Resources Engineering. In addition, the authors acknowledge the francial support ofthe Tufts University Center for Environmental Manage: tment. The authors are also grateful to J. RM. Hosking for providing useful comments on an early version of this manuscript. ‘The views amd conclusions contained inthis dovusnent are those of the authors and should not be interpreted as necessriy represent ing the oficial polices, either expressed or implied of the U.S. Government REFERENCES Fennessey, N. M., and R. M. Vogel, Regional flw-duration curves Tor ungated ses in Matsactunetts, J. Water Reovwr Plann Manage., Div. Am. Soc. Civ. Eng, 1164), $30-S49, 1990, Fischer, R./A., Moments and produet moments of sampling distr- butions, R. London Math. Soc. 2(30), 199. 1929. Greenwood, J. A.,J-M. Landwehr, N.C. Matalas and J. R. Wallis Probability weighted moments: Definition and relation to param celery of several distributions expressible in inverse form, Water Resour. Res. 15(3), 1089-1088, 1979. ‘Voce, an FeNvessty: E Moms Dinas Hazen, A., Flood Flows, Joba Wiley, New York, 1930. Hosking, 3. RUM., Lemuinents: Anaiysis and esdimation of distri: ‘butions using linear combinations of order statistic, J. R- Stat Soe, Ser. B. 5202), 108-124, 1990, Hosking, J. R- M., Approximations for use in constructing L:mo- ment rato diagrams, Res, Rep. RC-16635, 3 pp. 1BM Res. Div. Tad, Walson Res. Cent.. Yorktown Heights. N.Y. 199ta, Hosking, 1K. M.. Fortran routines for use with the method of emoments, Version , Res. Rep. RC-17087 (#75864), 117 pp. IDM Res. Div, T.J. Watson Res. Cont. Yorktown Heighte, N.Y., 19918. Hosking, J. R. M., Moments or L moments? An example comparing ‘wo measures af alstributonal shape, Am Sta. #0), 140-189, 982. Hosking, J. R.M..andJ.R. Walls, An “indexlood" procedure for ‘egional rainfall frequency analysis, Eos Trans. AGU, 68, 312, 1987 Mocking, J. RM. nnd 1. R. Wall, Parameter and quantile ‘stmation forthe generalized Pareto distribution, Techometrics, 2013), 339-349, 1987. Hosking, J. R. M., and J. Re Wallis, Some statistics useful ia feglonal frequency analysis. Res. Rep. RC 17096, 23 pp., IBM Res. Div. TJ. Watson Res. Cent., Yorktown Heighs. N.Y, 19 Jain, D.,and V.P. Singh, Estimating parameters of EVI distribution {or food frequency analysis, Water Resour. Bull, 23(1, 38-71, 1987, Kirby. W., Algebraic boundedness of sample staisties, Water Resour Bee 1042). 220-997. 1074 Kotz, 8, and N. L. Johnson, Encyclopedia of Statistical Sciences, ‘ol, 9. pp. S13-Sid, John Wiley, New York, 1988, Landweh J. Mi, N.C. Matalas, and J. R. Wai, Probability weighted moments compared with some traditional techniques in ‘estimating Gumbel. parameters and quantiles, Water Resour Fees 1313), 1099-1008, 19. MeCuen, Ro H., Statistical Methods for Engineers, pp. 61-63, Prentice-Hall, Englewood Cliffs, N. J, 1985 Stedinger. J. R., R. M. Vogel, and E” Foufoula-Georsiou, Fre- ‘quency analysis of extreme events. in Handbook of Applied Hydrology. chaps. 18, edtorin-chie D.'A. Maidment, MeCraw- Hill, New York, 1993 Vogel, R. M., and C. N. Kroll, Generalized low-flow frequency ‘elatonships for ungaged sites in Massachusetts, Water Reso Bull. 26(2). 241-253, 199, Vogel, R. M.,and C. N. Kroll, Regional geohydrolosic zeomorphie Feluionships Tor the estimation of low.tiow satshes. Water Resour. Res., 289), 2481-2458, 1992 Vogel, R. M.. and D. E. MeMartn, Probability plot goodness-of-fit tnd’ skewsess. estimation procedures for the Pearson type 3 distribution, Water Resour. Res., 2712), 3149-3188, IL Vogel, R. M, W. 0. Thomas, and T. A. MeMahon, Floodflow ffequeney model selection in Southwestern U-S.A\. J. Water Resour. Plann. Manoge., Div. Am. Soc. Civ- Eng. 1190). 353-366, 19980 Vogel, R.'M., T. A. MeMahon, and F. H. 8. Chiew, Floofow frequency model selection in Australia, J. Hydrol, in press, 19936, Walls, J. R., N.C. Matalas, and JR, Slack, Just a moment, Water Resour. Res., 142). 11-219. 1974, Walls, JR, Catastrophes. computing and containment: Living in ‘ur Festiess habitat, Spec. Se. Tech. 11a), 296-315, 1988, Walls, J. Ry Regional frequency studies using L moment, Ree. ‘Rep. 14597, 17 pp. BM Res. Div., T. J. Watson Res. Cent., ‘Yorktown Heighis, N. ¥. NN. M. Fennessey and R. M. Vogel, Department of Civil and Environmental Engineering, Tufts University, Medford, MA O215S. (Received August 19, 1992: revised February 4, 1983; accepted February $, 193)