1 Submitted to Water Resources Research, 1997.

Regional Frequency Analysis Methods for Evaluating Changes in Hydrologic Extremes

A. Allen Bradley
Iowa Institute of Hydraulic Research and Department of Civil and Environmental Engineering, University of Iowa, Iowa City, Iowa

A common assumption in frequency analysis is that hydrologic extremes  oods or heavy precipitation are generated by a random process. This implies that natural climatic variability does not change the distribution of extreme events. A regional frequency analysis approach is proposed to test the hypothesis of randomness over secular time scales. Observed regional occurrences of extreme events are compared to those from a random process. Signi cant departures may indicate nonrandomness due to climatic variability. Application of the approach to a region in the Southern Plains indicates nonrandomness in annual maximum precipitation.

Abstract

1. Introduction
Engineers and hydrologists estimate probabilities associated with large oods or heavy rainfall by analyzing hydrologic records. A probability distribution is chosen to model the frequency of extreme events, and the distribution is t to historical data using statistical techniques for examples, see U. S. Water Resources Council, 1982; Stedinger et al., 1993 . An important assumption made with this approach is that the process is random. This means that extreme events are independent and identically distributed with some unchanging underlying probability distribution. The treatment of hydrologic extremes as a random process implies that natural climatic variability does not a ect the occurrence of extreme events. Yet for some regions, there is compelling evidence that hydrologic extremes respond to large-scale climatic forcings. Eltahir 1996 found that 25 of the natural variability of the Nile River ood can be associated with the El Ni~o Southern Oscillation ENSO. n He concludes that the relationship provides a physical explanation for the Hurst phenomenon Hurst, 1951; Mandelbrot and Wallis, 1968; Kleme, 1974 s for the Nile ood series. Ely et al. 1994 observed that large winter oods in the southwestern United States occur more frequently during multiple-year periods of the low phase of the Southern Oscillation El Ni~o events, and that major oods are absent durn ing the high phase. They conclude that global-scale climate anomalies have a strong impact on the occurrence of winter oods in the region. Numerous other studies have established linkages between natural climatic variability and hydrologic variables on monthly and annual time scales Lins, 1985; Ropelewski and Halpert, 1987; Cayan and Peterson, 1989; Redmond and Koch, 1991; Leathers et al., 1991; Cayan and Webb, 1992; Kahya and Dracup, 1993; Dracup and Kahya, 1994; Dettinger and Cayan, 1995; Mann et al., 1995; Lall and Mann, 1995; among others . A few studies have considered the e ects of climatic variability on the probability of hydrologic extremes. Knox 1993 found that relatively modest changes in annual temperatures 2 C  during the Holocene have resulted in large changes in the frequency of extreme oods in the Upper Mississippi Valley. The e ects of climatic variability over secular time scales have also been explored. Knox 1984 estimated ood quantiles for the Mississippi River at St. Paul 1867-1981 for four climatic periods de ned

2 by independent investigations of atmospheric circulation patterns. Large di erences were observed in extreme ood quantiles for the four climatic periods. Webb and Betancourt 1992 observed an increase in ood magnitudes on the Santa Cruz River Arizona in recent decades, due in large part to an increase in oods from dissipating tropical storms during El Ni~o years. As a result, the estimated magnitude of n the 100-year return period event increased by a factor of ve. Hirschboeck 1988 proposed a hydroclimatological model of nonstationarity for ood distributions. She conceptualized the underlying ood distribution as a mixture of ood distributions associated with distinct atmospheric forcings. Interannual and interdecadal variations in the frequency atmospheric circulation patterns would then lead to nonstationarity. Other investigations that have considered the issue of nonrandomness in hydrologic extremes include Wall and Englot 1985 , Booy and Lyle 1989 , Krasovskaia and Gottschalk 1993 , and Lisi and Villi 1997 , among others. Although the assumption of randomness may be clouded by natural climatic variability, for many problems in water resources, it remains a useful operational model. However, for some problems e.g., predicting the impacts of climate change on extreme events, the relationship between hydrologic extremes and climate must be investigated. To what extent is there a climatic signal in records of extreme hydrologic events? Are the variations in extreme events su ciently nonrandom to distinguish them from a random stationary process? A serious obstacle in answering these questions is the shortness of hydrologic records. However, if climate variability is responsible for changing the underlying distribution of extreme events, the change should a ect a fairly large region due to the mismatch between scales for climatologic and hydrologic phenomena Hirschboeck, 1988 . Therefore, an e ective means for testing whether there have been changes in the frequency of hydrologic extremes is to use a regional approach. In this paper, a regional frequency analysis approach is proposed to test the null hypothesis of randomness over secular time scales. The approach builds on recent developments in regional ood frequency analysis. An index ood procedure is used to model hydrologic extremes. A random process can then be simulated based on the assumed regional model. The observed occurrences of extreme events are then compared to those predicted by the random model of extremes. Deviations from the random

3 model provide evidence of nonrandomness in hydrologic extremes. The regional frequency model and the proposed test for nonrandomness are described in the next two sections. The approach is then applied to a region in the Southern Plains of the United States. where qp is the estimated regional quantile function. ^

3. Tests for Nonrandomness

The regional model assumes that the regional distribution does not change through time. Therefore, in each year j , the non-dimensionalized sample data fZij ; i = 1; M g represent a sample drawn from the regional distribution. Let Zj p be a nonparametric estimate of the regional quantile qp in year j. Zj p is found based on the ranked sample data for year j . Exceedance probabilities for the ranked data are estimated using the Weibull plotting position. The time series fZj p; j = 1; N g will be referred to as the annual quantile series. Any number of statistical tests for nonrandomness can be applied to the annual quantile series. Kendall's correlation coe cient can be used test for trends in Zj p. This is a nonparametric test that can detect monotonic trends Hirsch et al. 1991; Hirsch et al. 1993 . For each value fZj p; j = 1; N , 1g, one counts the number of the subsequent values that are greater than Zj p, and the number of values that are less than Zj p. The test statistic S is equal to the number of values that are greater minus the number of values that are less. The standardized test statistic Z is: 8 pS,1 S 0; V S  S = 0; 4 Z= 0 : pSV+1S S 0;  where V S  = nn , 12n + 5=18. For samples sizes greater than 10, the test statistic Z is approximately normally distributed. Another approach for testing for nonrandomness exploits the regional framework. The basic idea is to use the regional frequency model to generate realizations of Zj p for a stochastic process that mimics the process assumed for the region. Observed occurrences of Zj p are then compared with those for the stochastic process. Let Sjk p be the k-year moving average of Zj p:
k Sj p = j +k ,1 X

2. Regional Frequency Analysis

Many di erent techniques are used to regionalize information on oods and extreme rainfall. One approach is the index ood procedure, originally developed for ood frequency analysis Dalrymple, 1960 , but more recently used for precipitation frequency analysis Guttman et al., 1993 . The procedures presented below are described in detail in Hosking and Wallis 1993 , Stedinger et al. 1993 , and Hosking and Wallis 1997 . In the index ood approach, the quantile function Qi p for the i-th site in the region is: Qi p = i qp; 1 where p is the exceedance probability 0  p  1, i is a site-speci c scaling factor, or index ood", and qp is the regional quantile function. This implies that quantile functions at all sites in the region are identical, apart from the site-speci c scaling factor i.e., a homogeneous region Hosking and Wallis, 1993 . It also assumes that the process is stationary i.e., no watershed or climate changes. The scaling factor is usually estimated by the site-speci c sample mean i . Sample data for a site fXij ; j = 1; :::; nig ^ can be non-dimensionalized using: Zij = Xij =i ; j = 1; :::; ni : ^ 2 The parameters of the regional quantile function are commonly estimated using the method of probability weighted moments PWM Greenwood et al., 1979; Hosking, 1986 . First, the sample L-moments are computed for each of the M sites using the unbiased PWM estimators Hosking, 1986; Hosking, 1990 . Regional L-moments are computed as the weighted average of the sample L-moments for the M sites. The weighting factor is proportional to the sample length ni . The regional L-moments are used as estimates of the L-moments of the assumed regional distribution. The form of the regional distribution may be selected using a goodness-of- t measure Vogel, 1986; Vogel and McMartin, 1991; Chowdhury et al., 1991; Hosking and Wallis, 1993 . The quantiles for individual sites are estimated by: ^ Qi p = i qp; ^^ 3

Zi p : k i=j

5

Three test statistics from the observed series can be used to test for nonrandomness:
k k Smin = minfSj p; j = 1; N , k + 1g;

6

4
k k Smax = maxfSj p; j = 1; N , k + 1g;

7 8

and
k Srange = Smax , Smin :

The use of moving averages of Zj p makes it possible to evaluate nonrandomness over a range of time scales. One possible cause of nonrandomness could be short-term climate anomalies that change the distribution of extreme events. For example, the 1993 oods in the Midwest may be an example of a shortterm climate anomaly several months that resulted in an increased probability of extreme events Kunkel et al., 1994; Mo et al., 1994; Bell and Janowiak, 1995 . Another possible cause of nonrandomness may be secular changes in the distribution of extreme events. For example, stationary climatic periods with di erent distributions would lead to nonrandomness over secular time scales. Similar ideas have motivated the shifting-level models used to describe the Hurst phenomenon Kleme, 1974; Potter, 1976; Boes and s Salas, 1978 . Each of the three test statistics can be compared to the random model of hydrologic extremes. Monte Carlo simulation is used to derive the distribution of k k k Smin , Smax , and Srange for a random process. The distributions are then used to compute signi cance levels. This approach has similarities to that proposed by Livezey and Chen 1983 , which uses Monte Carlo methods to determine the spatial signi cance from local hypothesis tests applied to a set of sites. In the next section, these techniques are used to test for nonrandomness in extreme precipitation for a region in the Southern Plains.

4. Case Study
Figure 1 shows the locations of the 43 hourly precipitation stations in the Southern Plains study area. These stations were selected because 1 the stations were in operation continuously during the 42-year study period from 1949-1990, and 2 the distance between adjacent stations is at least 50 km. For each station, annual maximum precipitation series were extracted for durations ranging from 1 to 72 hours. For all durations, the peak occurrence of annual maxima is in May. However, for short durations  3 hours, there is a steady decrease from the peak in May through the summer and into the fall. For longer durations, a secondary peak in September appears. The two peak months become more dominant as the

precipitation duration increases  24 hours. These results are consistent with previous studies of extreme rainstorms in this region by Bradley and Smith 1994 and Smith et al. 1994 . The proposed region was evaluated using the guidance provided Hosking and Wallis 1993 and Hosking and Wallis 1995 . The sample L-moments for each station were computed and compared using the heterogeneity measure H. This measure compares the degree of variability of the sample L-moments with that for a homogeneous region. The results are shown in Table 1. The region was found to be acceptably homogeneous H 1 for durations of 6-hours or longer. The region was found to be possibly heterogeneous 1  H 2 for 2- to 3-hour precipitation durations, and de nitely heterogeneous H  2 for the 1-hour duration. Despite concerns of heterogeneity for short durations, the 43-station region was used for all durations to maintain consistency in the regional de nition. The selection of the probability distribution for the regional quantile function was based on the goodnessof- t measure G Hosking and Wallis, 1993 . For all durations, the generalized extreme value GEV distribution was selected as the best t. However, for the 72-hour duration, jGj was slightly greater than the 1.65 value recommended. Since the deviation is quite small, the GEV distribution was also used for the 72-hour duration. Table 1 shows the estimated parameters of the tted distribution and estimated regional quantiles for each duration. In each year of the 42-year period, Zj p was estimated for exceedance probabilities of 0.5, 0.25, 0.1, 0.05, and 0.0227. These correspond to nonparametric annual quantile estimates of the 2-, 4-, 10-, 20-, and 44-year return period events. Note that since the annual sample contains 43 sites, Zj 0:0227 is simply the series of the maximum non-dimensionalized value observed in the region for each year. Figure 2 shows Zj 0:50 and Zj 0:10 for 1- and 24-hour annual maximum precipitation. The linear regression for each series suggests that the estimated annual quantiles increase over time. Note that the increases for the 24-hour precipitation are much greater than for the 1-hour precipitation.

4.1 Trend Tests for Annual Maxima

Kendall's correlation coe cient was used to test for statistically signi cant trends in Zj p for each duration and exceedance probability. All the time series have a positive value of S , indicating that ex-

5 treme precipitation increases over time for all durations. The signi cance level for rejecting the null hypothesis that there is no trend i.e., that the correlation of the values Zj p with time is zero is shown in Table 2. Many of the annual quantile series have statistically signi cant trends at the 5 level. Trends for the 44-year event are signi cant for all durations except 1 and 3 hours. For the 20-year event, the trends are signi cant for long durations  6 hours. Most of the trends for the 4- and 10-year events are signi cant at the 10 level or less. However, for the 2-year event, the only signi cant trend at the 10 level is for the 24-hour duration. Overall, it is clear that annual maximum precipitation is increasing over the study region over the period of record. Statistically signi cant trends at the 5 level were detected for one or more of the annual quantile series for nearly all durations. All the signi cant trends occurred for 4-year events or greater exceedance probabilities of 0.25 or less. In the next sections, other forms of nonrandomness in the extreme precipitation process are explored.
k Signi cance levels for the three test statistics, Smin , k , and S k , were computed using Monte Carlo Smax range

4.2 Test for Nonrandomness

simulation. A GEV regional quantile function with parameters estimated from the Southern Plains analysis was assumed to be the true quantile function. Usually, correlation of hydrologic extremes between sites is ignored in regional frequency analysis notable exceptions are the work of Hosking and Wallis 1988 and Stedinger and Tasker 1985 . However, annual maximum precipitation for gages in close proximity are correlated. Furthermore, intersite correlation increases as precipitation is aggregated over longer durations. Figure 3 shows computed correlation coe cients for 24-hour annual maximum precipitation for the Southern Plains region. Also shown is a logarithmic function tted to the data. Note the tendency for larger coe cients at smaller intersite distances. The e ects of intersite correlation were included in the Monte Carlo simulations. The simulations were patterned after those described in Hosking and Wallis 1988 . First, intersite dependence was modeled using a logarithmic correlation function. Next, samples for the 43 sites were randomly generated from a multivariate normal distribution with covariance de ned by the correlation function, and then transformed to generalized extreme value deviates. The annual quantile series Zj p and the moving averages Sjk p were com-

puted for each sample. In addition, the three statisk k k tics, Smin , Smax , and Srange , were evaluated for the 42-year long samples. The procedure was repeated for 5000 samples to derive the distribution for the test statistics. Note that in using this procedure, the correlation function should be estimated using data that have been transformed to normality. However, Hosking and Wallis 1988 note that the di erences are not signi cant for a GEV distribution with skewness between one and three. Figure 4 shows the 5 signi cance levels based on a k k one-sided test for Smin 0:1 and Smax0:1, for k ranging from 1 to 42 years, for the 24-hour precipitation. Signi cance levels are computed assuming independence between sites i.e., no intersite correlation and assuming the logarithmic correction function shown in Figure 3. The random model of hydrologic exk tremes implies that 95 of the time, values of Smin k will fall below will fall above the lower limit, and Smax the upper limit. Values exceeding the limits occur by chance only 5 of the time. For observed values of k k Smin and Smin that exceed the limits, one can reject the null hypothesis of a random process at the 5 signi cance level i.e., conclude that there is evidence for nonrandomness. Note that intersite correlation has a very signi cant e ect on the 5 signi cance levels, especially for very short moving average windows. Figure 5 shows the annual quantile series for the 10-year 24-hour event Zj 0:1, as well as its 5-year moving average Sj5 0:1. Also shown are the 5 signi cance levels for the minimum and maximum values of the 5-year moving average for no intersite correlation and with the logarithmic correlation model. Clearly, the minimum of the 5-year moving average in the 1950s is much less than would be expected for a random process with no intersite correlation. Yet even for a region with intersite correlation, the 5-year moving average is still signi cant at the 5 level. Therefore, one could reject the null hypothesis of randomness for this series at the 5 signi cance level.

4.3 Regional Results

The statistical methods were applied to annual quantile estimates for all durations. For each duration, a logarithmic correlation function was used to model the intersite correlation. Signi cance levels were then computed for the observed values of the three test statistics. None of the observed values of k Smax exceeded the 5 signi cance limit. However, k many values of Smin were signi cant at the 5 level;

6
k a smaller number of the values of Srange were also signi cant. Selected results are shown in Tables 3 through 6. Table 3 shows the estimated signi cance levels for an exceedance probability of 0.0227 44-year event k for the minimum test statistic Smin for moving averages ranging from 1 to 30 years. Only those values that are signi cant at a 10 level or less are shown. The minimum test statistic is statistically signi cant at the 5 level for 12-, 48-, and 72-hour durations. Recall that for all durations except 1 hour, the trends in the 44-year event are statistically signi cant at the 5 level. Although the signi cant values for long moving average windows for the 48- and 72-hour durations are a direct result of the long term increase in annual quantile series, the trends observed for durations of 24 hours or less are insu cient to distinguish these series from a random process. Unlike the results for 44-year event, the agreement between the signi cant trends and signi cant departures from the random model is generally better for exceedance probabilities of 0.05 or greater 2- to 20year events. Table 4 shows the estimated signi cance levels for an exceedance probability of 0.1 10-year k event for the range test statistic Srange . For a moving average of 30 years, the results are signi cant at the 10 level for durations of 24-hours or less, and significant at the 5 level for most of the durations. These results are due to the long-term increase in Zj 0:1. Recall that statistically signi cant trends in Zj 0:1 were also observed for durations of 24 hours or less. For exceedance probabilities of 0.05 or greater 2to 20-year events, the results show many statistically signi cant departures from the random process for short moving average windows. Table 5 shows the estimated signi cance levels for an exceedance probability of 0.1 10-year event for the minimum test k statistic Smin . For all durations except 1 hour, the minimum values are signi cant at the 5 level for moving average windows of 1 or 5 years. In each case, the minimums are associated with the period from 1952 to 1956, which has low estimated annual quantiles for all exceedance probabilities. For the 1and 24-hour durations see again Figure 2, the two lowest estimated quantiles within the region occur in 1952 and 1956. The results of the test indicate that such an extended period with low estimated annual quantiles would be unusual for the random model of hydrologic extremes. Table 6 shows the estimated signi cance levels for an exceedance probability of 0.5 2-year event for the k minimum test statistic Smin . Although the trends in the estimated 2-year events are not statistically signi cant, the test again shows signi cant departures from the random model for moving average windows of 1 to 5 years. Note that the results are signi cant for the 5-year moving average at the 1 level or less for durations of 2- to 24-hours. In summary, the results show evidence of nonrandomness in extreme precipitation for all durations based on the proposed test. Most occurrences of signi cant statistics were for the minimum test statistics k Smin . Signi cant values were often associated with moving average window of 1 to 5 years. A period with low annual quantiles in the 1950s was responsible. Some values of the minimum and range test statistics were also signi cant for moving average window of 20 to 30 years. These results are due to the long-term trends in extreme precipitation.

5. Discussion
There are several possible explanations for the apparent nonrandomness detected in extreme precipitation based on the regional test: 1 the signi cant departure from the random process results from limitations of the random model; 2 the signi cant departure arises due to non-climatic causes; or 3 the process is not random and the change in the distribution is related to natural climatic variability. Each of these explanations will be considered below.

5.1 Limitations of the Random Model

The use of the regional frequency model of hydrologic extremes assumes that 1 the region is homogeneous, 2 the form of the regional distribution and the distribution parameters are known, and 3 the form of the intersite correlation function and the model parameters are known. Studies by Lettenmaier and Potter 1985 , Lettenmaier et al. 1987 , Hosking and Wallis 1988 , and Potter and Lettenmaier 1990 , have shown that even with regional heterogeneity, regional frequency estimation is superior to at-site approaches. Still, regional heterogeneity will a ect the distribution of Zj p and its moving averages. Based on a measure of homogeneity H Hosking and Wallis, 1993 , heterogeneity for durations of 1 to 3 hours is a serious concern. Sample L-moments do show systematic variations across the region. Speci cally, L-CV tends to decrease from west to east. Figure 6 shows the most severe case 1-hour duration. Systematic variations are not observed for higher order sample

7 L-moments, although the larger sampling variability for these moments may mask heterogeneity. Additional Monte Carlo experiments were carried out to evaluate the e ects of regional heterogeneity on the signi cance levels. First, a sloping plane trend surface was t to the sample L-CVs. The tted value of the L-CV at each site was assumed to be the true value. Next, samples for the 43 sites were randomly generated and the three test statistics were computed from the annual quantile series. The procedure was repeated for 5000 samples to derive signi cance levels for the test statistics. Figure 7 shows the results for the 10-year 1-hour precipitation. The signi cance levels are nearly identical to those for a homogeneous region. The results for other durations, where the regional trends in the sample L-CV are less signi cant, were similar. Therefore, the test results for this region are not signi cantly altered by the apparent regional heterogeneity. In this analysis, the form of the regional quantile function was selected based on a goodness-of- t measure G Hosking and Wallis, 1993 . However, the use of the incorrect probability distribution in regional frequency estimation results in biased estimates of quantiles Lettenmaier and Potter, 1985 . For the Monte Carlo simulation of the random process, the bias would have the greatest e ect on the upper tail of the distribution. Yet in this study, signi cant departures from the random model were more common for less extreme annual quantiles 2- to 10-year events. The signi cance for these quantiles should be more robust to the distributional assumption. A simple model was used to simulate the e ects of intersite correlation for the Southern Plains region. Hosking and Wallis 1988 have noted that the intersite correlation for oods may depend on ood magnitude. Although many investigations on rainfall modeling have dealt with the form of the spatial correlation for individual storms for examples, see Laughlin, 1981; Waymire et al., 1984; Bell, 1987 , little is known about the form of the correlation for annual maxima. Clearly, the correlation structure of annual maxima would not be the same as for a single rainstorm. Figure 8 summarizes the number of days per year containing an annual maximum, a measure of the number of individual storms for each duration. The average number ranges from just over 30 1-hour duration to 23 72-hour duration for the 43 sites. Therefore, in each year, the annual maxima for sites across the region are produced by many independent storm events. Figure 8 also shows the maximum number of annual maxima in a single day in each year, a measure of the regional impact of a single storm. The maximum number averages between 4 1-hour duration and 7.2 48-hour duration. For the study period, the largest number of annual maxima in a single day was 15 48hour duration. Given the importance of the intersite correlation model on the derived signi cance levels, further research is needed to de ne the form and estimate the parameters of the correlation function.

5.2 Non-Climatic Causes

Another possible cause for the apparent nonrandomness may be non-climatic factors. Most important, precipitation records may not be homogeneous. Changes in the location of gages or the environment surrounding the gages is known to a ect the homogeneity of gages. Changes in the gaging equipment may also a ect the measurement of heavy rainfall rates. For instance, the installation of windshields at existing gage sites would increase the precipitation catch, and could result in an apparent increase in extreme precipitation. Furthermore, recording practices may a ect the quality of the data sets collected over time. Potter 1979 examined the e ect of nonhomogeneity for annual precipitation in the northeastern United States. He found that nonhomogeneous records resulted in in ated estimates of the autocorrelation function and the Hurst coe cient. He noted that these results could lead to a false conclusion that the time series have very long memory. The precipitation records were examined for homogeneity using the bivariate test described by Potter 1981 . This test compares the annual precipitation series for each site with the annual precipitation from a regional series. The areal-average annual precipitation based on a network of 1999 National Weather Service NWS cooperative observer gages within the study area was used as the regional series. The test can detect nonhomogeneity that produces a shift in the mean of the gage series e.g., the result from a gage relocation or the installation of a windshield. It is assumed that the regional series is homogeneous. Only two of the 43 gages would appear to be nonhomogeneous based on this test. For one, a shift in the mean of the series was detected in the mid-1970s. For the other, a shift was indicated in the last year of the record. Still, these results do not suggest that the apparent nonrandomness is a result of nonhomogeneous precipitation records.

8 If the extreme precipitation process in the Southern Plains is not random, and the change is related to climatic variability, concomitant changes in other hydroclimatological variables might also be expected. Figure 9 shows the areal-average annual precipitation for the Southern Plains region during the 42-year study period based on the network NWS cooperative observer gages. Note that there is an increase in annual precipitation over time, similar to that observed for the annual quantile series. However, the trend is not signi cant at the 5 level based on Kendall's correlation coe cient. Still, the trend is consistent with long-term precipitation trends for the Southern Plains observed by Lettenmaier et al. 1994 and Karl et al. 1996 . More important, Karl et al. 1996 showed that the most pronounced trend during this century is the increase in the fraction of precipitation in the form of heavy rainfall  1-inch per day. Clearly, the increases in the annual quantile series, which represent even larger precipitation amounts, are consistent with this observation. The 5-year period from 1952 to 1956 contained the lowest annual quantiles on record, and this produced many statistically signi cant results with the test for nonrandomness. From Figure 9, it is clear that the period from 1952 to 1956 was also the driest during the study period. Three years were among the four driest on record; the other two years had below-average precipitation. Figure 10 compares the monthly precipitation for the two driest years, 1952 and 1956, with monthly average precipitation for the region. During the warm season months of May through October, when most annual maxima occur, both years have below average precipitation in each month. Still, the most striking similarity is the very low precipitation accumulations in late summer and early fall, which corresponds to the secondary peak in annual maximum for longer durations. A similar observation was made by Eddy 1982 in a climatological analysis of precipitation in Oklahoma. For the ve driest years on record which include 1952 and 1956, Eddy 1982 noted that the principal cause of the below average precipitation was the absence of heavy rainfall  1:0 inch per day during the fall. To determine whether the variations in annual quantiles are related to the variations in annual precipitation, the correlation between the two variables was examined. Figure 11 shows the regression for Zj 0:5 2-year event for 24-hour annual maximum precipitation and the annual precipitation. Although

5.3 Climatic Causes of Nonstationarity

there is a great deal of scatter, there is still a clear relationship  = 0.746. Table 7 shows the correlation coe cients for each exceedance probability and duration. The relationships are signi cant for all exceedance probabilities at the 5 level or less. In general, the correlation is lower for more extreme quantiles and shorter durations. The results of this analysis suggest a linkage between extreme events and natural climatic variability on a regional basis. The atmospheric circulations that bring wetter conditions to the region also tend to favor the occurrence of more extreme events and vice versa, even for durations as short as 1 hour. Although a better statistical relationship between annual quantiles and regional hydroclimatic variables is certainly possible and should be pursued, it is clear that such relationships would be useful in predicting the impacts of regional climate change on the occurrence of extreme precipitation.

6. Summary and Conclusions

A regional frequency analysis approach is proposed to evaluate the changes in hydrologic extremes over secular time scales. Non-dimensionalized regional data are used to construct time series of annual quantile estimates. Statistical tests for trends and nonrandomness are applied to the annual quantile series. The test for nonrandomness is based on the regional model of hydrologic extremes. Observed annual quantiles are compared to those for a random process. The results are used to detect statistically signi cant departures from a random process. The proposed regional approach should be useful for identifying the e ect of climate variability on extreme events. The results may also provide a means for the assessment of hydrologic extremes under changing climatic conditions. The regional approach was applied to a study area in the Southern Plains of the United States. Signi cant trends increases in annual quantile series were observed over time. Nonrandomness in the extreme precipitation process was detected based on the proposed method. Many factors may in uence the observed departures from the random model of hydrologic extremes, including regional heterogeneity, distributional assumptions, and intersite correlation. Further research is needed to quantify their impact on the signi cance associated with the observed results. Still, there is strong supporting evidence for climate-related non-

randomness in extreme precipitation in the Southern Plains region. The increases in annual quantiles over time are consistent with the trend in annual precipitation and a trend for a larger fraction of precipitation from heavy daily amounts. Statistically signi cant departures from the random model were also associated with a 5-year period of anomalously dry conditions in the Southern Plains. Finally, strong relationships between the variations in annual quantiles and annual precipitation were observed for durations ranging from 1 to 72 hours.

References

Acknowledgments. This project was partially supported by a seed grant from the Center for Global and Regional Environmental Research CGRER, the University of Iowa Old Gold Fellowship, and grant CMS-9702715 from the National Science Foundation. The work was performed at the Iowa Institute of Hydraulic Research IIHR Computational Laboratory for Hydrometeorology and Water Resources. The author thanks Rocky Durrans, Peter Rasmussen, and an anonymous reviewer, for their careful reviews and helpful suggestions.

Bell, G. D., and J. E. Janowiak, Atmospheric circulation associated with the Midwest Floods of 1993, Bulletin of the American Meteorological Society, 765, 681 695, 1995. Bell, T. L., A space-time stochastic model of rainfall for satellite remote-sensing studies, Journal of Geophysical Research, 92D8, 9631-9643, 1987. Boes, D. C., and J. D. Salas, Nonstationarity in the mean and the Hurst phenomenon, Water Resources Research, 141, 135 143, 1978. Booy, C., and L. M. Lye, A new look at ood risk determination, Water Resources Bulletin, 255, 933 943, 1989. Bradley, A. A., and J. A. Smith, The hydrometeorological environment of extreme rainstorms in the Southern Plains, Journal of Applied Meteorology, 3312, 1418 1431, 1994. Cayan, D. R., and D. H. Peterson, The in uence of North Paci c atmospheric circulation on stream ow in the West, in Geophysical Monograph, 55, 375 397, 1989. Cayan, D. R., and R. H. Webb, El Ni~o Southern Osn cillation and stream ow in the western United States, in El Ni~o: Historical and Paleoclimatic Aspects of the n Southern Oscillation, H. Diaz and V. Markgraf, editors, Cambridge University Press, 29 68, 1992. Chowdhury, J. U., J. R. Stedinger, and L.-H. Lu, Goodnessof- t tests for regional GEV ood distributions, Water Resources Research, 277, 1765 1776, 1991. Dalrymple, T., Flood frequency analysis, U. S. Geological Survey Water Supply Paper, 1543-A, 1960. Dettinger, M. D., and D. R. Cayan, Large-scale atmospheric forcing of recent trends toward early snowmelt runo in California, Journal of Climate, 83, 606 623, 1995. Dracup, J. A., and E. Kahya, The relationships between U.S. stream ow and La Ni~a events, Water Resources n Research, 307, 2133 2141, 1994. Eddy, A., 1982. Operational weather modi cation. 5. A rainfall climatology for Oklahoma. Oklahoma Climatological Survey, 1982. Eltahir, E. A., El Ni~o and the natural variability in ow n of the Nile River, Water Resources Research, 321, 131 137, 1996. Ely, L. L., Y. Enzel, and D. R. Cayan, Anomalous North Paci c atmospheric circulation and large winter oods in the Southwestern United States, Journal of Climate, 76, 977-987, 1994. Guttman, N. B., J. R. M. Hosking, and J. R. Wallis, Regional precipitation quantile values for the continental U. S. computed from L-moments. Journal of Climate, 612, 2326 2340, 1993. Greenwood, J. A., J. M. Landwehr, N. C. Matalas, and J. R. Wallis, Probability weighted moments: De nitions and relation to parameters of several distributions expressible in inverse form, Water Resources Research,

10
156, 1049 1054, 1979. Hurst, H. E., Long-term storage capacity of reservoirs, Transactions of the American Society of Civil Engineers, 116, 770 808, 1951. Hirsch, R. M., R. A. Alexander, and R. A. Smith, Selection of methods for the detection and estimation of trends in water quality, Water Resources Research, 27, 803 813, 1991. Hirsch, R. M., D. R. Helsel, T. A. Cohn, and E. J. Gilroy, Statistical treatment of hydrologic data, in Handbook of Hydrology, D. R. Maidment, editor, McGraw-Hill, Inc., 1993. Hirschboeck, K. K., Flood hydroclimatology, Flood Geomorphology, In V. R. Baker, R. C. Kochel, and P. C. Patton, editors, John Wiley & Sons, Inc., 1988. Hosking, J. R. M., The theory of probability weighted moments, Research Report RC-12210, IBM Research Division, T. J. Watson Research Center, Yorktown Heights, N. Y., 1986. Hosking, J. R. M., L-moments: Analysis and estimation of distributions using linear combinations of order statistics, Journal of the Royal Statistical Society, B, 522, 105 124, 1990. Hosking, J. R. M., and J. R. Wallis, The e ect of intersite dependence on regional ood frequency analysis, Water Resources Research, 244, 588 600, 1988. Hosking, J. R. M., and J. R. Wallis, Some statistics useful in regional frequency analysis, Water Resources Research, 292, 271 281, 1993. Hosking, J. R. M., and J. R. Wallis, Correction to Some statistics useful in regional frequency analysis," Water Resources Research, 311, 251, 1995. Hosking, J. R. M., and J. R. Wallis, Regional Frequency Analysis, Cambridge University Press, 1997. Kahya, E., and J. A. Dracup, U.S. stream ow patterns in relation to the El Ni~o Southern Oscillation, Water n Resources Research, 298, 2491 2503, 1993. Karl, T. R., R. W. Knight, D. R. Easterling, and R. G. Quayle, Indices of climate change for the United States, Bulletin of the American Meteorological Society, 772, 279 292, 1996. Kleme, V., The Hurst phenomenon: A puzzle?, Water s Resources Research, 10, 675 688, 1974. Knox, J. C., Fluvial response to small scale climate changes, In J. E. Costa and P. J. Fisher Eds. , Developments and Applications of Geomorphology, J. E. Costa and P. J. Fisher, editors, Springer-Verlag, Berlin, 318 342, 1984. Knox, J. C., Large increases in ood magnitude in response to modest changes in climate. Nature, 361, 430 432, 1993. Krasovskaia, I. and L. Gottschalk, Frequency of extremes and its relation to climate uctuations, Nordic Hydrology, 24, 1 12, 1993. Kunkel, K. E., S. A. Changnon, and J. R. Angel, Climatic aspects of the 1993 Upper Mississippi River Basin ood,
Bulletin of the American Meteorological Society, 755, 811 822, 1994. Lall, U., and M. E. Mann, The Great Salt Lake: A barometer of interannual climatic variability, Water Resources Research, 3110, 2503 2515, 1995. Laughlin, C. R., On the e ect of temporal sampling on the observation of mean rainfall, Precipitation Measurements from Space, Workshop report, edited by D. Altas and O. W. Thiele, NASA publication, Goddard Space Flight Center, Greenbelt, Maryland, 1981. Leathers, D. J., B. Yarnal, and M. A. Palecki, The Paci c North American teleconnection pattern and United States climate, Part I: Regional temperature and precipitation associations, Journal of Climate, 4, 517 528, 1991. Lettenmaier, D. P., J. R. Wallis, and E. F. Wood, E ect of regional heterogeneity on ood frequency estimation, Water Resources Research, 232, 313 324, 1987. Lettenmaier, D. P., E. F. Wood, and J. R. Wallis, Hydroclimatological trends in the continental United States, 1948 88, Journal of Climate, 74, 586 585, 1994. Lettenmaier, D. P., and K. W. Potter, Testing ood frequency estimation methods using a regional ood generation model, Water Resources Research, 21, 1903 1914, 1985. Lins, H. F., Stream ow variability in the United States: 1931-78, Journal of Climate and Applied Meteorology, 24, 463 471, 1985. Lisi, F., and V. Villi, Statistical considerations on the randomness of annual maximum daily rainfall. Journal of the American Water Resources Association, 332, 431442, 1997. Livezey, R. E. and W. Y. Chen, Statistical eld signi cance and its determination by Monte Carlo techniques, Monthly Weather Review, 111, 46 59, 1983. Mandelbrot, B. B., and J. R. Wallis, Noah, Joseph, and operational hydrology, Water Resources Research, 45, 909 918, 1968. Mann, M. E., U. Lall, and B. Saltzman, Decadal-tocentury scale climate variability: Insights into the rise and fall of the Great Salt Lake, Geophysical Research Letter, 22, 937 940, 1995. Mo, K., J. Nogues-Paegle, and J. Paegle, Physical mechanisms of the 1993 summer oods, Journal of the Atmospheric Sciences, 52, 879 895, 1994. Potter, K. W., Evidence for nonstationarity as a physical explanation of the Hurst phenomenon, Water Resources Research, 12, 1047 1052, 1976. Potter, K. W., Annual precipitation in the northeast United States: Long memory, short memory, or no memory?, Water Resources Research, 15, 340 346, 1979. Potter, K. W., Illustration of a new test for detecting a shift in mean in precipitation series, Monthly Weather Review, 1099, 2040 2045, 1981. Potter, K. W., and D. P. Lettenmaier, A comparison of

11
regional ood frequency estimation methods using a resampling method, Water Resources Research, 263, 415 424, 1990. Redmond, K. T., and R. W. Koch, Surface climate and stream ow variability in the western United States and their relationship to large-scale circulation indices, Water Resources Research, 27, 2381 2399, 1991. Ropelewski, C. F., and M. S. Halpert, Global and regional scale precipitation patterns associated with the El Ni~o Southern Oscillation, Monthly Weather Ren view, 115, 1606 1626, 1987. Smith, J. A., A. A. Bradley, and M. L. Baeck, The spacetime structure of extreme storm rainfall in the Southern Plains, Journal of Applied Meteorology, 3312, 1402 1417, 1994. Stedinger, J. R., and G. D. Tasker, Regional hydrologic regression, 1. Ordinary, weighted and generalized least squares compared, Water Resources Research, 219, 1421 1432, 1985. Stedinger, J. R., R. M. Vogel, and E. Foufoula-Georgiou, Frequency analysis of extreme events, in Handbook of Hydrology, D. R. Maidment, editor, McGraw-Hill, Inc., 1993. U. S. Water Resources Council, Guidelines for determining ood ow frequencies, Bulletin 17-B of the Hydrology Committee, 1982. Vogel, R. M., The probability plot correlation coe cient test for normal, lognormal, and Gumbel distributional hypotheses, Water Resources Research, 224, 587 590, 1987. Vogel, R. M., and D. E. McMartin, Probability plot goodness-of- t and skewness estimation procedures for the Pearson type III distribution, Water Resources Research, 2712, 3149 3158, 1991. Wall, D. J., and M. Englot, Correlation of annual peak ows for Pennsylvania Streams. Water Resources Bulletin, 213, 459 464, 1985. Waymire, E., V. K. Gupta, and I. Rodriguez-Iturbe, A spectral theory of rainfall intensity at the meso- scale, Water Resources Research, 20, 1453 1465, 1985. Webb, R. H., and J. L. Betancourt, Climatic variability and ood frequency of the Santa Cruz River, Pima County, Arizona, U.S. Geological Survey Water-Supply Paper 2379, 1992.

A. A. Bradley, Iowa Institute of Hydraulic Research and Department of Civil and Environmental Engineering, University of Iowa, Iowa City, IA 52252.

A This preprint was prepared with AGU's L TEX macros v4. File nal1 formatted November 23, 1997.

12

Figure 1. Southern Plains study area. Circles show the locations of the 43 hourly precipitation stations. Figure 2. Time series of the annual quantile estimates Zj p for exceedance probabilities of 0.5 2-year event

and 0.1 10-year event a 1-hour annual maximum precipitation b 24-hour annual maximum precipitation. The trends in the series are shown by the linear regression heavy solid line.

Figure 3. Correlation coe cients for 24-hour annual maximum precipitation versus distance between stations.

Crosses show the sample correlation coe cients. The dashed lines show the 5 signi cance levels for the test of the hypothesis that the correlation coe cient is zero. The heavy solid line shows the logarithmic correlation function assumed for the Monte Carlo simulations. k k Figure 4. The 5 signi cance levels for Smin and Smax based on a single-sided test for the 10-year 24-hour event. Dashed lines show signi cance levels assuming independence between sites. Solid lines show levels assuming k k a logarithmic intersite correlation function. Values of Smin below the levels, and values of Smax above the levels, occur by chance only 5 of the time for a random process.

Figure 5. Annual quantile series Zj 0:1 10-year event for the 24-hour precipitation line with circles and the

5-year moving average heavy line. The short dashed lines show the 5 signi cance levels for the 5-year moving average assuming independence between sites. The heavy dashed lines show the 5 signi cance levels assuming a logarithmic intersite correlation function.

Figure 6. Sample L-CV for 1-hour annual maximum precipitation for sites circles versus longitude. The trend
across the region is indicated by the linear regression heavy solid line.
k k Figure 7. The 5 signi cance levels for Smin and Smax based on a single-sided test for the 10-year 1-hour

event. Dashed lines show signi cance levels for a homogeneous region. Solid lines show levels for a heterogeneous region. Regional heterogeneity was modeled using a spatially varying L-CV estimated from the sample data. The k signi cance levels for Smin are nearly identical for the two cases.

Figure 8. Averages of 1 the number of days per year with an annual maximum solid line with circles, and 2 the maximum number of annual maxima for a single day in a year dashed line with squares. Averages are based on the 42-year study period. Figure 9. Annual precipitation time series for the Southern Plains region. The trend in annual precipitation is shown by the linear regression heavy solid line. The observed trend is not signi cant at the 5 level based on a test using Kendall's correlation coe cient. Figure 10. Monthly precipitation for the Southern Plains region. The monthly precipitation for 1952 solid line
with circles and 1956 dashed line with squares are compared to the 42-year average heavy solid line. precipitation. The sample correlation is 0.746. The solid line shows the linear regression.

Figure 11. Relationship between the annual quantiles Zj 0:5 2-year event for the 24-hour duration and annual

13

Table 1. Regional frequency analysis for the Southern Plains region. H is the hetero-

geneity measure. Highlighted values of H indicate that the region may be heterogeneous for the given duration. G is the goodness-of- t statistic for the generalized extreme value distribution GEV. The estimated parameters of the GEV distribution are , , and . The estimated quantiles of the regional distribution are qp. ^ Duration hour 1 2 3 6 12 24 48 72 H 3.21 1.63 1.32 0.46 -0.65 -1.65 -1.58 -0.57 G -1.50 -1.40 -1.15 -1.42 -1.61 -1.32 -1.45 -1.70 0.831 0.822 0.820 0.818 0.819 0.811 0.811 0.819 0.281 0.280 0.281 0.279 0.285 0.288 0.289 0.289 -0.025 -0.056 -0.059 -0.073 -0.055 -0.073 -0.073 -0.050
 q0:5 q0:25 q0:1 q0:05 ^ ^ ^ ^

.934 .926 .924 .921 .925 .918 .918 .926

1.186 1.183 1.184 1.181 1.186 1.187 1.187 1.189

1.481 1.493 1.498 1.499 1.501 1.516 1.517 1.504

1.697 1.726 1.734 1.742 1.738 1.767 1.768 1.739

Table 2. Signi cance levels  for Kendall's trend test for the non-dimensionalized annual quantile series. Trends that are signi cant at the 5 level or less are highlighted.
Duration hour 1 2 3 6 12 24 48 72 Return Period 2-yr 4-yr 10-yr 20-yr 44-yr 57.31 1.71 4.73 34.02 32.40 66.47 3.10 2.94 17.55 3.19 84.53 7.73 10.17 11.36 5.37 55.84 5.51 10.40 3.10 2.04 14.95 3.19 1.23 0.61 4.16 8.68 1.57 0.77 0.21 0.44 13.48 2.86 13.20 3.19 2.63 37.42 13.76 18.97 2.94 3.75

Table 3. Signi cance levels  for the test for nonrandomness for
k Smin 0:0227 44-year event. Statistics signi cant at the 10 level are

shown.

Duration hour 1 2 3 6 12 24 48 72

Moving Average Window 1-yr 3-yr 5-yr 10-yr 15-yr 20-yr 25-yr 30-yr 7.28 9.18 2.80 5.86 2.84 7.60 1.54 7.06 0.18 6.90 8.96

5.38 8.38

4.10 3.66

2.60 2.32

4.84 5.78

14

Table 4. Signi cance levels  for the test for nonrandomness for
shown. Duration Moving Average Window hour 1-yr 3-yr 5-yr 10-yr 15-yr 20-yr 25-yr 30-yr 1 4.20 2.48 2.22 2 7.78 6.32 1.96 3 7.54 6 8.64 4.08 12 9.42 6.12 24 8.66 7.24 5.84 2.40 48 72

k Srange 0:1 10-year event. Statistics signi cant at the 10 level are

Table 5. Signi cance levels  for the test for nonrandomness for
k Smin 0:1 10-year event. Statistics signi cant at the 10 level are

shown.

Duration hour 1 2 3 6 12 24 48 72

1-yr 8.84 2.18 0.76 9.40 9.58 6.46 1.52 0.80

Moving Average Window 3-yr 5-yr 10-yr 15-yr 20-yr 25-yr 30-yr 6.80 6.94 4.36 9.34 3.72 9.52 3.06 9.62 1.14 4.96 0.68 7.50 4.86

Table 6. Signi cance levels  for the test for nonrandomness for
k Smin 0:5 2-year event. Statistics signi cant at the 10 level are

shown.

Duration hour 1 2 3 6 12 24 48 72

1-yr 6.34 9.86 0.30 2.38 6.70 1.50 2.52 1.26

Moving Average Window 3-yr 5-yr 10-yr 15-yr 20-yr 25-yr 30-yr 1.32 6.04 0.60 0.62 0.24 1.10 0.20 4.68 0.72 2.34 0.18 8.04 8.22 6.92 1.82 3.18

15

Table 7. Sample correlation coe cient between annual quantile series Zj p and the annual precipitation. All the coe cients are statistically signi cant at the 5 level or less.
Duration hour 1 2 3 6 12 24 48 72 2-yr .687 .607 .673 .649 .698 .746 .691 .705 Return Period 4-yr 10-yr 20-yr 44-yr .566 .476 .520 .472 .637 .543 .512 .363 .754 .673 .461 .309 .708 .642 .591 .313 .690 .736 .641 .471 .728 .718 .553 .480 .656 .491 .416 .391 .631 .527 .400 .352

102W

100W

98W

96W

94W

38N

38N

36N

36N

34N

34N

32N
102W
Figure 1.

32N

100W

98W

96W

94W

Southern Plains study area. Circles show the locations of the 43 hourly precipitation stations.

2
2.2 2 1.8 1.6 Z (p) 1.4 1.2 1 0.8 0.6 50 60 70 Year 80 90
j

(a) 1-Hour Precipitation

2-year 10-year

2.2 2 1.8 1.6 Z (p) 1.4 1.2 1 0.8 0.6 50

Figure 2.

(b) 24-Hour Precipitation

2-year 10-year

60

70 Year
Z p

80

90

Time series of the annual quantile estimates j   for exceedance probabilities of 0.5 2-year event and 0.1 10-year event a 1-hour annual maximum precipitation b 24-hour annual maximum precipitation. The trends in the series are shown by the linear regression heavy solid line.

24-Hour Precipitation
0.6 Correlation Coefficient 0.4 0.2 0 -0.2 -0.4 0
Figure 3.

100

200 300 400 500 600 Intersite Distance (km )

700

800

Correlation coe cients for 24-hour annual maximum precipitation versus distance between stations. Crosses show the sample correlation coe cients. The dashed lines show the 5 signi cance levels for the test of the hypothesis that the correlation coe cient is zero. The heavy solid line shows the logarithmic correlation function assumed for the Monte Carlo simulations. 10-Year 24-Hour Precipitation
2.4 2 Sk(0.1) 1.6 1.2 0.8 0 5 10 15 20 k
S

No Intersi te Correl ati on Logari th mi c Correl ati on Model

Smax

Smin

25

30

35

40

Figure 4.

k k The 5 signi cance levels for min and max based on a single-sided test for the 10-year 24-hour event. Dashed lines show signi cance levels assuming independence between sites. Solid lines show levels assuming k k a logarithmic intersite correlation function. Values of min below the levels, and values of max above the levels, occur by chance only 5 of the time for a random process.
S S S

4
2 1.8 1.6 Z (0.1) 1.4 1.2 1 0.8 50
Figure 5.
Z :

10-Year 24-Hour Precipitation

60

70 Year

80

90

Annual quantile series j 0 1 10-year event for the 24-hour precipitation line with circles and the 5-year moving average heavy line. The short dashed lines show the 5 signi cance levels for the 5-year moving average assuming independence between sites. The heavy dashed lines show the 5 signi cance levels assuming a logarithmic intersite correlation function. 1-Hour Precipitation
0.3

0.25 L-CV

0.2

0.15

0.1 102
Figure 6.

101

100

99 98 97 Longitude (W )

96

95

94

Sample L-CV for 1-hour annual maximum precipitation for sites circles versus longitude. The trend across the region is indicated by the linear regression heavy solid line.

10-Year 1-Hour Precipitation

2.2 2 Sk(0.1) 1.8 1.6 1.4 1.2 1 0
Figure 7.

Homogeneous Regi on Regi onal Heterogenei ty

Smax

Smin 5 10 15 20 k
S

25

30

35

40

k k The 5 signi cance levels for min and max based on a single-sided test for the 10-year 1-hour event. Dashed lines show signi cance levels for a homogeneous region. Solid lines show levels for a heterogeneous region. Regional heterogeneity was modeled using a spatially varying L-CV estimated from the sample data. The k signi cance levels for min are nearly identical for the two cases. Annual Maxima Averages
S S

35 30 25

Num ber

20 15 10 5 0 1 2

Number of days per year wi th an annual maxi mum Maxi mum number per year of annua l maxi ma i n a si ngl e day

3 6 12 24 Duration (hours)

48

72

Averages of 1 the number of days per year with an annual maximum solid line with circles, and 2 the maximum number of annual maxima for a single day in a year dashed line with squares. Averages are based on the 42-year study period.
Figure 8.

6
45 40 Precipitation (in) 35 30 25 20 15 1950
Figure 9.

Annual Precipitation

1960

1970 Year

1980

1990

Annual precipitation time series for the Southern Plains region. The trend in annual precipitation is shown by the linear regression heavy solid line. The observed trend is not signi cant at the 5 level based on a test using Kendall's correlation coe cient. Southern Plains
5 4 3 2 1 0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Precipitation (inches)

Av erage 1952 1956

Month

Figure 10.

Monthly precipitation for the Southern Plains region. The monthly precipitation for 1952 solid line with circles and 1956 dashed line with squares are compared to the 42-year average heavy solid line.

7
1.3 1.2 1.1 1 Z (0.5) 0.9 0.8 0.7 0.6 0.5 20
Figure 11.

24-Hour Precipitation

25 30 35 40 Annual Precipitation (in)

Z :

45

Relationship between the annual quantiles j 0 5 2-year event for the 24-hour duration and annual precipitation. The sample correlation is 0.746. The solid line shows the linear regression.