Guidelines
For
Determining
Bulletin7#VB
of the
Hydrology Subcommittee
Revised September 1981
Editorial Corrections March 1982
INTERAGENCY ADVISORY COMMITTEE
ON WATER DATA
US. Department of the interior
Geological Survey
Office of Water Data Coordination
Reston, Virginia 22092
FOREWORD
An accurate estimate of the flood damage potential is a key element to
an effective, nationwide flood damage abatement program. Further, there is
an acute need for a consistent approach to such estimates because management
of the nation's water and related land resources is shared among various
levels of government and private enterprise. To obtain both a consistent
and accurate estimate of flood losses requires development, acceptance, and
widespread application of a uniform, consistent and accurate technique for
determining floodflow frequencies.
In a pioneer attempt to promote a consistent approach to floodflow
frequency determination, the U.S. Water Resources Council in December 1967
published Bulletin No. 15, "A Uniform Technique for Determining Flood Flow
Frequencies." The technique presented therein was adopted by the Council
for use in all Federal planning involving water and related land resources.
The Council also recommended use of the technique by::State, local government,
and private organizations. Adoption was based upon the clear understanding
that efforts to develop methodological improvements in the technique would
be continued and adopted when appropriate.
An extension and update of Bulletin No. 15 was published in March 1976
as Bulletin No. 17, "Guidelines for Determining Flood Flow Frequency." It
presented the currently accepted methods for analyzing peak flow frequency
data at gaging stations with sufficient detail to promote uniform applica
tion. The guide was a synthesis of studies undertaken to findmethod
ological improvements and a survey of existing literature on peak flood
flow determinations.
* The present guide is the second revision of the original publication
*
i
*
and improves the methodologies. It revises and expands some of the *
techniques in the previous editions of this Bulletin and offers a further
explanation of other techniques. It is the result of a continuing effort
to develop a coherent set of procedures for accurately defining flood
potentials. Much additional study is required before the two goals
of accuracy and consistency will be fully attained. All who are interested
in improving peak floodflow frequency determinations are encouraged
to submit comments, criticism and proposals to the Office of Water
Data Coordination for consideration by the Ilydroloqv Subcommittee.
Federal agencies are requested to use these guidelines in all planning
activities involving water and related land resources. State, local
and private organizations are encouraged to use these guidelines also
to assure more uniformity, compatibility, and comparability in the frequency
values that all concerned agencies and citizens must use for many vital
decisions.
This present revision is adopted with the knowledge and understanding
T hat review of these procedures will continue. When warranted by experience
*
and by examination and testing of new techniques, other revisions will
be publlshed.
ii
Ll
HYDROLOGY SUBCOMMITTEE
*
Member
Robert E. Rallison
Robert G. Delk
Walter J. Rawls
Agency
Soi 1 Conservation
Forest Service
Science Education
Administration
Service
Department
AgricultureII
II
Vernon
Roy G.
K. Hagen
Huffman
Corps of EngineersII
Arty
Allen
John
F. Flanders
F. Wilier
NOAA, National
ServiceII
Weather Commerce
II
Truman Goins Housing and Urba
Development
Porter Ward
David F. Gudgel
Don Willen
Ewe11 H. Mohler, Jr.
Geological Survey
Bureau of Reclamation
Office of Surface Mining
Office of Water Research
and Technology
InteriorII
II
it
Sidney J, Spiegel
Pat Uiffy
Leo Fake
Victor Berte'
Irene L. Murphy
Bureau of Indian Affairs
Bureau of Mines
Fish and Wildlife Service
National Park Service
Heritage, Conservation and
Recreation Service
D. c.
Philip
woo
L. Thompson
Federal HighwayII
Administrat on TransportationII
Timothy
Robert
Stuart
Horn
Environmental
ProtectionII
Agent
Steve
Patrick
Parker
Jefferson
Federal Energy
Regulatory
CommissionII
Brian
William
Mrazik
S. Bivins
Federal Emergent
Management Agent,
I,
Edward F. Hawkins Nuclear Regula
tory Commission
. . .
111
*
Member
Donald W. Newton
Larry 14. Richardson
Ron Scullin
.Y
HYDROLOGY SUBCOMMITTEE w,. .
Agency
Con't
Department
Tennessee Valley
Authocity
Water Resources
Council
Member
Roger Cronshey
Roy G. Huffman
John F. Miller*
WORK GROUP ON REVISION OF BULLETLN 17
Agency
Soil Conservation Service
Corps of Engineers
NOAA, National Weather
Service
Department
Agriculture
Army
Commerce
William H. Kirby
Wilbert 0. Thomas,
Frederick A. Bertle
Donald W. Newton
Jr.
GeologicalII
Bureau of
Survey
Reclamation
Interior6,
11
Tennessee
Authority
Valley
*
* Chairman
YMembership as of September 1981
iV
The following pages contain revisions from material presented in
"Guidelines for Determining Flood Flow Frequency."
1, 4, 82, and 131
The revised material is included on the lines enclosed by the +
The following pages of Bulletin 17 have been deleted:
132 through 1335
The following pages contain revisions from the material in either
Bulletin 17 or 17A,
i, ii, iii, iv, v, vi, vii, 1, 3, 10, 11, 12, 13, 14, 15, 17, 18, 19,
20, 26, 71, l2, l3, l4, 23, 27, 28, 41, 51, 52, 53, 54,
61, 62, 63, 65, 66, 67, 71, 72, 73, 74, 75, 76, 77, 78,
79, 9l through 910, 101, 102, 103, 122 through 1237 and 141
The revised material is included on the lines enclosed by the *
The following page of Bulletin 77 and 77A has been deleted from 17B:
Editorial corrections to Bulletin 17B were incorporated into this
V
CONTENTS
Page
46
Foreword. . . . . ..I.......................................
Hydrology Subcommittee .,...............................
i . . .
111
Page Revisions to Bulletin 17 and 17A .................. V *
I. Introduction ...........................................
II. Summary ................................................
A. Information to be Evaluated ........................
B. Data Assumptions ...................................
* C. Determination of the Frequency Curve ...............
D. Reliability Applications ...........................
E. Potpourri ..........................................
F. Appendix ...........................................
III. Information to be Evaluated ............................
A. Systematic Records .................................
B. Historic Data ......................................
C. Comparisons with Similar Watersheds ................
D. Flood Estimates From Precipitation .................
IV. Data Assumptions .......................................
A. Climatic Trends ....................................
B. Randomness of Events ...............................
C. Watershed Changes ..................................
D. Mixed Populations ..................................
E. Reliability
V. Determination
A. Series
B. Statistical
The
of Flow Estimates ......................
of Frequency Curve .......................
Selection ...................................
Treatment ..............................
Distribution ...............................
:: Fitting
3. Estimating
* 4. Weighting
Broken
the Distribution .......................
Generalized Skew ....................
the Skew Coefficient .................
Record ..................................
5. Incomplete Record
7: Zero Flood Years .............................................................
a. Mixed Populations ..............................
9. Outliers .......................................
10. Historic Flood Data ............................
C. Refinements to Frequency Curve .....................
1. Comparisons with Similar Watersheds ............
2. Flood Estimates From Precipitation .............
vi
VI. Reliability Application ................................. pa%
22
A. Confidence Limits ................................... 23
* 6. Risk ................................................ 24
C. Expected Probability ................................ 24
VII. Potpourri ...............................................
A. Nonconforming Special Situations ...................
25
25
B. Plotting Position ................................... 26
C. Future Studies ...................................... 27
*
Appendices
1. References .................................................
2. Glossary and Notation ......................................
3. Table of K Values .........................................
+) 4. Outlier Test K Values ...................................
5. Conditional Probability Adjustment .........................
6. Historic Data ..............................................
7. TwoStation Comparison ..................................... 71
8. Weighting of Independent Estimates ......................... 8l
Confidence Limits ..........................................
1:: Risk ....................................................... 1:~;
Expected Probability ....................................... 11l
1:: Flow Diagram and Example Problems .......................... 12l
13. Computer Program ........................................... 13l
14. "Flood Flow Frequency Techniques" Report Summary ........... 141
vii
I. Introduction
In December 1967, Bulletin No. 15, 'A Uniform Technique for Determining
Flood Flow Frequencies," was issued by the Hydrology Committee of the
Water Resources Council, The report recommended use of the Pearson Type
III distribution with log transformation of the data (logPearson Type
III distribution) as a base method for flood flow frequency studies.
As pointed out in that report, further studies were needed covering various
aspects of flow frequency determinations.
+ In March 1976, Bulletin 17, "Guidelines for Determining Flood Flow
Frequency" was issued by the Water Resources Council. The guide was an
extension and update of Bulletin No. 15. It provided a more complete
guide for flood flow frequency analysis incorporating currently accepted
technical methods with sufficient detail to promote uniform application.
It was limited to defining flood potentials in terms of peak discharge
and exceedance probability at locations where a systematic record of peak
flood flows is available. The recommended set of procedures was selected
from those used or described in the literature prior to 1976, based on
studies conducted for this purpose at the Center for Research in Water
Resources of the University of Texas at Austin (summarized in Appendix
14) and on studies by the Work Group on Flood Flow Frequency. +
=i% The "Guidelines" were revised and reissued in June 7977 as Bulletin
17A. Bulletin 17B is the latest effort to improve and expand upon the
earlier publications. Bulletin 17B provides revised procedures for weighting
a station skew value with the results from a generalized skew study, detect
ing and treating outliers, making two station comparisons, and computing con
fidence limits about a frequency curve. The Work Group that prepared this
revision did not address the suitability of the orlginal distribution
or ,the generalized skew map. #
Major problems are encountered when developing guides for flood flow
frequency determinations. There is no procedure or set of procedures that
can be adopted which, when rigidly applied to the available data, will
accurately define the flood potential of any given watershed. Statistical
analysis alone will not resolve all flood frequency problems. As discussed
in subsequent sections of this guide, elements of risk and uncertainty
are inherent in any flood frequency analysis. User decisions must be
based on properly applied procedures and proper interpretation of results
considering risk and uncertainty. Therefore, the judgment of a profes
sional experienced in hydrologic analysis will enhance the usefulness
of a flood frequency analysis and promote appropriate application.
It is possible to standarize many elements of flood frequency analysis,
This guide describes each major element of the process of defining the
. . ..^ . _
flood potential at a specific location in terms of peak discharge and
exceedance probability. Use is confined to stations where available
records are adequate to warrant statistical analysis of the data. Special
situations may require other approaches. In those cases where the proce
dures of this guide are not followed, deviations must be supported by
appropriate study and accompanied by a comparison of results using the
recommended procedures.
As a further means of achieving consistency and improving results,
the Work Group recommends that studies be coordinated when more than
one analyst is working currently on data for the same location. This
recommendation holds particularly when defining exceedance probabilities
for rare events, where this guide allows more latitude.
Flood records are limited, As more years of record become available
at each location, the determination of flood potential may change.
Thus, an estimate may be outdated a few years after it is made. Additional
flood data alone may be sufficient reason for a fresh assessment of
the flood potential. When making a new assessment, the analyst should incor
porate in his study a review of earlier estimates. Where differences
appear, they should be acknowledged and explained.
I I. Summary
This guide describes the data and procedures for computing flood
flow frequency curves where systematic stream gaging records of sufficient
length (at least 10 years) to warrant statistical analysis are available
as the basis for determination. The procedures do not cover watersheds
2
where flood flows are appreciably altered by reservoir regulation or
where the possibility of unusual events, such as dam failures, must be
considered. The guide was specifically developed for the treatment of
annual flood peak discharge. It is recognized that the same techniques
‘could also be used to treat other hydrologic elements, such as flood
volumes. Such applications, however, were not evaluated and are not
intended.
The guide is divided into six broad sections which are summarized
below:
A. Information to be Evaluated
The following categories of flood data are recognized: systematic
records, historic data, comparison with similar watersheds, and flood
estimates from precipitation. Mow each can be used to define the flood
potential is briefly described.
13. Data Assumptions
A brief discussion of basic data assumptions is presented as a reminder
to those developing flood flow frequency curves to be aware of potential
data errors. Natural trends, randomness of events, watershed changes,
mixed populations , and reliability of flow estimates are briefly discussed.
c. Determination of the Frequency Curve
This section provides the basic guide for determination of the fre
quency curve. The main thrust is determination of the annual flood series.
Procedures are also recommended to convert an annual to partialduration
flood series.
The Pearson Type III distribution with log transformation of the
flood data (logPearson Type III) is recommended as the basic distribution
for defining the annual flood series. The method of moments is used to de
termine the statistical parameters of the distribution from station data.
4Generalized relations are used to modify the station skew coefficient.
Methods are proposed for treatment of most flood record problems encoun
‘%ered. Proce dures are described for refining the basic curve determined
from statistical analysis of the systematic record and historic flood data
to incorporate information gained from comparisons with similar watersheds ~
and flood estimates from precipitation.
3
uo Kellabl Ilt;y AppllCatlOnS
Procedures for computing
provided along with those for
ability adjustments.
confidence
calculating
limits
risk and
to the
for
frequency curve
making expected
are
prob
E. Potpourri
This section provides infoguide, including a discussion
positions, and suggested future
rmation
of nonconstudies.
of
forming
interest
special
but not essential
situations,
to
plotting
the
F. Appendix
The appendix provides a list of references, a glossary and list of
ymbols, tables of K values, the computational details for treating most
B
of the recommended procedures , information about how to obtain a computer
program for handling the statistical analysis and treatment of data, and a c
summary of the report ("Flood Flow Frequency Techniques") describing studies
made at the University of Texas which guided selection of some of the pro
cedures proposed,
III, Information to be Evaluated
When developing a flood flow frequency curve, the ana7yst should con
sider all available information. The four general types of data which can
be included in the flood flow frequency analysis are described in the follow
ing paragraphs. Specific applications are discussed in subsequent sections.
A. Systematic Records
Annual peak discharge information is observed systematically by many
Federal and state agencies and private enterprises. Most annual peak
records are obtained either from a continuous trace of river stages or from
periodic observations of a creststage gage. Creststage records may provide
information only on peaks above some preselected base. A major portion of
these data are available in U.S. Geological Survey (USGS) Water Supply
Papers and computer files, but additional information in published or
unpublished form is available from other sources.
A statistical analysis of these data is the primary basis for the
determination of the flow frequency curve for each station.
B. Historic Data
At many locations, particularly where man has occupied the flood
plain for an extended period, there is information about major floods
which occurred either before or after the period of systematic data
collection, This information can often be used to make estimates of
peak discharge. It also often defines an extended period during which
the largest floods, either recorded or historic, are known. The USGS
includes some historic flood information in its published reports and
computer files. Additional information can sometimes be obtained from
the files of other agencies or extracted from newspaper files or by
intensive inquiry and investigation near the site for which the flood
frequency information is needed,
Historic flood information should be obtained and documented
whenever possible, particularly where the systematic record is relatively
short. Use of historic data assures that estimates fit community experi
ence and improves the frequency determinations.
C. Comparison With Similar Watersheds
Comparisons between computed frequency curves and maximum flood
data of the watershed being investigated and those in a hydrologically
similar region are useful for identification of unusual events and for
testing the reasonableness of flood flow frequency determinations.
Studies have been made and published [e.g., (l), (2), (3), (4)1* which
permit comparing flood frequency estimates at a site with generalized
estimates for a homogeneous region. Comparisons with information at
stations in the immediate region should be made, particularly at gaging
stations upstream and downstream, to promote regional consistency and
help prevent gross errors.
*Numbers in parentheses refer to numbered references in Appendix 1.
5
D. Flood Estimates From Precipitation
Flood discharges estimated from climatic data (rainfall and/or
snowmelt) can be a useful adjunct to direct streamflow measurements.
Such estimates, however, require at least adequate climatic data and a
valid watershed model for converting precipitation to discharge.
Unless such models are already calibrated to the watershed, considerable
effort may be required to prepare such estimates.
Whether or not such studies are made will depend upon the availabilit,
of the information, the adequacy of the existing records, and the exceedar
probability which is most important,
IV. Data Assumptions
Necessary assumptions for a statistical analysis are that the array
of flood information is a reliable and representative time sample of
random homogeneous events. Assessment of the adequacy and applicability
of flood records is therefore a necessary first step in flood frequency
analysis, This section discusses the effect of climatic trends, randomnes
of events, watershed changes, mixed populations, and reliability of flow
estimates on flood frequency analysis.
A. Climatic Trends
There is much speculation about climatic changes. Available
evidence indicates that major changes occur in time scales involving
thousands of years. In hydrologic analysis it is conventional to
._. .
assume flood flows are not affected by climatic trends or cycles.
Climatic time invariance was assumed when developing this guide.
B. Randomness of Events
In general, an array of annual maximum peak flow rates may be
considered a sample of random and independent events, Even when statis
tical tests of the serial correlation coefficients indicate a significant
deviation from this assumption, the annual peak data may define an unbiase
estimation of future flood activity if other assumptions are attained.
The nonrandomness of the peak series will, however, increase the degree
6
of uncertainty in the relation; that is, a relation based upon nonrandom
data will have a degree of reliability attainable from a lesser sample
of random data (5), (6).
C. Watershed Changes
It is becoming increasingly difficult to find watersheds in which
the flow regime has not been altered by man's activity. Man's activities
which can change flow conditions include urbanization, channelization,
levees, the construction of reservoirs, diversions, and alteration of
cover conditions.
Watershed history and flood records should be carefully examined to
assure that no major watershed changes have occurred during the period of
record. documents which accompany flood records often list such changes.
All watershed changes which affect record homogeneity, however, might
not be listed; unlisted, for instance, might be the effects of urbaniza
tion and the construction of numerous small reservoirs over a period of
several years. Such incremental changes may not significantly alter the
flow regime from year to year but the cumulative effect can after several
years.
Special effort should be made to identify those records which are
not homogeneous. Only records which represent relatively constant
watershed conditions should be used for frequency analysis.
13. Mixed Populations
At some locations flooding is created by different types of events.
For example, flooding in some watersheds is created by snowmelt, rainstorms,
or by combinations of both snowmelt and rainstorms. Such a record may
not be homogeneous and may require special treatment.
E. Reliability of Flow Estimates
Errors exist in streamflow records, as in all other measured
values. Errors in flow estimates are generally greatest during maximum
flood flows. Measurement errors are usually random3 and the variance
introduced is usually small in comparison to the yeartoyear variance
in flood flows. The effects of measurement errors, therefore, may
7
normally be neglected in flood flow frequency analysis. Peak flow
estimates of historic floods can be substantially in error because of the
uncertainty in both stage and stagedischarge relationships.
At times errors will be apparent or suspected. If substantial, the
errors should be brought to the attention of the data collecting agency
with supporting evidence and a request for a corrected value, A more
complete discussion of sources of error in streamflow measurement is
found in (7).
V. Determination of Frequency Curve
A. Series Selection
Flood events can be analyzed using either annual or partialduration
series. The annual flood series is based on the maximum flood peak for
each year. A partialduration series is obtained by taking all flood
peaks equal to or greater than a predefined base flood.
If more than one flood per year must be considered, a partial
duration series may be appropriate. The base is selected to assure that
all events of interest are evaluated including at least one event per
time period. A major problem encountered when using a partialduration
series is to define flood events to ensure that all events are independent,
It is common practice to establish an empirical basis for separating
flood events. The basis for separation will depend upon the investigator
and the intended use. No specific guidelines are recommended for defining
flood events to be included in a partial series.
A study (8) was made to determine if a consistent relationship
existed between the annual and partial series which could be used&to
convert from the annual to the partialduration series. Based on this
study as summarized in Appendix 14, the Work Group recommends that the
partialduration series be developed from observed data. An alternative
but less desirable solution is to convert from the annual to the partial
duration series. For this, the first choice is to use a conversion
factor specifically developed for the hydrologic region in which the
8
gage is located. The second choice is to use published relationships
[e.g., WI
l
Except for the preceding discussion of the the partialduration
series, the procedures described in this guide apply to the annual flood
series.
6. Statistical Treatment
1. The DistributionFlood events are a succession of natural
events which, as far as can be determined, do not fit any one specific
known statistical distribution. To make the problem of defining flood
probabilities tractable it is necessary, however, to assign a distribution.
Therefore, a study was sponsored to find which of many possible distribu
tions and alternative fitting goods mr&l best m@t the purposes of this
guide. This study is summarized in Appendix 14. The Work Group concluded
from this and other studies that the Pearson Type III distribution with
log transformation of the data (logPearson Type III distribution)
should be the base method for analysis of annual series data using a
generalized skew coefficient as described in the following section.
2. Fitting the DistributionThe recommended technique for fitting
a logPearson Type III distribution to observed annual peaks is to
compute the base 10 logarithms of the discharge, Q, at selected exceedance
probability, P, by the equation:
Log Q=X+KS (1)
where x and S are as defined below and K is a factor that is a function
of the skew coefficient and selected exceedance probability. Values of
K can be obtained from Appendix 3.
The mean, standard deviation and skew coefficient of station data
may be computed using the following equations:
S = (34
0.5
= [ (XX;; 1fX)'/N ]
(W
13
G = NX(XX (44
(N l)(N 2)S3
= N*( ZX3) 3N(C X)(X X2) I 2Lr: Xl3 (4b)
N(Nl)(N2)S3
in which:
X = logarithm of annual peak flow
N = number of items in data set
x = mean logarithm
S = standard deviation of logarithms
G = skew coefficient of logarithms
Formulas for computing the standard errors for the statistics x, S,
and G are given in Appendix 2. The precision of values computed with
equations 3b and 4b is more sensitive than with equations 3a and 4a
to the number of significant digits used in their calculation, When
the available computation facilities only provide for a limited number
of significant digits, equations 3a and 4a are preferable.
* 3. Estimating Generalired SkewThe skew coefficient of the station
record (station skew) is sensitive to extreme events; thus it is difficult
to obtain accurate skew estimates from small samples. The accuracy of the
estimated skew coefficient can be improved by weighting the station skew
with generalized skew estimated by pooling information from nearby sites.
The following guidelines are recommended for estimating generalized skew.&
10
Guidelines on weighting station and generalized skew are provided in the
next section of this bulletin,
The recommended procedure for developing generalized skew coefficients
requires the use of at least 40 stations, or all stations within a JOO
mile radius. The stations used should have 25 or more years of record.
It is recognized that in some locations a relaxation of these criteria
may be necessary. The actual procedure includes analysis by three methods:
1) skew isolines drawn on a map; 2) skew prediction equation; and 3)
the mean of the station skew values. Each of the methods are discussed
separately.
To develop thecysoline map, plot each station skew value at the cen
._
troid of its drainage basin and examine the plotted data for any geographic
or topographic trends. If a pattern is evident, then isolines are drawn
and the average of the squared differences between observed and isoline
values, meansquare error (MSE), is computed. The MSE will be used in
appraising the accuracy of the isoline map. If no pattern is evident,
then an isoline map cannot be drawn and is therefore, not further considered.
A, prediction equation should be developed that would relate either
the station skew coefficients or the differences from the isoline map
to predictor variables that affect the skew coefficient of the station
record. These would include watershed and climatologic variables. The
prediction equation should preferably be used for estimating the skew
coefficient at stations with variables that are within the range of data
used to calibrate the equation. The MSE (standard error of estimate
squared) will be used to evaluatethe accuracy of the preciction equation.
Determine the arithmetic mean and variance of the skew coefficients
for all stations. In some cases the variability of the runoff regime
may be so large as to preclude obtaining 40 stations with reasonably
homogeneous hydrology. In these situations, the arithmetic mean and
variance of about 20 stations may be used to estimate the generalized
skew coefficient. The drainage areas and meteorologic, topographic, and
geologic characteristics should be representative of the region around
the station of interest.
Select the method that provides the most accurate skew coefficient
11
* estimates. Compare the MSE from the isoline map to the MSE for the pre
diction equation. The smaller MSE should then be compared to the variance
of the data. If the MSE is significantly smaller than the variance, the
method with the smaller MSE should be used and that MSE used in equation 5
for MSEc If the smaller MSE is not significantly smaller than the vari
ance, neither the isoline map nor the prediction equation provides a
more accurdte estimate of the skew coefficient than does the mean Vale.
The mean skew coefficient should be used aS 'it provides tne most accurate
estimate and the variance should be used in equation 5 for aEgo
In the absence of detailed studies the generalized skew (c) can be
read from Plate I found in the flyleaf pocket of this guide. This map
of generalized skew was developed when this bulletin was first introduced
and has not been changed. The procedures used to develop the statistical
analysis for the individual stations do not conform in all aspects to
the procedures recommended in the current guide. However, Plate I is
still considered an alternative for use with the guide for those who prefer
not to develop their own generalized skew procedures.
The accuracy of a regional generalized skew relationship is generally
not comparable to Plate I accuracy. While the average accuracy of Plate I
is given, the accuracy of subregions within the United States are not
given. A comparison should only be made between relationships that cover
approximately the same geographical area. Plate I accuracy would be
directly comparable to other generalized skew relationships that are
applicable to the entire country.
4. Weighting the Skew CoefficientThe station and generalized
skew coefficient can be combined to form a better estimate of skew for
a given watershed. Under the assumption that the generalized skew is
unbiased and independent of station skew, the meansquare error (MSE)
of the weighted estimate is minimized by weighting the station and
generalized skew in inverse proportion to their individual meansquare
errors. This concept is expressed in the following equation adopted
from Tasker (39) which should be used in computing a weighted skew co
efficient:
MS+(G) + MSEC(Q
Gw =
MSEg + MSEC
12
i( where Gw = weighted skew coefficient
G = station skew
% = generalized skew
MSEc = meansquare error of generalized skew
= meansquare error of station skew
MSEG
Equation 5 can be used to compute a weighted skew estimate regardless
of the source of generalized skew, provided the MSE of the generalized
skew can be estimated. When generalized skews are read from Plate I,
the value of MSEc = 0.302 should be used in equation 5. The MSE of the
station skew for logPearson Type III random variables can be obtained
from the results of Monte Carlo experiments by Wallis, Matalas, and Slack
(40). Their results show that the MSE of the logarithmic station skew
is a function of record length and population skew. For use in calculat
ing Gwl this function (MSEG) can be approximated with sufficient
accuracy by the equation:
[A B ~Lw10(W10)9]
MSEG"10 (6)
Where A = 0.33 f O.OSlGl if IGI LO.90
0.52 f 0.3OlGI if IGi >0.90
B = 0.94 0.26IGI if IGI 51.50
0.55 if IGJ >I.50
in which IGJ is the absolute value of the station skew (used as an
estimate of population skew) and N is the record length in years. If
the historic adjustment described in Appendix 6 has been applied, the
historically adjusted skew,%, and historic period, H, are to be used
for G and N, respectively, in equation 6. For convenience in manual
computations, equation 6 was used to produce table 1 which shows MSEG
values for selected record lengths and station skews.
13
TABLE 1, SUbiMARY OF MEAN SQUARE ERROR OF STATION SKEW AS A FUNCTIONOF RECORD LENGTH AND STATION SKEW. Jt
0.6
0.1
3.2
(3:; 3
0
0 . 5
0 l et
0.7
3c
it.9
1 l 0
1.1
z 12
I.3
1.4
1:; F
1
1.7
1.E
I.9
2.0
2. f
2.2
2.3
2.4
2.5
2.6
2.7
2 .P
2.9
3.0
*
* Application of equation 6 and table 1 to stations with absolute skew
values (logs) greater than 2 and long periods of record gives relatively
little weight to the station value. Application of equation 5 may also
give improper weight to the generalized skew if the generalized and station
skews differ by more than 0.5. In these situations, an examination of
the data and the floodproducing characteristics of the watershed should
be made and possibly greater weight given to the station skewe *
5. Broken RecordAnnual peaks for certain years may be missing
because of conditions not related to flood magnitude, such as gage
removal. In this case, the different record segments are analyzed as
a continuous record with length equal to the sum of both records, unless
there is some physical change in the watershed between segments which may
make the total record nonhomogeneous.
6. Incomplete RecordAn incomplete record refers to a streamflow
record in which some peak flows are missing because they were too low
or too high to record, or the gage was out of operation for a short
period because of flooding. Missing high and low data require different
treatment.
When one or more high annual peaks during the period of systematic
record have not been recorded, there is usually information available
from which the peak discharge can be estimated. In most instances the
data collecting agency routinely provides such estimates. If not, and
such an estimate is made as part of the flood frequency analysis, it
should be documented and the data collection agency advised.
At some crest gage sites the bottom of the gage is not reached
*in some years. For this situation use of the conditional probability
+e
adjustment is recommended as described in Appendix 5.
7. Zero Flood YearsSome streams in arid regions have no flow
for the entire year. Thus, the annual flood series for these streams
will have one or more zero flood values. This precludes the normal
statistical analysis of the data using the recommended logPearson Type III
++distribution because the logarithm bf zero is minus infinity. The condi
tional probability adjustment is recommended for determining frequency
curves for records with zero flood years as described in Appendix 5. #
15
8. Mixed PopulationFlooding in some w,atersheds is created by
different types of events. This results in flood frequency curves with
abnormally large skew coefficients reflected by abnormal slope changes
when plotted on logarithmic normal probability paper. In some situations
the frequency curve of annual events can best be described by computing
separate curves for each type of event. The curves are then combined.
Two examples of combinations of different types of floodproducing
events include: (1) rain with snowmelt and (2) intense tropical storms
with general cyclonic storms. Hydrologic factors and relationships oper
ating during general winter rain flood are usually quite different from
those operating during spring snowmelt floods or during local summer
cloudburst floods. One example of mixed population is in the Sierra
Nevada region of California. Frequency studies there have been made
separately for rain floods which occur principally during the months
of November through March, and for snowmelt floods, which occur during
the months of April through July. Peak flows were segregated by cause
those predominately caused by snowmelt and those predominately caused
by rain. Another example is along the Atlantic and Gulf Coasts, where
in some instances floods from hurricane and nonhurricane events have
been separated, thereby improving frequency estimates.
When it can be shown that there are two or more distinct and genera
independent causes of floods it may be more reliable to segregate the
flood data by cause, analyze each set separately, and then to combine
the data sets using procedures such as described in (11). Separation
by calendar periods in lieu of separation by events is not considered
hydrologically reasonable unless the events in the separate periods are
clearly caused by different hydrometeorologic conditions. The fitting
procedures of this guide can be used to fit each flood series separately
with the exception that generallzed skew coefficients cannot be used
unless developed for specific type events being examined.
tions
ally
one
If the
cannot
meaningful
population.
flood
be
events
identified
criterion,
that are believed
and separated
the record shall
by
to comprise
an objective
be treated
two
and
as
or
coming
more popul
hydrologic
from
16
Ifi9. OutliersOutliers are data points which depart significantly
from the trend of the remaining data, The retention, modification,
deletion of these outiiers can significantly affect the statistical
parameters computed from the data, especially for small samples. All
procedures for treating outliers ultimately require judgment involving
both mathematical and hydrologic considerations. The detection and
treatment of high and low outliers are described below, and are outlined
on the flow chart in Appendix 12 (figure 123),
If the station skew is greater than +0.4, tests for high outliers
are considered first, If the station skew is less than 0.4 tests for
low outliers are considered first. Where the station skew is between
2 0.4, tests for both high and low outliers should be applied before
eliminating any outliers from the data set,
The following equation is used to detect high outliers:
xH
= x + KWS (7)
where XH = high outlier threshold in log units
x = mean logarithm of systematic peaks (X's) excluding zero flood
events9 peaks below gage base, and outliers previously
detected.
S = standard deviation of X's
= K value from Appendix 4 for sample size N
KN
If the logarithms of peaks in a sample are greater than XH in equation
7 then they are considered high outliers. Flood peaks considered high
outliers should be compared with historic flood data and flood information
at nearby sites. If information is available which indicated a high
outlier(s) is the maximum in an extended period of time, the outlier(s)
is treated as historic flood data as described in Section V.B,lO. If
useful hl'storic information is not available to adjust for high outliers,
then they should be retained as part of the systematic record. The treat
ment of all historic flood data and high outliers should be well documented
in the analysis.
*
17
* The following equation is used to detect low outliers:
XL = x KNS h4
where XL = low outlier threshold in log units and the other terms are a:
defined for equation 7.
If an adjustment for historic flood dai;a has previously been made,
then the following equation is used to detect low outliers:
=x KH: (8b)
xL
where XL = low outlier threshold in log units
= K value from Appendix 4 for period used to compute% and?
KH
% = historically adjusted mean logarithm
Y = historically adjusted standard deviation
If the logarithms of any annual peaks in a sample are less than XL in
equation 8a or b, then they are considered low outliers. Flood peaks
considered low outliers are deleted from the record and the conditional
probability adjustment described in Appendix 5 is applied.
If multiple values that have not been identified as outliers by th
recommended procedure are very close to the threshold value, it may be
desirable to test the sensitivity of the results to treating these valu
as outliers.
Use of the K values from Appendix 4 is equivalent to a onesided t
that detects outliers at the 10 percent level of significance (38). Th
K values are based on a normal distribution for detection of single out
liers. In this Bulletin, the test is applied once and all values above
the equation 7 threshold or below that from equation 8a or b are consid
outliers. The selection of this outlier detection procedure was based
testing several procedures on simulated logPearson Type III and observ
flood data and comparing results. The population skew coefficients for
the simulated data were between 4 1.5, with skews for samples selected
from these populations rangina between 3.67 and +3.25. The skew value
'for the observed data were between 2.19 and t2.80. Other test procedures
evaluated included use of station, generalized, weighted, and zero skew.
The selected procedure performed as well or better than the other pro
cedures while at the same time being simple and easy to apply. Based on
these results, this procedure is considered appropriate for use with the
logPearson Type III distribution over the range of skews, 4 3.
10. Historic Flood Data Information which indicates that any flood
peaks which occurred before, during, or after the systematic record
are maximums in an extended period of time should be used in frequency
computations. Before SUCII data are used, the reliability of the data,
the peak discharge magnitude, changes in watershed conditions over the
extended period of time, and the effects of thasc on the computed frequency
curve must all be evaluated by the analyst. The adjustment described
in Appendix 6 is recommended when historic data are used. The underlying
assumption to this adjustment is that the data from the systematic record
is representative of the intervening period between the systematic and
historic record lengths. Comparison of results from systematic and
historically adjusted analyses should be made.
The hjstoric information should be used unless the comparison
of the two analyses, the magnitude of the observed peaks, or other
factors suggest that the historic data are not indicative of the ex
tended record. All decisions made should be thoroughly documented.
C. Refinements to Frequency Curve
The accuracy of flood probability estimates based upon statistical
analysis of flood data deteriorates for probabilities more rare than
those directly defined by the period of systematic record. This is
partly because of the sampling error of the statistics from the station
data and partly because the basic underlying distribution of flood
data is not known exactly,
Although other procedures0 for estimating floods on a watershed
and flood data from adjoining watersheds can sometimes be used for evalu
ating flood levels at high flows and rare exceedance probabilities;
19
procedures for doing so cannot be standardized to the same extent as the
procedures discussed thus far. The purpose for which the flood frequency
information is needed will determine the amount of time and effort that
can justifiably be spent to obtain and make comparisons with other water
sheds, and make and use flood estimates from precipitation. The remainder
of the recommendations in this section are guides for use of these
additional data to refine the flood frequency analysis.
The analyses to include when determining the flood magnitudes with
0.01 exceedance probability vary with length of systematic record as shown
by an X in the following tabulation:
Length of Record Available
* Analyses to Include 10 to 24 25 to 50 50 or more
Statistical Analysis X X X
Comparisons with Similar Watersheds X X 
Flood Estimates from Precipitation X 4
All types of analyses should be incorporated when defining flood
magnitudes for exceedance probabilities of less than 0.01. The following
sections explain how to include the various types of flood information
in the analysis.
1. Comparisons with Similar WatershedsA comparison between flood
and storm records (see3 e.g., (12)) and flood flow frequency ana!yses at
nearby hydrologically similar watersheds will often aid in evaluating
and interpreting both unusual flood experience and the flood frequency
analysis of a given watershed, The shorter the flood record and the more
unusual a given flood event, the greater will be the need for such com
parisons,
Use of the weighted skew coefficient recommended by this guide is
one form of regional comparison. Additional comparisons may be helpful
and are described in the following paragraphs.
20
Several mathematical procedures have been proposed for adjusting
a short record to reflect experience at a nearby longterm station,
Such procedures usually yield useful results only when the gaging stations
are on the same stream or in watersheds with centers not more than 50
miles apart. The recommended procedure for making such adjustments is
given in Appendix 73 The use of such adjustments is confined to those
situations where records are short and an improvement in accuracy of
at least 10 percent can be demonstrated.
Comparisons and adjustment of a frequency curve: based upon flood
experience in nearby hydrologically similar watersheds can improve mc:st
flood frequency determinations. Comparisons of statistical parameters
of the distribution of flows with selected exceedance probabilities can
be made using prediction equations [e.g., (13), (14), (15), (16)], the
index flood method (17), or simple drainage area plots. As these estimates
are independent of the station analysis, a weighted average of the two
estimates will be more accurate than either alone. The weight given
to each estimate should be inversely proportional to its variance as
described in Appendix 8. Recommendations of specific procedures for
regional comparisons or for appraising the accuracy of such estimates
are beyond the scope of this guide. In the absence of an accuracy
appraisal, the accuracy of a regional estimate of a flood with 0.01
exceedance probability can be assumed equivalent to that from an analysis
of a loyear station record.
2. Flood Estimates from PrecipitationFloods estimated from observed
or estimated precipitation (rainfall and/or snowmelt) can be used in
several ways to improve definition of watershed flood potential. Such
estimates, however% require a procedure (e.g*) calibrated watershed
model, unit hydrograph,rainfallrunoff relationships) for converting pre
cipitation to discharge. Unless such procedures are available, considerable
effort may be required to make these flood estimates. Whether or not
such effort is warranted depends upon the procedures and data available
and on the use to be made of the estimate.
Observed watershed precipitation can sometimes be used to estimate
a missing maximum event in an incomplete flood record,
21
Observed watershed precipitation or precipitation observed at nearby
stations in a meteorologically homogeneous region can be used to generate
a synthetic record of floods for as many years as adequate precipitation
records are available. Appraisal of the technique is outside the scope
of this guide. Consequently, alternative procedures for making such
studies, or criteria for deciding when available flood records should
be extended by such procedures have not been evaluated.
Floods developed from precipitation estimates can be used to adjust
frequency curves, including extrapolation beyond experienced values.
Because of the many variables, no specific procedure is recommended
at this time. Analysts making use of such procedures should first stand
ardize methods for computing the flood to be used and then evaluate
its probability of occurrence based upon flood and storm experience
in a hydrologically and meteorologically homogeneous region. Plotting
of the flood at the exceedance probability thus determined provides
a guide for adjusting and extrapolating the frequency curve. Any adjust
ments must recognize the relative accuracy of the flood estimate and
the other flood data.
VI. Reliability Application
The preceding sections have presented recommended procedures for
determination of the flood frequency curve at a gaged location. When
applying these curves to the solution of water resource problems, there
are certain additional considerations which must be kept in mind. These
are discussed in this section.
It is useful to make a distinction in hydrology between the concepts
of risk and uncertainty (18).
Risk is a permanent population property of any random phenomenon
such as floods. If the population distribution were known for floods,
then the risk would be exactly known. The risk is stated as the probabil
ity that a specified flood magnitude will be exceeded in a specified
period of years. Risk is inherent in the phenomenon itself and cannot
be avoided.
22
Because use is made of data which are deficient, or biased, and
because population properties must be estimated from these data by
some technique, various errors and information losses are introduced
into the flood frequency determination. Differences between the population
properties and estimates of these properties derived from sample data
constitute uncertainties. Risk can be decreased or minimized by various
water resources developments and measures, while uncertainties can
be decreased only by obtaining more or better data and by using better
statistical techniques.
The following sections outline procedures to use for (a) computing
confidence limits which can be used to evaluate the uncertainties inherent
in the frequency determination, (b) calculating risk for specific time
periods, and (c) adjusting the frequency curve to obtain the expected
probability estimate. The recommendations given are guides as to how
the procedures should be applied rather than instruction on when to
apply them. Decisions on when to use each of the methods depend on
the purpose of the estimate.
A, Confidence Limits
The user of frequency curves should be aware that the curve is
only an estimate of the population curve; it is not an exact representation.
A streamflow record is only a sample. How well this sample will predict
the total flood experience (population) depends upon the sample size,
its accuracy, and whether or not the underlying distribution is known.
Confidence limits provide either a measure of the uncertainty of the
estimated exceedance probability of a selected discharge or a measure of
the uncertainty of the discharge at a selected exceedance probability.
ConFidence limits on the discharge can be computed by the procedure
described in Appendix 9.
Application of confidence !iitilits in reaching water resource planning
decision depends upon the needs of the user. This discussion is presented
to emphasize that the frequency curve developed using this guide is
only today's best estimate of the flood frequency distribution. As
more data become available, the estimate will normally be improved
and the confidence limits narrowed.
23
B. Risk
As used in this guide, risk is defined as the probability that
one or more events will exceed a given flood magnitude within a specifiec
period of years. Accepting the flow frequency curve as accurately
representing the flood exceedance probability, an estimate of risk
may be computed for any selected time period. For a lyear period
the probability of exceedance, which is the reciprocal of the recurrence
interval T, expresses the risk. Thus, there is a 1 percent chance that
the looyear flood will be exceeded in a given year. This statement
however, ignores the considerable risk that a rare event will occur
during the lifetime of a structure. The frequency curve can also be
used to estimate the probability of a flood exceedance during a specifiec
time period. For instance, there is a 50 percent chance that the flood
with annual exceedance probability of 1 percent will be exceeded one
or more times in the next 70 years.
Procedures for making these calculations are described in Appendix
10 and can be found in most standard hydrology texts or in (19) and (20)
C. Expected Probability
The expected probability is defined as the average of the true
probabilities of all magnitude estimates for any specified flood frequent
that might be made from successive samples of a specified size [(B),
(21)]. It represents a measure of the central tendency of the spread
between the confidence limits.
The study conducted for the Work Group (8) and summarized in
Appendix 14 indicates that adjustments [(21),(Z)] for the normal distri
bution are approximately correct for frequency curves computed using
the statistical procedures described in this guide. Therefore, the
committee recommends that if an expected probability adjustment is made,
published adjustments applicable to the normal distribution be used.
It would be the final step in the frequency analysis. It must be docu
mented as to whether or not the expected probability adjustment is
made. If curves are plotted, they must be appropriately labeled,
24
It should be recognized when using the expected probability adjust
ment that such adjustments are an attempt to incorporate the effects
of uncertainty in application of the curve. The basic flood frequency
curve without expected probability is the curve used in computation
of confidence limits and risk and in obtaining weighted averages of
independent estimates of flood frequency discharge.
The decision about use of the expected probability adjustment is
a policy decision beyond the scope of this guide. It is most often used
in estimates of annual flood damages and in establishing design flood
criteria.
Appendix 11 provides precedures for computing the expected proba
bility and further description of the concept.
VII. Potpourri
The following sections provide information that is of interest
but not essential to use of this guide,
A, Nonconforming Special Situations
This guide describes the set of procedures recommended for defining
flood potential as expressed by a flood flow frequency curve. In the
Introduction the point is made that special situations may require other
approaches and that in those cases where the procedures of this guide
are not followed, deviations must be supported by appropriate study,
including a comparison of the results obtained with those obtained using
the recommended procedures.
It is not anticipated that'many special situations warranting other
approaches will occur. Detailed and specific recommendations on analysis
are limited to the treatment of the station data including records of
historic events. These procedures should be followed unless there are
compelling technical reasons for departing from the guide procedures.
These deviations are to be documented and supported by appropriate study3
including comparison of results. The Hydrology Subcommittee asks that
these situations be called to its attention For consideration in
future modifications of this guide.
25
The map of skew (Plate I) is a generalized estimate. Users are
encouraged to make detailed studies for their region of interest using
the procedures outlined in Section V.B.3.
Major problems in flood frequency analysis at gaged locations are
encountered when making flood estimates for probabilities more rare than
defined by the available record. For these situations the guide described
the information to incorporate in the analysis but allows considerable
latitude in analysis.
t3. Plotting Position
Calculations specified in this guide do not require designation
of a plotting position. Section V.B.TO., describing treatment of historic
data, states that the results of the analysis should be shown graphically
to permit an evaluation of the effect on the analysis of including historic
data. The merits of alternative plotting position formulae were not
studied and no recommendation is made.
A general formula for computing plotting positions (23) is
pJ!!xL(9)
(Nab+l)
where
m = the orderedsequence of flood values with
the largest equal to 1
* IJ = number of items in data set and a and b depend
*
upon the distribution. For symmetrical
distributions a=b and the formula reduces to
p=(m_a) (10)
(IWa+l)
26
The Weibull plotting position in which a in equation 10 equals
0 was used to illustrate use of the historic adjustment of figure 63
and has been incorporated in the computer program referenced in Appendix
13, to facilitate data and analysis comparisons by the program user.
This plotting position was used because it is analytically simple and
intuitively easily understood (18, 24).
Weibull Plotting Position formula:
p=m (11)
N+f
C. Future Studies
This guide is designed to meet a current, everpressing demand
that the Federal Government develop a coherent set of procedures for
accurately defining flood potentials as needed in programs of flood
damage abatement. Much additional study and data are required before
the twin goals of accuracy and consistency will be obtained. It is
hoped that this guide contributes to this effort by defining the essential
elements of a coherent set of proedures for flood frequency determination.
Although selection of the analytical procedures to be used in each step
or element of the analysis has been carefully made based upon a review
of the literature, the considerable practical experience of Work Group
members, and special studies conducted to aid in the selection process,
the need for additional studies is recognized. Following is a list
of some additional needed studies identified by the Work Group.
1. Selection of distribution and fitting procedures
(a) Continued study of alternative distributions and
fitting procedures is believed warranted.
(b) Initially the Work Group had expected to find that
the proper distribution for a watershed would vary
depending upon watershed and hydrometeorological
conditions. Time did not permit exploration of
this idea.
27
(c) More adequate criteria are needed for selection
of a distribution.
(d) Development of techniques for evaluating
homogeneity of series is needed.
2. The identification and treatment of mixed distributions.
3. The treatment of outliers both as to identification and
computational procedures.
4. Alternative procedures for treating historic data.
5. More adequate computation procedures for confidence limits
to the Pearson III distribution.
6. Procedures to incorporate flood estimates from precipitation
into frequency analysis.
7. Guides for defining flood potentials for ungaged watersheds
and watersheds with limited gaging records.
8. Guides for defining flood potentials for watersheds altered
by urbanization and by reservoirs*
28
Appendix 1
REFERENCES
1. Brater, E. F. and J. D. Sherrill,
Fre uencies of Floods in Michi an
* %r3+ ty o Mlc igan, Ann r or,
Univers
2. Gann, E. E., "Generalized FloodFrequency Estimate for Urban Areas
in Missouri," USGS OpenFile Report, 18 pp., 1971.
3. Thomas, C. A., W. A. Harenberg, and 3. M, Anderson, "Magnitude and
Frequency of Floods in Small Drainage Basins in Idaho," USGS Water
Resources Inv. 773, NTIS, Springfield, VA., 1973.
4. Todorovic, R. and E. Zelenhasic, "A Stochastic Model for Flood
Analysis," Water Resources Research, Vol. 6, No. 6, 1970, pp. 1641
1648.
5, Carrigan, P. H., Jr., and C. S. Hutzen, "Serial Correlation of
Annual Floods," International Hydrology Symposium, Fort Collins,
September 1967.
6. Matalas, N. C., "Autocorrelation of Rainfall and Streamflow Minimums,"
U.S. Geological Survey Prof. Paper 434B, 1963.
7. Pacific Southwest Interagency Commission Report of the Hydrology
Subcommittee, "Limitations in Hydrologic Data as Applied to Studies
in Water Control Management," February 1966.
8. Beard, L. R,, Flood Flow Frequency Techni ues Center for Research
in Water Resources, The Unlverslty +o Texas at Austin, 1974.
9. Langbein, W. B., "Annual Floods and the Partial Duration Series"
Transactions, American Geophysical Union, Vol. 30, 1949, p. 879.
10. Hardison, C. H., "Generalized Skew Coefficients of Annual Floods in
the United States and Their Application" Water Resources Research,
Vol. 10, No. 5, pp. 745752.
11. U.S. Army Engineer District, Sacramento, California, Civil Works
Investigations Project CW151, Flood Volume StudiesWest Coast.
Research Note No. 1, "Frequency of New England Floods," July 1958,
ll
12, Q&of; Engineers, U.S. Army, "Storm Rainfall in the United
n Washington, 1945.
13. Benson, M. A., "Evaluation of Methods for Evaluating the Occurrence
of Floods," USGS Water Supply Paper 1580A, 1962,
14. Benson, M. A., "Factors Influencing the Occurrence of Floods in a
llh$d Region of Diverse Terrain, " USGS Water Supply Paper 15808,
.
15, Benson, M. A., "Factors Affecting the Occurrence of Floods in the
Southwest," USGS Water Supply Paper 1580D, 1964.
16, Bock, P,, I. Enger, G, P, Malhotra, and D. A, Chisholm, "Estimating
Peak Runoff Rates from Ungaged Small Rural Watersheds," National
Cooperative Highway Research Program Report 136, Highway Research
Board, 1972.
17. Dalrymple, T., Flood Frequent Anal ses Manual of H drolo : Part
pt "Flood Flow mques," f+ ater
U G SumPaper b8$i96~
18. Yevjevich, Vujica, Probabilit and Stat;s;;;so+H drolo
Resources Publications,'l?X+o~ns, C 1 d , e Water
19. Gumbel, E. J., "The Calculated Risk in Flood Control," Applied
Science Research, Section A, Vol. 5, 1955, The Hague.
20. Riggs, H. C., "Frequency of Natural Eventseli Journal of the Hydraulics
* Division, & &, Vol. 87, No. HYl, 1961,mx.34
21. Beard, L. R., "Probability Estimates Based on Small NormalDistribu
tion Samples," Journal of Geophysical Research, July 1960.

22. Hardison, C., and M. Jennings, "Bias in Computed Flood Risk,"
* Journal of the Hydraulics Division, PFoc. ASCE, Vol. 98, No. HY3, Marc!
l~i% Ecussion ana*, =415427.
23, Harter, H. L., "Some Optimization Problems in Parameter Estimation,"
edited by Jagdish S. Rustagi, Optimizing Methods $IStatistics,
Academic Press, New York,
24, Chow, V. T,, Handbook of
New York, 196mpp,8zB
25. Hardison, C., "Accuracy
Professional Paper 650D,
1971, pp. 3262.
A lied Hydrology, McGrawHill
an!hz9*
of Streamflow Characteristics,"
1969, pp* D210D214.
Book Co.,
USGS
12
26. Wilson, E. B. and M. M. Hilferty, "The Distribution of ChiSquare,"
Proc., National Academy of Science, Vol. 17, No. 12, December 1931,
pp. 684688.
27. McGinnis, David G., and William H. Sammons, Discussion of Paper
by Payne, Neuman, and Kerri. "Daily Stream Flow Simulation,"
Journal of the Hydraulics Division, Proc. ASCE, Vol. 96, No. HY5,
May 1970.
28. Jennings, M. E. , and M. A. Benson, "Frequency Curves for Annual
Flood Series with Some Zero Events or Incomplete Data," Water
Resources Research, Vol. 5, No. 1, 1969, pp. 276280.
29. Matalas, N., and B. Jacobs, "A Correlation Procedure for Augmenting
Hydrologic Data: USGS Professional Paper 434E, 1964.
30. Gilroy, E. 3. personal communication to C. Hardison, 1974.
31. Beard, L. R., Statistical Methods in Hydrology U. S. Army Corps
of Engineers, Civil Works Investigation Projeci CW151, 1962.
32. Natrella, M. G., Experimental Statistics, National Bureau of Standards
Handbook 91, 1963.
33. Hardison, C. H., personal communication, 1974
34. U.S. Corps of Engineers, "Regional Frequency Computation," General
ized Computer Program, The Hydrologic Engineering Center, July
1972.
35. Water Resources Council, Hydrology Committee, A Uniform Technique
for Determining Flood Flow Frequencies, BulletTn 15, Washington,
D.C., 1967. 
36. Harter, H. L., and A. H. Moore, "A Note on Estimation from a Type
I ExtremeValue Distribution," Technometrics, Vol. 9, NO. 2, May
1967, pp. 325331.
37. Thorn, H. C. S., "A Note on the Gamma Distribution," Monthly Weather
Review, Vol. 86, 1958, pp. 117122.
* 38. Grubbs, Frank E, and Glenn Beck, "Extension of Sample Sizes and
Percentage Points for Significance Tests of Outlying Observations,"
Technometrics, Vol. 14, No. 4, November 1972, pp. 847854.
39. Tasker, Gary D., "Flood Frequency Analysis with a Generalized Skew
Coefficient," Water Resources Research, Vol. 14, No. 2, April1978,
pp. 373376.
40. Wallis, J. R., N. C. Matalas, and J. R. Slack, “Just a Moment,"
Water Resources Research, Vol. 10, No. 2, April 1974, pp. 211219.
l3
* 41. Resnikoff, G. J., and G. J. Lieberman, Tables of the NonCentral
tDistribution, Stanford University Press, Stanford, California,
1957.
42. Zelen, k, and N. C. Severo, "Probability Functions," Handbook
of Mathematical Functions, Applied Mathematics Series No. 55,
U. S. National Bureau of Standards, 1904.
43, Owen, D. B., "Factors for OneSided Tolerence Limits and for
Variables Sampli,ng Plans," Sandia Corporation Monograph SCR607,
March, 1963.
=x
l4
Appenaix z
GLOSSARY AND NOTATION
Glossary
The terms used in this guide include definitions taken from refer
ences listed in the Bibliography or from "Nomenclature for Hydraulics,"
Manual 43, American Society of Civi'l Engineers, 1962, and from definitions
especially prepared for this guide. For more technical definitions of
statistical terms* see "Dictionary of Statistical Terms" by M. G. Kendall
and W. R. Buckland, Hafner Publishing Company, New York, 1957.
TERM Definition
Annual Flood The maximum momentary peak discharge in each year
of record. (Sometimes the maximum mean daily
discharge is used,)
Annual FZood A list of annual floods.
b!MVkS
Annual Series A general term for a set of any kind of data in
which each item is the maximum or minimum in a year.
Array A list of data in order of magnitude; in flood
frequency analysis it is customary to list the
largest value first, in a lowflow frequency analysis
the smallest fdrst.
Broken Record A systematic record which Is divided into separ
ate continuous segments because of deliberate
discontinuation of recording for significant periods
of time.
2l
coef;faczent of n numerical measure or inaex 0T tne lack OT sym
Skezl?ne88 metry In a frequency distribution. Function of
the thtrd moment of magnitudes about their mean, a
measure of asymmetry. Also called "coefficient of
skew" or "skew coefficient."
Confidence
Limits
Computed values
a parameter that
bility the range
parameter lies.
on both sides of an estimate
show for a specified proba
in which the true value of
of
the
Distribution Function describjng the
which events of various
relative
magnitudes
frequency
occur.
with
Distribution
Free
Requiring no assumptions about the kind of
bility distribution a set of data may have,
proba
Exceedance
Frequency
The percentage of values that exceed a specified
magnitude, 100 times exceedance probability.
Exceedance
ProbabiZity
Probability
specified
one year
that a random event will
magnitude in a given time
unless otherwise indicated.
exceed
period,
a
usually
Expected
Probability
The average of the true probabilities of all
magnitude estimates for any specified flood fre
quency that might be made from successive samples
a specified size.
of
Generalized
Coefficient
Sketi A skew coefficient
integrates values
derived
obtained
by
at
a procedure which
many locations,
Homogeneity Records from the same populations.
22
Incomp Ze te
Record
LeveZ of
Significance
MeanSquare
Ex~or
Method of
Mome&s
Nonparame tric
NopmaZ
lxs haLtion
A streamflow record in which some peak flows
are missing because they were too low or
high to record or the gage was out of operation
for a short period because of flooding.
The probability of rejecting a hypothesis
when it is in fact true. At a "10percent"
level of significance the probability is
l/10.
Sum of the squared differences between the
true and estimated values of a quantity divided
by the number of observations. It can also be
defined as the bias squared plus the variance
of the quantity. *
A standard statistical computation for estim
ating the moment of a distribution from the
data of a sample.
The same as distributionfree.
A probability distribution that is symmetrical
about the mean, median, and mode (bellshaped).
It is the most studied distribution in sta
tistics, even though most data are not exactly
normally distributed, because of its value
in theoretical work and because many other
distributions can be transformed into normal.
It is also known as Gaussian, The Laplacean,
The GaussLaplace, or the LaplaceGauss dis
tribution, or the Second Law of Laplace.
23
OUt;ZiC?X Outliers (extreme events) are data points
which depart from the trend of the rest of data.
Parameter A characteristic
or standard
descriptor,
deviation.
such as a mean
Percent Chance A probability multiplied by 100.
PopuZation The entire
from which
The total
floods at
lation of
the floods
(usually infinite) number of data
a sample is taken or collected.
number of past, present, and future
a location on a river is the popu
floods for that location even if
are not measured or recorded.
Recuxxence
Interval (Return
Period, Excsed
ante Interva 2 I
The average time interval between actual
occurrences of a hydrological event of a
given or greater magnitude. In an annual
flood series, the average interval in which
a flood of a given size is exceeded as an
annual maximum. In a partial duration series,
the average interval between floods of a given
size, regardless of their relationship to
the year or any other period of time. The
distinction holds even though for large floods
recurrence intervals are nearly the same for
both series.
Sarrp i!e An element, part,
Every hydrologic
longer record,
or fragment
record is a
of a
sample
"popUlat.iOn."
of a much
Skew Coefficient See "coefficient of skewness."
2J
Standard A measure of the dispersion or precision
Deviation of a series of statistical values such
as precipitation or streamflow. It is
the square root of the sum of squares
of the deviations from the arithmetic
mean divided by the number of values
or events in the series. It is now
standard practice in statistics to divide
by the number of values minus one in
order to get an unbiased estimate of
the variance from the sample data.
Standard Error An estimate of the standard deviation
of a statistic, Often calculated from
a single set of observations. Calculated
like the standard deviation but differing
from it in meaning.
Student's t A distribution used in evaluation of
Distribut (0.01908) (.0.18746) 6
(76) (7573 (0.02413) [ 1:
(3)(1.68182)(43)(0.02487)(0~07140) + (1.68182)(44)(0.0000154) + (0.70802)]
"G = 0.0418
*
66
EXPLANATION
Histortcal peaks (1897,1919 and 1927) plotted as
largest m 77 year perrod (18971973).
0 Recorded peaks (44 years, 19301973)plotted
such that each point represents 1.68 years
in the longer 77 year period.
Points plotted by Weibull Plotting Position formula.
Weighted Log Pearson Type III compged curve.
N=44, Z=3, H=77 G = 0.04
5=3+71581 Gw=0.004
;=0.289
I
1,000
#
k I
99.99 99.9 99.8 99.5
EXCEEDANCEPROBABILITY, IN PERCENT
BIG SANDY RIVER AT BRUCETON, TN
ANNUAL PEAKS
HISTORICALLY WEIGHTED LOG
PEARSPN TYPE Ill
FIGURE 63
._ 
FIB
5
2 1 0.5 0.2
Appendix 7
TWO STATION COMPARISON
INTRODUCTION
The procedure outlined herein is recommended for use in adjusting
the logarithmic mean and standard deviation of a short record on
the basis of a regression analysis with a nearby longterm record.
The theoretical basis for the equations provided herein were developed
by Matalas and Jacobs (29).
The first step of the procedure is to correlate observed peak
flows for the short record with concurrent observed peak flows for
the long record. The regression and correlation coefficients, respectively,
can be computed by the following two equations:
xxlyl I=xlDl/Nl
b = (7l)
XX; @Xl 12/Nl
(72)
where the terms are defined at the end of this Appendix.
If the correlation coefficient defined by equation 72 meets
certain criteria, then improved estimates of the short record mean and
standard deviation can be made. Both of these statistics can be improved
when the variance of that statistic is reduced. As each statistic is
evaluated separately, only one adjustment may be worthwhile. The criterion
and adjustment procedure for each statistic are discussed separately.
In each discussion, two cases are considered: (1) entire short record
contained in the long record, (2) only part of the short record contained
in the long record. The steps for case 2 include all of those for
case 1 plus an additional one.
CRITERION AND ADJUSTMENT PROCEDURE FOR MEAN
*
The variance of the adjusted mean (v) can be determined by equation 73:
q
Var(V) =
(1r')
(73)
l&T
I )J
Since (Syl)2/Nl is the variance of yl, the shortrecord mean, V will be
1r’
a better estimate of the true mean than yl if the term r2 N in
1
equation 73 is positive. Solving this relationship for r yields
equation 74. If the correlation coefficient satisfies equation 74,
r > l/(N1 2)l" (74)
then an adjustment to the mean is worthwhile. The right side of this
inequality represents the minimum critical value of r. Table 7l contains
minimum critical values of r for various values of Nl. The adjusted
logarithmic mean can be computed using equation 75a or 75b.
J=YpN2 (75a)
N1+N2
y = Yl + b(=K3 xl) (75b)
Equation 75b saves recomputing a newF2 at the long record station
for each short record station that is being correlated with the long
record station. While the adjusted mean from equation 75a or 75b
may be an improved estimate of the mean obtained from the concurrent
period, it may not be an improvement over the entire short record mean
in case 2. It is necessary to compare the variance of the adjusted
mean (equation 73) to the variance of the mean cv,) for the entire short
record period (N3). Compute the varfance of the mean Y3 using equation
76:
s 2
Var(Y3) = (
Y3 ) (7M *
N3
72
*
where S is the standard deviation of the logarithms of flows for the
y3
short record site for the period N3. If the variance of equation 76 is
smaller than the variance of v given in equation 73, then use y3 as the
final estimate of the mean. Otherwise, use the value of y computed in
equation 75a or 75b.
EQUIVALENT YEARS OF RECORD FOR THE MEAN
As illustrated in equations 73 and 76, the variance of the mean
is inversely proportional to the record length at the site. Using
equation 73 it can be shown that the equivalent years of record, N,,
for the adjusted mean is:
N, =
(77)
It may be seen from equation 77 that when there is no correlation
(r=O), then N, is less than N,. This indicates that the correlation
technique can actually decrease the equivalent years of record unless
r satisfies equation 74. For perfect correlation (r=l), then
= N, + N2, the total record length at the long record site.
Ne
Although N, is actually the equivalent years of record for the
mean, it is recommended that N, be used as an estimate of the equivalent
years of record for the various exceedance probability floods dn the
computation of confidence limits and in applying the expected probability
adjustment.
CRITERION AND ADJUSTMENT PROCEDURE FOR THE STANDARD DEVIATION
The variance of the adjusted variance Sy2 (square of the standard de
viation) can be determined by equation 78:
73
* E(S )$
Var(Sy2)= N l j.
“1 C1 (78)
1(N,+N2,)2
where A, B, and C are defined below and the other terms are defined
at the end of the appendix. In equation 78, 2(S )4/(N,,) is the
Yl
variance of S ' (the shortrecord variance). If the second term
in equation 78
Yl
is negative, then the variance of S 2 will be less
Y
2
than the variance of S Solving this relationship for r yields the
Yl'
following equation:
Irl >[B z dB2:4AC ]112 (79)
2A
,, (N2+2)(N,6)(N,8) 8(N,4) 2N2(N, 4) 2 N N (N 4)'
+12 1
where =
(Nl3)(N,5) (N13) (N,3)2
(N,3)2(N,2]
4(N,4)
+ (Tip
B 6(N2+2)(N,6) 2(N;N,14) + 2N2(N, 4) (N, 5)
= (N,3)(N,6) * (N,3) (Nl3)2
2(N,4)(N,+3) 2N,N,(N,4)2
(N13) (N13)2(N,2)
2(N,+l) WJ2+21 ++I) (2N,+N22)
C = NJ3 + (NJ3)(Nl5) N11
2N2N,41 + 2(N,4)(N,+l) + N,~~2(N,4)
+ (N,3)2 (Nl33
(N, 3J2(N, 2)
74
*
The right side of the inequality (79) represents the minimum critical
value of r. Table 7l gives approximate minimum critical values of
r for various values of Nla The table values are an approximation as
they are solutions of equation 79 for a constant N2. The variations
in N2 only affect the table values slightly.
If the correlation coefficient satisfies equation 79, then the
adjusted variance can be computed by equation 710:
S2y = +j [ (Nl1 )Sy12 + (N2l)b2~x22
+ N2(Nl4)(Nl1)
(lr2)S ' + !!& b2 (x _ ,,)2](740)
(Nl3) (N,2) N1+N2 2
Yl
The adjusted standard deviation Sy equals the square root of the adjusted
variance in equation 710. The third term in brackets in equation
710 is an adjustment factor to give an unbiased estimate of Sy2.
This adjustment is equivalent to adding random noise to each estimated
value of flow at the shortterm site.
While the adjusted variance from equation 710 may be an improved
estimate of the variance (standard deviation) obtained from the con
current period, Jt may not be an improvement over the entire short
record variance (standard deviation) in case 2. It is necessary to
compare the variance of the adjusted variance (equation 78) to the
variance of the variance (S 2, for the entire period (N3)" ': Compute
Y3
the variance of the shortrecord variance (S 2, using equation 711.
y3
(711)
N3 1
75
*
where all terms are previously defined. If the variance of equation
711 is smaller than the variance of Sy2 given in equation 78, then
use S as the final estimate of the standard deviation. Otherwise,
y3
use the value of Sy determined from equation 710.
FURTHER CONSIDERATIONS
The above equations were developed under the assumption that the
concurrent observations of flows at the short and longterm sites have
a joint normal probability distribution with a skewness of zero. When
this assumption is seriously violated, the above equations are not exact
and this technique should be used with caution, In addition, the reli
ability of r depends on the length of the concurrent period, NT. To
obtain a reliable estimate of r, N, should be at least 10 years.
Notice that it is not necessary to estimate the actual annual peaks
from the regression equation but only the adjusted logarithmic mean and
standard deviation. The adjusted skew coefficient should be computed
by weighting the generalized skew with the skew computed from the short
record site as described in Section V.B.4.
7s
*
NOTATION
Nl = Number OF years when flows were concurrently observed at the two sites
N2 = Number of years when flows were observed at the longer record site
but not observed at the short record site
N3 = Number of years'of flow at the short record site
Ne= Equivalent years of record of the adjust ed mean
S = Standard deviation of the logarithm of f lows for the extended periodY at the short record site
S
x1
= Standard
during
deviation
concurrent pof logarithm
eriod
of flows at the long record site
S"
"2
= Standard
for the
deviation
period when
of logarithm
flows were
of
not
flows
observed
at the
at the
long record site
short record site
S
y1
= Standard
for the
deviation
concurrent
of the
period
logarithm of flows at the short record site
s
Y2
= not used
s
'3
= Standard deviation
short record site
of logarithm of flows for the entire period at the
= Logarithms of flows from long record during concurrent period
xl
= Mean logarithm of flows at the long record site for the concurrent
5
period
x2 = Mean logarithm of flows at the long record site for the period when
flow records are not available at the short record site
x3 = Mean logarithm of flows for the entire period at the long record site
yl
= Logarithms of flows from short record during concurrent period
v = Mean logarithm of flows for the extended period at the short record
site
5 = Mean logarithm of flows for the period of observed flow at the short
record site (concurrent period)
y2 = not used
77
*
5 = Mean logarithm of flows for the entire period at the short record site
b = RegreSsion coefficient for Y, on X,
r = Correlation coefficient of the flows at the two sites for concurrent
periods
*
TABLE 7l MINIMUM r VALUES FOR IMPROVING
MEAN OR STANDARD DEVIATIONESTIMATES
CONCURRENT MEAN STANDARD
RECORD DEVIATION
10 0.35 0.65
11 0.33 0.62
12 0.32 0*59
13 0.30 0.57
14 0.29 0.55
15 0.28 0.54
16 0.27 0,52
17 0.26 0.50
18 0.25 0.49
19 0,24 0,48
20 0.24 0.47
21 0.23 0.46
22 0.22 0.45
23 0.22 0.44
24 0.21 0.43
25 0.21 0.42
26 0.20 0.41
27 cJ.20 0.41
28 0.20 0.40
29 0.19 0.39
30 0.19 0.39
31 0.19 0.38
32 0.18 0.37
33 0.18 0.37
34 OJ8 0.36
35 0.17 0.36 +
79
Appendix 8
WEIGHTINGOF INDEPENDENT ESTIMATES
The following procedure is suggested for adjusting flow frequency
estimates based upon short records to reflect flood experience in
nearby hydrologically similar watersheds, using any one of the various
*generalization methods mentIoned in V.C.1, The procedure is based upon
*
the assumption that the estimates are independent, which for practkal
purposes is true in most situations.
If two independent estimates are weighted inversely proportional to
their variance, the variance of the weighted average, z, is less than
the variance of either estimate. According to Gilroy (30), if
x(VJ + Y(Vx)
=
z (8J 1
vy + vx
then
.
vv
XY
v, = Vx + Vy + 2rdm (82)
ox + Vy12
:L I
in which Vx, Vy, and Vz are the var%ances of x, y, and z respectively,
and r is the cross correlation coefficient between values of x and
values of y. Thus, if two estimates are independent, r is zero and
vz = vxvy (83)
vx + v
y <:’
As the variance of flood events at selected exceedance probabilities
computed by the Pearson Type III procedure is inversely proportional to
the number of annual events used to compute the statistics (25), equation
(83) can be written
(WJx) (C/NY)
C/N, = (84)
C/N, + C/NY
in which C is a constant, N, and NY are the number of annual events used
to compute x and y respectively, and N, 'fs the number of events that
would be required to give a flood event at the selected exceedance
probabilities with a variance equivalent to that of z computed by
equation 8l. Therefore,
8l
Nz=N, f NY (85)
From equation 81,
(86) +
Equation 86 can be used to weight independent estimates of the logarithms
of flood discharges at selected probabilities and equation 85 can be used
to appraise the accuracy of the weighted average. As a flood frequency
discharge estimated by generalization tends to be independent of that obtained
from the station data, such weighting is often justified particularly if the
stations used in the generalization cover an area with a radius of over 100
miles or if their period of record is long in comparison with that at the
station for which the estimate is being made, For generalizations based on
stations covering a smaller area or with shorter records, the accuracy of
the weighted average given by equation 86 is less than given by equation 85.
For cases where the estimates from the generalization and from the
station data are not independent, the accuracy of the weighted estimate is
reduced depending on the cross correlation of the estimates.
Given a peak discharge of 1,000 cfs with exceedance probability of 0.02
from a generalization with an accuracy equivalent to an estimate based on a
loyear record, for example, and an independent estimate of 2,000 cfs from
+lS annual peaks observed at the site, the weighted average would be given
by substitution in equation 86 as follows: +
lO(log 1000) f 15(log 2000) = 3 181
Log Q,,, = 25 .
from which Q.,, is 1,520 cfs. By equation 85 this estimate is as good
as would be obtained from 25 annual peaks.
If an expected probability adjustment is to be applied to a weighted
estimate, the adjustment to probability should be the same as that appli
cable to samples from normal distributions as described in Appendix 11, but
N should be that for a sample size that gives equivalent accuracy. Thus,
in the preceding example, the expected probability adjustment would be that
for a sample of size 25 taken from a normal distribution,
82
*
Appendix 9
CQNFIUENCE LIMITS.
The record of annual peak flows at a site is a random sample of the
underlying population of annual peaks and can be used to estimate the
frequency curve of that population. If the same size random sample could
be selected from a different period of time, a different estimate of the
underlying population frequency curve probably would result. Thus, an
estimated flood frequency curve can be only an approximation to the true
frequency curve of the underlying population of annual flood peaks. To
gauge the accuracy of this approximation , one may construct an interval
or range of hypothetical frequency curves that, with a high degree of
confidence, contains the population frequency curve. Such intervals are
called confidence intervals and their end points are called confidence
limits.
This appendix explains how to construct confidence intervals for
flood discharges that have specified exceedance probabilities. To this
end, let Xt denote the true or population logarithmic discharge that has
exceedance probability P. Upper and lower confidence limits for X; ) with
confidence level c, are defined to be numbers Up ,(X) and Lp ,(X), based
on the observed flood records, X, such that the ;pper confidgnce limit
up,c(x) 1ies above X6 with probability c and the lower limit Lp ,(X) lies
below X6 with probability c. That is, the confidence limits ha;e the
property that
Probability (up,c(xl L q.!i~ = c (gla)
Probability bp ,(x1 5. xfq = c (glb)
,
9I
*
Explicit formulas for computing the confidence limits are given below;
the above formulas simply explain the statistical meaning of the confidence
limits.
The confidence limits defined above are called onesided confidence
limits because each of them describes a bound or limit on just one side
of the population pprobability discharge, A twosided confidence interval
can be formed from the overlap or union of the two onesided Intervals,
as follows:
Probability { Lp ,(X) f x*p 2 up,c(x) 1 = zc1 (92)
,
Thus, the union of two onesided g&percent confidence intervals
is a twosided gopercent interval. It should be noted that the two
sided interval so formed may not be the narrowest possible interval
with that confidence level; nevertheless, it is considered satisfactory for
use with these guidelines.
It may be noted in the above equations that Up ,1X) can lie above
*
xp if and only if Up c (XI lies above a fraction (1IP) of all possible
floods in the population. In quality control terminology, UP ,(X1 would
be called an upper tolerance limit, at confidence level c,,;fo; the
proportion (1P) of the population. Similarly, Lp ,(X1 would be a lower
tolerance limit for the proportion (Pb, Because tie tolerance limit
terminology refers to proportions of the population, whereas the confidence
limit terminology refers directly to the discharge of interest, the
confidencelimit terminology is adopted in these guidelines.
Explicit formulas for the confidence limits are derived by specifying
the general form of the limits and making additional simplifying assump
tions to analyze the relationships between sample statistics and population
statistics. The general form of the confidence limits is specified as:
up,c(x) = x f s K”Pc (93a)
( 9 >
Lp,clx) = x+ s K"P c (93b)
9
( 1
92
*
in which x and S are the logarithmic mean and standard deviation of
the final estimated log Pearson Type III frequency curve and KF c and KL
9 P9C
are upper and lower confidence coefficients.
The confidence coefficients approximate the noncentral tdistribution.
The noncentral tvariate can be obtained in tables (41, 321, although
the process is cumbersome when G, is nonzero. More convenient Is the use
of the following approximate formulas (32, pp2156, based on a large sample
approximation to the noncentral tdistribution (42):
Kit KG,,P+l( K:wgP’=ab
(94a)
9 a
KGwgP 4"'Gw.P
KL (94b)
P,c = a
in which
2
a=lzC (95)
TJm
(g6)
and zc is the standard normal deviate (zeroskew Pearson Type III deviate)
with cumulative probability c (exceedance probabflity lc). The systematic
record length N is deemed to control the statistical reliability of
the estimated frequency curve and is to be used for calculating confidence
limits even when historic information has been used to estimate the
frequency curve.
The use of equations 93 through 96 is illustrated by calculating
95percent confidence limits for X; Olg the 0.01 exceedance probability
flood, when the estimated frequency'curve has logarithmic mean, standard
deviation, and skewness of 3*00, 0.25, and 0.20, respectively based on 50
years of systematic record.
*
93
*
z
= 1.645 = 2.4723
C
KG\4$p
= (1.645)*
a = 0.9724
'98
b = (2.4723) 2 _ (l$
= 6.058
e
K; o1 o g5 = 2.4723 I d(;.V;;" (0.9724)(6.058)
. , l 0
P 3.026
Kb o1 o g5 = 2.4723 4(2.4723+ (0 9724)(6 058)
. , l
0.9724 ’ l
= 2.059
Uo.ol, o.95 (X) = 3.00 + (0.25)(3.026) = 3.756
Loo,, 0.95 (Xl = 3.00 + (0.25)(2.059) = 3.515
The corresponding limits in natural units (cubic feet per second)
are 3270 and 5700; the estimated 0.01 exceedance probability flood is
4150 cubic feet per second.
Table 9l is a portion of the noncentral t tables (43) for
a skew of zero and can be used to compute KUp c and KL ,, c for selected
values of P and c when the distribution of loiarithms oi the annual
peaks is normal (i.e., Gw=O).
An example of using table 91 to compute confidence limits is as
follows: Assume the 95percent confidence limits aru desired for X*o olg
0
the 0.01 exceedance probability flood for a frequency curve with logarithmic
mean, standard deviation and ske\rrness of 3.00, 0.25 and O.OU, respectively,
based on 50 years of systematic record,
94
*Ku 0.01, 0.95 = 2.862 Found by entering table 9l with confidence
level 0.05, systematic record length 50 and
exceedance probability 0.01.
KL0.01, 0.95 = 1.936 Found by entering table 9l with confidence
level 0.95,systematic record length 50 and
exceedance probability 0.01.
"o,ol, o.g5 (X) = 3.00 +0.25(2.862) = 3.715
Loeol, o.g5 (x> = 3.00 + 0.25tl.936) = 3.484
The corresponding limits in natural units (cubic feet per second)
are 3050 and 5190; the estimated 0.01 exceedance probability flood is
3820 cubic feet per second.
95
Appendix 9 Notation
up ,c (X> = upper confidence limit in log units
LP,ctx) = lower confidence limit in log units
P = exceedance probability
C = confidence level
*
xP
= population logarithmic discharge for exceedance probability P
x = mean logarithm of peak flows
S = standard deviation of logarithms of annual peak discharges
KGw,P
= Pearson Type III coordinate expressed in number of standard
deviations from the mean for weighted skew (Gw) and exceedance
probability (P).
GW
= weighted skew coefficient
KUP,c
= upper confidence coefficient
KLP,c
= lower confidence coefficient
N = systematic record length
i!
C = is the standard normal deviate *
96
* TABLE 9l
CONFIDENCE LIMIT DEVIATE VALUES FOR NORMAL DISTRIBUTION
EXCEEDANCE PROBABILITY
ConfiSystematic
dence Record
Level Length
N .002 ,005 .OlO .020 .040 ,100 .200 .500 .800 .900 .950 .990
.Ol 10 6.178 5.572 5.074 4.535 3.942 3.048 2.243 .892 .107 .508 .804 1.314
15 5.147 4.639 4.222 3.770 3.274 2.521 1.841 .678 .236 .629 . 929 1.458
20 4.675 4.212 3.832 3.419 2.965 2.276 1.651 .568 .313 .705 1.008 1.550
25 4.398 3.960 3.601 3.211 2.782 2.129 1.536 .498 .364 .757 1.064 1.616
30 4.212 3.792 3.447 3.071 2.658 2.030 1.457 .450 .403 .797 1.107 1.667
40 3.975 3.577 3.249 2.893 2.500 1.902 1.355 .384 .457 .854 1.169 1.741
50 3.826 3.442 3.125 2.781 2.401 1.821 1.290 .340 .496 .894 1.212 1.793
60 3.723 3.347 3.038 2.702 2.331 1.764 1.244 .309 .524 . 924 1.245 1.833
70 3,647 3.278 2.974 2.644 2.280 1.722 1.210 285 .545 .948 1.272 1.865
80 3.587 3.223 2.924 2.599 2.239 1.688 1.183 0265 .563 . 968 10 293 1.891
90 3.538 3.179 2,883 2.561 2.206 1.661 1.160 .250 .578 .984 1.311 1.913
100 3.498 3.143 2.850 2.531 2.179 1.639 1.142 .236 .591 .998 1.326 1.932
.05 10 4.862 4.379 3,981 3.549 3.075 2.355 1.702 .580 .317 .712 1.017 1,563
15 4,304 3.874 3,520 3,136 2.713 2.068 1.482 ,455 .406 ,802 1.114 1.677
20 4.033 3.628 3.295 2.934 2.534 1.926 1.370 .387 a460 ,858 1.175 1.749
25 3.868 3.478 3.158 2.809 2.425 1.838 1.301 .342 . 497 .898 1.217 1.801
30 3,755 3.376 3.064 2.724 2.350 1.777 1.252 .310 .525 .928 1,250 1.840
40 3.608 3.242 2.941 2.613 2.251 1.697 1.188 .266 .565 .97O 1.297 1.896
50 3.515 3.157 2.862 2.542 2.188 1.646 1,146 .237 .592 1.000 1.329 1.936
60 3.448 3.096 2.807 2.492 2.143 1.609 1.116 .216 .612 1.022 1.354 1.966
70 3.399 3.051 2,765 2.454 2.110 1.581 1.093 .199 .629 1.040 1.374 1.990
80 3.360 3;016 2.733 2.425 2.083 1.559 1.076 .I86 ,642 1.054 1.390 2.010
90 3.328 2.987 2.706 2.400 2.062 1.542 1.061 .175 .652 1,066 1.403 2.026
100 3.301 2.963 2.684 2.380 2.044 1.527 1.049 .166 .662 1.077 1.414 2.040
* TABLE 9l (CONTINUED)
CONFIDENCE LIMIT DEVIATE VALUES FOR NORMAL DISTRIRIJTION
EXCEEDANCE PQORARILITY
ConfiSystematic
dence Record
Level Length
N .002 .005 .OlO .02O ,040 .lOO .200 .500 .800 .900 .950 ,990
.lO 10 4.324 3.889 3.532 3.144 2.716 2.066 1.474 .437 .429 .828 1.144 1.715
15 3.936 3.539 3.212 2.857 2.465 1.867 1.320 .347 .499 .901 1.222 1.808
20 3.743 3.364 3.052 2.712 2.338 1.765 1.240 .297 .541 .946 1.271 1.867
25 3.623 3.255 2.952 2.623 2.258 1,702 1.190 .264 .570 .978 1.306 1.908
30 3.541. 3.181 2.884 2.561 2.204 1.657 1.154 .239 .593 1,002 1.332 1.940
40 3.433 3.082 2.793 2,479 2,131 1.598 1.106 .206 .624 1.036 1.369 1.986
50 3.363 3.019 2.735 2.426 2.084 1.559 1.075 al84 .645 1.059 1.396 2,018
60 3.313 2,974 2.694 2.389 2.051 1.532 1.052 .l67 .662 1.077 1.415 2.042
70 3.276 2.940 2.662 2.360 2.025 1.511 1.035 .I55 ,674 1,091 1.431 2,061
80 3.247 2.913 2,638 2,338 2.006 1.495 1,021 144 ,684 1,103 1.444 2.077
90 3.223 2.891 2.618 2.319 1.989 I.481 1.010 ,136 ,693 1.112 1.454 2.090
100 3.203 2.873 2.601 2.305 1.976 I.470 1.001 .l29 .701 1.120 1.463 2.101
.25 10 3.599 3.231 2.927 2.596 2.231 1.671 1.155 .222 .625 1.043 1,382 2.008
15 3.415 3.064 2.775 2.460 2.112 1.577 1.083 .179 .661 1.081 1.422 2.055
20 3.320 2.978 2.697 2.390 2.050 1.528 1.045 .154 .683 1.104 1.448 2.085
25 3.261 2,925 2.648 2.346 2.011 1.497 1.020 137 .699 1.121 1.466 2.106
30 3.220 2.888 2.614 2.315 1.984 1.475 1.002 .125 .710 1.133 1.479 2.123
40 3.165 2.838 2.568 2.274 1.948 1.445 .978 ,108 .726 1.151 1.499 2.147
50 3.129 2.805 2.538 2.247 1.924 1.425 .962 .096 .738 1.164 1.513 2.163
60 3.105 2.783 2.517 2,227 1.907 1.411 950 .088 .747 1.173 1.523 2.176
70 3.085 2.765 2.501 2.213 1.893 1.401 .942 .081 .753 1.181 1.532 2.186
80 3.070 2.752 2.489 2.202 1.883 1.392 .935 .076 .759 1.187 1.538 2,194
90 3.058 2.740 2.478 2.192 1.875 1.386 .929 .071 .763 1.192 1.544 2.201
100 3.048 2.731 2.470 2.184 1.868 1,380 .925 .068 .767 1.196 1.549 2.207
*
CONFIDENCE LIMIT
mm 91 (CONTINUED)
DEVIATE VALUES FOR NORMAL DISTRIBUTION
*
Confi
dence
Level
Systematic
Record
Length
EXCEEDANCE F'RORABILITY
N .002 .005 .OlO .020 .040 .lOO .200 .500 .800 .900 ,950 .990
.75 10
15
20
25
30
40
50
60
70
80
90
100
2.508
2.562
2.597
2.621
2.641
2.668
2.688
2.702
2.714
2.724
2.731
2.739
2.235
2.284
2.317
2.339
2.357
2.383
2.400
2.414
2.425
2.434
2.441
2.447
2.008
2.055
2.085
2.106
2.123
2.147
2.163
2.176
2.186
2.194
2.201
2.207
1,759
1.803
1.831
1.851
1.867
1.888
1.903
1.916
1.925
1,932
1.938
1.944
1.480
1.521
1.547
1.566
1.580
1.600
1.614
1.625
1.634
1.640
1.646
1.652
1.043
1.081
1.104
1.121
1.133
1.151
1.164
1.173
1.181
1.187
1.192
1.196
.625
,661
.683
.699
.710
.726
.738
.747
.753
.759
.763
.767
.222
. 179
.154
.137
.125
.108
.096
.088
.081
.076
.07f
.068
1.155
1.083
1.045
1.020
1.002
.978
.962
‘a 950
. 942
.935
.929
.925
1.671
1.577
1.528
1.497
1.475
1.445
1.425
1.411
1.401
1.392
1.386
1.380
2.104
1.991
1.932
1.895
1.869
1.834
1.811
1.795
1.782
1.772
1.764
1.758
2.927
2.775
2.697
2.648
2.614
2.568
2.538
2.517
2.501
2.489
2.478
2.470
90 10
15
20
25
30
40
50
60
70
80
90
100
2.165
2.273
2.342
2.390
2.426
2,479
2.517
2.544
2.567
2.585
2,600
2.613
1.919
2.019
2.082
2.126
2.160
2,209
2.244
2.269
2.290
2.307
2.321
2.333
1.715
1.808
1.867
1.908
1.940
1.986
2.018
2,042
2.061
2.077
2.090
2.101
1.489
1.576
1.630
1.669
1.698
1.740
1.770
1.792
1.810
1.824
1.836
1.847
1.234
1.314
1.364
1.400
1.427
1.465
1.493
1.513
1.529
1.543
1,553
1.563
.828
.9Ol
.946
.978
1.002
1.036
1.059
1.077
1.091
1.103
1.112
1.120
.429
.499
,541
,570
.593
.624
.645
.662
.674
.684
.693
.7Ol
.437
.347
.297
.264
.239
.206
.184
.167
.155
.144
.136
. 129
1,474
1.320
1.240
1.190
1.154
1.106
1.075
1.052
1.035
1.021
1.010
1.001
2.066
1.867
1.765
1.702
1.657
1.598
1.559
1.532
1.511
1.495
1.481
1,470
2.568
2.329
2.208
2.132
2.080
2.010
1.965
1.933
1.909
1.890
1.874
1.861
3.532
3.212
3.052
2.952
2.884
2.793
2.735
2.694
2.662
2.638
2.618
2.601
*
CONFIDENCE LIMIT
TABLE 91 (CONTINUED)
DEVIATE VALUES FOR NORMAL DISTRIBUTION
EXCEEDANCE PROBABILITY
*
Confi
dence
Level
Systematic
Record
Length
N .002 ,005 ,010 .020 .040 .I00 .200 .500 .800 .900 .950 .990
.95 10
15
20
25
30
40
50
60
70
80
90
100
1.989
2.121
2.204
2.264
2,310
2.375
2.421
2.456
2.484
2.507
2.526
2.542
1.757
1.878
1.955
2.011
2.053
2.113
2.156
2.188
2.214
2,235
2.252
2.267
1.563
1.677
1.749
1.801
1.840
1.896
1,936
1.966
1.990
2.010
2.026
2.040
1.348
1.454
1.522
1.569
1.605
1.657
1.694
1.722
1.745
1.762
1.778
1.791
1.104
1.203
1.266
1.309
1.342
1.391
1.424
1.450
1.470
1.487
1.500
1.512
.712
.802
.858
.898
.928
.970
1.000
1.022
1.040
1.054
1.066
1.077
.317
.406
.460
.497
.525
.565
.592
.612
.629
a642
,652
.662
,580
.455
.387
.342
.310
.266
.237
.216
.199
.186
.175
.166
1.702
1.482
1.370
1.301
1.252
1.188
1,146
1.116
1.093
1,076
1,061
1,049
2.355
2.068
1.926
1.838
1.777
1.697
1.646
1.609
1.581
1.559
1,542
1.527
2.911
2.566
2.396
2.292
2.220
2.125
2.065
2.022
1,990
1.964
1.944
1.927
3.981
3.520
3.295
3.158
3.064
2,941
2,862
2.807
2.765
2.733
2.706
2.684
.99 10
15
20
25
30
40
50
60
70
80
90
100
1.704
1.868
1.974
2.050
2.109
2.194
2.255
2.301
2.338
2.368
2.394
2.416
1,492
1.645
1.743
1.813
1.867
1,946
2.002
2.045
2.079
2.107
2.131
2.151
1.314
1.458
1.550
1.616
1.667
1.741
1.793
1.833
1.865
1.891
1.913
1.932
1.115
1.251
1,336
1.399
1.446
1,515
' 1.563
1.600
1.630
1.653
1.674
1.691
.886
1.014
1.094
1.152
1.196
1.259
1.304
1.337
1,365
1.387
1.405
1.421
.508
.629
.705
.757
,797
.854
.894
.924
.948
.968
.984
.998
107
.236
.313
.364
403
.457
.496
.524
.545
.563
.578
.591
.a92
.678
.568
. 498
.450
.384
.340
. 309
.285
.265
.250
.236
2.243
1.841
1.651
1.536
1,457
1.355
1.290
1.244
1,210
1.183
1.160
1.142
3.048
2.521
2.276
2.129
2.030
1.902
1.821
1.764
1.722
1.688
1.661
1.639
3.738
3.102
2.808
2.633
2.515
2.364
2.269
2.202
2.153
2,114
2.082
2.056
5.074
4.222
3.832
3.601
3.447
3.249
3.125
3,038
2.974
2.924
2.883
2.850
*
Appendix 10
RISK
This appendix describes the recommended procedures for estimating
the risk incurred when a location is occupied for a period of years, As
used in this guide, risk is defined as the probability that one or more
events will exceed a given flood magnitude within a specified period of
years.
Two basic approaches may be used to compute risk, nonparametric
methods [(e.g., (19)] and parametric methods [(e.g., (20)) e Parametric
methods which use the binomial distribution require assuming that the
annual exceedance frequency is exactly known. The difference between
methods is not great, particularly 1n the range of usual interest;
consequently, use of the binomial distribution is recommended because of
ease of comprehension and application.
The binomial expression for estimating risk is:
RI = *PI (1P)
NI (10l)
in which RI is the estimated risk of obtaining in N years exactly I
number of flood events exceeding a flood magnitude with annual exceedance
probability P.
When I equals 0 equation 10l reduces to:
RO = (1P)N (loa)
in which R, is the estimated probability of nonexceedance of the selected
flood magnitude in N years. From this the risk R of one or more exceedance
becomes
R (1 or more) = 1 (~MP)~ (103)
Risk of 2 or more exceedances, R (2 or more), is
R(2 or more) = RR, = RNP (lP)N*l (104)
* Some solutions are illustrated by the following table"and figure *
10l.
10l
*
TIME
50
60
iti
1;:
110
120
150
200
TIME
10
20
iii
6";
70
80
12
110
120
150
200
NOTE:
BINOMIAL RISK TABLE
** RISK (PERCENT) ** ** RISK (PERCENT) **
P=O.lOO P=O.O50
NONE ONE OR TWO OR NONE ONE OR TWO OR
MORE MORE MORE MORE
35 65 60 / 40 9
12 88 tif 1; 64
4 82 ~~
;69 92 13 i; 60
; 8 92 72
0 1;1: ii'9 5 ii; 81
100
i 100 1:: % 87
98 91
0 100 100 ; ;; 94
0 100 100
100 100 0 100 ;ii
ii 100 100
100 100 0 100 190:
ii 100 100 0 100 100
** RISK (PERCENT) ** ** RISK (PERCENT) **
P=O.O40 P=O.O20
NONE ONE OR TWO OR NONE ONE OR TWO OR
MORE MORE MORE MORE
66 34 82 2 2
44 56 1; 67 6
29 71
20 80 2 ii 45 :%
13 87 ii 5:
60 70 26
91 70
ii 94 78
4 96 83 ;; 76 48
80 ii
88
; ii; 91 :"3 8487 i"o
1 99
;: 11 89 ii;
!I 1:: z 91 80
0 100 lFl"o 2 iii 91
TABLE VALUES ARE ROUNDEDTO NEAREST PERCENT
IO2
*
BINOMIAL RISK TABLE
TIME ** RISK F"R"o:No" **
b
NONE ON; OR TWO OR
MORE MORE
:: 0
ii
3: 26 i
33 6
:; 4359 lf
70
80 2 1;
60 23
1:: 63
110 33 67 5:
120
150 ;2" E 4":
200 13 87 60
TIME ** RISK (PERCENT) **
P=O.O02
NONE ONE OR TWO OR
MORE MORE
f
6
1:
:i
15
16
18
** RISKp (J’EOROC$NT’
**
NONE O;E *OR TWO OR
MORE MORE
ii"0 1:
0
8682 ii :
78
E 4"
;t 30
67 33 :
64 fig 8
58 42
61 1;
i; 45 :7"
37 ;3 26
** RISK (PERCENT) **
P=O.OOl
ONE OR TWO OR
MORE MORE
NOTE: TABLE VALUES ARE ROUNDEDTO NEAREST PERCENT
*.
103
1:
20
30
40
50
60
70
80
“,Y
92
93
94
95
96
97
98
99
Figure 10l. RISK OF ONE OR MORE FLOOD EVENTS EXCEEDING
A FIOOD OF GIVEN ANNUAL EXCEEDANCE FREQUENCY OF YEARS
WITHIN A F'EEUOD
1 o4
Appendix 11
EXPECTED PROBABILITY
The principle of gambling based upon estimated probabilities can be
applied to water resources development decisions. However, because
probabilities must be inferred from random sample data, they are uncertain
and mathematical expectation cannot be computed exactly as errors due to
uncertainty do not necessarily compensate. For example, if the estimate
based on sample data is that a certain flood magnitude will be exceeded
on the average once in 100 years, it is possible that the true exceedance
could be three or four more times per hundred years, but it can never be
less than zero times per hundred years. The impact of errors in one
direction due to uncertainty can be quite different from the impact of
errors in the other direction. Thus, it is not adequate to simply be
too high half the time and too low the other half. It is necessary to
consider the relative impacts of being too high or too low,
It is possible to delineate uncertainty with considerable accuracy
when dealing with samples from a normal distribution. Therefore, when
flood flow frequency curves conform fairly closely to the logarithmic
normal distribution, it is possible to delineate uncertainty of frequency
or probability estimates of flood flows.
Figure 11l is a generalized representation of the range of uncertainty
in probability estimates based on samples drawn from a normal population.
The vertical scale can represent the logarithm of streamflow. The
curves show the likelihood that the true frequency of any flood magnitude
exceeds the value shown on the frequency scale, The curve labeled .50
is the curve that would be used for the best frequency estimate of a log
normal population. From this curve a magnitude of 2 would be exceeded
on the average 30 times per thousand events. The figure also shows a 5
percent chance that the true frequency is 150 or more times per thousand
or a 5 percent chance that the true frequency is two times or less per
thousand events.
If a magnitude of 2,O were selected at 20 independent locations,
the best estimate for the frequency is 3 exceedances per hundred years
for each location. The estimated total exceedance for all 20 locations
would be 60 per 100 years. However, due to sampling uncertainties, true
frequencies for a magnitude of 2.0 would differ at each location and
total exceedances per 100 years at the 20 locations might be represented
by the following tabulation.
Exceedances Per 100 Years at Each of 20 Locations*
20 5 3 .9
12 5 2 .8
10 4 2 .5 Total Exceedances = Approximately 90
8 4 2 .3
7 3 1 .1
*Determined from Figure 11l using 0.05 parameter value increments
from .025 through .975.
The total of these exceedances is about 90 per 100 years or 30 more than
obtained using the best probability estimate as the true probability at
each location. If, however, the mathematically derived expected proba
bility function were used instead of the traditional "best" estimate we
could read the expected probability curve of Figure llltto obtain the
value of about 4.5 exceedances per 100 events. This value when applied
to each of the 20 locations would give an estimate of 90 exceedances per
100 years at all 20 locations. Thus, while the expected probability
estimate would be wrong in the high direction more frequently than in
the low direction, the heavier impacts of being wrong in the low direction
would compensate for this. It can be noted, at this point, that expected
probability is the average of all estimated true probabilities,
If a flood frequency estimate could be accurately knownthat is,
the parent population could be definedthe frequency distribution of
observed flood events would approach the parent population as the
number of observations approaches infinity. This is not the case where
probabilities are not accurately known. Howeverp if the expected
probabilities as illustrated in Figure 11l can be computed, observed
112
flood frequency for a large number of independent locations will approach
the estimated flood frequency as the number of observations approaches
infinity and the number of locations approaches infinity.
It appears that the answer to the question as to whether expected
probability should be used at a single location would be identical to
the answer to the question, "What is a fair wager for a single gamble?"
If the gamble must be undertaken, and ordinarily it must9 then the
answer to the above question is that the wager should be proportional to
the expected return. In determining whether the expected probability
concepts should apply for a single location, the same line of reasoning
would indicate that it should.
It has been shown (21) that for the normal distribution the expected
probability PN can be obtained from the formula
pN = Prob
c$jl b Kn (&)
l/2
I
(11l)
where K, is the standard normal variate of the desired probability
of exceedance, N is the sample size, and tN1 is the Student's tsta
tistic with Nl degrees of freedom.
The actual calculations can be carried out using tables of
the tstatistic, or the modified values shown in Table 11l (31).
To use Table 111, enter with the sample size minus 1 and read
across to the column with the desired exceedance probability. The
value read from the table is the corrected plotting position.
The expected probability correction may also be calculated
from the following equations (34) which are based on Table 11l.
For selected exceedance probabilities greater than 0.500 and a
given sample size, the appropriate PN value equals 1 minus the value in
Table 11l or the equations 112,
113
Exceedance Probability Expected Probability, PN
.OOOl .OOOl (1.0 + 1600/N1'72) (11*a)
,001 .OOl (1.0 + 280/N1*55) (112b)
.Ol a01 (1.0 + 26/N1*16) (ll2c)
.05 .05 (1.0 + 6/N1004) (112d)
.lO ,l (1.0 + 3/Nloo4) (112e)
.30 .3 (1.0 + 0.46/Noog25) (lL2f)
For floods with an exceedance probability of 0.01 based on
samples of 20 annual peaks, for example, the expected probability
of exceedance from equation 112~ ds (.Ol) (1.0 + 26/32.3) or 0.018.
Use of Table 11l gives 0.0174. Comparable equations for adjusting the
computed discharge upward to give a discharge for which the expected
probability equals the exceedance probability are available (22).
114
I I I
Note: Parameter IS relative frequency with which
true value exceeds the indicated value as the
number of random samples of this size ap
proaches Infinity.
w I i I/Y Y.
Figure 11l
PRO6ABILlTY ESTIMATES
FROM
NORMAL DISTRIBUTION SAMPLE
N=lO
II
99.9 99.8 99.5 99 98 95 90 80 60 30 20 10 5 2 1
EXCEEDANCEFREQUENCY,IN PERCENT
Table 11l
TABI OF PN VERSUS PO0
For use with samples drawn from a normal populatSon
?XOTE: pN values above are usable approximately with Pewson Type III
distributions having small skew coefficients.
Appendix 12
FLOW DIAGRAM AND EXAMPLE PROBLEMS
*
The sequence of procedures recommended by this guide for defining flood
potentials (except for the case of mixed populations) is described in
the following outline and flow diagrams.
A. Determine available data and data to be used.
1. Previous studies
2. Gage records
3. Historic data
4. Studies for similar watersheds
5. Watershed model
B. Evaluate data.
1. Record homogeneity
2. Reliability and accuracy
c. Compute curve following guide procedures as outlined in following
flow diagrams. Example problems showing most of the computational
techniques follow the flow diagram.
12l
ZEROCFLOOO
* INCOMPLETE RECORD COMPLETE RECORD
SEE APPENDIX 5, COMPUTE STATION
cotaITloNAL STAnSTlCS I
I
PROBAl3lLllY
ADJUSTMENT, FOR
OUTLIERS SEE
PAGES 17 TO 19
AFY, APPENDIX
BAND 6
COMPUTE EXTENDED
RECORD APPENDIX 7
4t IF SYSTEMATIC RECORD LENGTH IS
LESS THAN 50 YEARS THE ANALYST
SHOULD CONSIDER WHETHER THE
USE OF THE PROCEDURES OF
APPENDIX 7 IS APPROPRIATE.
NOR3 IS FURlHER ANALYSIS WARRANTED@
STEPS TO THIS POINT ARE BASIC
STEPS REQUIRED IN ANALYSIS OF
READILY AVAllpBLE STAnON AND
HISTORIC OATA. AT THIS POiNT A
DECISION SHOULD BE MADE AS TO
WHETHER FUTURE FURTHER REFINEFINAL CURVE
hwr w niiz FREQUENCY mwm
Lrzl
IS JUSTIFIED. MIS DECISION WlLL
DEPEMJ BOTH UPON TIME AND
EFFORT REQUlREO FOR REFINEMENT IF DESIRED
ANO UPON THE PURPOSE OF THE
bl
FREQUENCY ESTIMATE.
FLOW DIAGRAM FOR FLOOD FLOW FREQUENCY ANALYSIS
*
122
*FLOW DIAGRAM FOR HISTORIC AND OiJTLIER ADJUSTMENT
RECOMPUTE
sTm+m;;~s
LOW
OUTLIERS
1YES
\r
NO
1 YES
I
YES
RECOMPUTE
STATISTICS
ADJUSTEb FOR
HISTORIC PEAKW
HIQH OUTLIERS
APPENDIX 8
i
RECOMPUTE
ST;;+m;:~s
LOW
OUTLIERS
L
RECOMPUTE
ST;;+/W;;~S
LOW
OUTLI ERS
I
I
CONDITIONAL
PROBABILITY
ADJUSTMENT
APPENDZX 6
The following examples illustrate application of most of the
techniques recommended in this guide. Annual flood peak data for
four statifns (Table 12l) have been selected to illustrate the following:
1. Fitting the LogPearson Type III distribution
2. Adjusting for high outliers
3. Testing and adjusting for low outliers
4. Adjusting for zero flood years
The procedure for adjusting for historic flood data is given
in Appendix 6 and an example computation is provided. An example
has not been included specifically for the analysis of an incomplete
record as this technique is applied in Example 4, adjusting for zero
flood years. The computation of confidence limits and the adjustment
for expected probability are described in Example 1. The generalized
*skew coefficient used in these examples was taken from Plate I.
In actual practice, the generalized skew may be obtained from other
sources or a special study made for the region.
Because of round off errors in the computational procedures,
computed values may differ beyond the second decimal point.
*
These examples have been completely revised using the procedures
recommended in Bulletin 17B. Specific changes have not been indicated on
the following pages:
124
TABLE 12l
ANNUAL FLOOD PEAKS FOR FOUR STATIONS IN EXAMPLES
Fishkill Creek Floyd River
013735 066005
fear Example 1 Example 2
1929
I930
1931
1932
1933
1934
1935 1460
I936 4050
I937 3570
I938 2060
1939 1300
I940 1390
I941 1720
1942 6280
I943 1360
1944 7440
I945 2290 5320
1946 1470 1400
1947 2220 3240
1948 2970 2710
1949 3020 4520
I95C 1210 4840
1951 2490 8320
I952 3170 13900
1953 3220 71500
I954 1760 6250
1955 8800 2260
1956 8280 318
1957 1310 1330
195E 2500 970
1959 1960 1920
l96C 2140 15100
1961 4340 2870
1962 3060 20600
1963 1780 3810
1964 1380 726
1965 980 7500
1966 1040 7170
1967 1580 2000
I968 3630 829
1969 "I17300
197c 4740
1971 13400
1972 2940
197: 5660
*Not included in example computations.
Back Creek
016140
Yw
15500
4060
l
22000*
6;OO
3130
4160
6700
22400
3880
8050
4020
1600
4460
4230
3010
9150
5100
9820
6200
10700
3880
3420
3240
6800
3740
4700
4380
5190
3960
5600
4670
7080
4640
536
6680
8360
18700
5210
Orestimba Creel
112745
Example 4
T
L
4260
345
516
1320
1200
2150
3230
115
3440
3070
1880
6450
1290
5970
782
00
335
175
2920
3660
147
0
56::
1440
10200
5380
448
174:
8300
156
560
128
4200
0
5080
1010
584
151:
EXAMPLE 1
FITTING THE LOGPEARSON TYPE III DISTRIBUTION
a. Station Description
Fishkill Creek at Beacon, New York
USGS Gaging Station: 013735
Lat: 41"30'42", long: 73'56'58"
Drainage Area: 190 sq. mi.
Annual Peaks Available: 19451968
b. Computational Procedures
Step 1 List
the
data,
cubes.
transform to logarithms, and compute the squares and
Year
194194:
1947
1948
1949
TABLE 122
COMPUTATION OF SUMMATIONS
Annual Peak Logarithm
(cfs) (x)
2290 3.35984
1470 3.16732
2220 3.34635
2970 3.47276
3020 3.48001
X2
10:031:2
11.19806
12.06006
12.11047
X3
37.92764
31.77429
37.47262
41 .88170
42.14456
1950
1951
1952
1953
1954
1210
2490
3170
3220
1760
' 3.08279
3.39620
3.50106
3.50786
3.24551
9.50359
11.53417
12.25742
12,30508
lo,53334
29.29759
39.17236
42.91397
43.16450
34018604
1955
1956
1957
1958
1959
8800
8280
1310
2500
1960
3.94448
3.91803
3.11727
3.39794
3.29226
15.55892
15.35096
9,71737
11.54600
10.83898
61.37186
60.14552
30,29167
39.23260
35.68473
1960
1967
1962
1963
1964
2140
4340
3060
1780
1380
3.33041
3.63749
3.48572
3.25042
3e13988
11.09163
13.23133
12.15024
10.56523
9.85885
36.93968
48.12884
42.35235
34.34144
30.95559
1965
1966
1967
1968
N=24
980
1040
1580
3630
MS
2.99123
3.01703
3.19866
3.55991
C 80.84043
8.94746
9.10247
10.23143
12.67296
273.68646
26.76390
27.46243
32.72685
45.11459
931.44732
126
Example 1 Fitting the LogPearson Type I II Distribution (continued)
Step 2 Computation of mean by Equation 2:
= !%@$!t? = 3.3684 (12l)
Computation of standard deviation by Equation 3b:
S = px2
[
S
0.5
N(EX)2/N 
1
(80.84043)'/24
23
0.5 1 (12Z)
S = d= 0.2456
Computation of skew coefficient by Equation 4b:
G=
N2EX3) 3N(CX)EX2)
N(Nlj(N2)S3
f 2 &Xl3
= (2412(931.44732)
24124l)
3(24)(80.84043)(273.68646)
(24Z) 6.24561)3
+ 2(80.84043)3
= 536513.6563 1592995.0400 f 1056612.7341
(24) (23) (22) t.014816)
(123)
=
131.3504im = 0.7300
127
Example I Fitting the LogPearson Type III Distribution (continued)
Step 3 Check for Outliers:
xH = i+ KNS
= 3.3684 + 2.467 1.2456) = 3.9743 (124)
QH = antilog (3.9743) = 9425 cfs
The largest recorded value does not exceed the threshold value. Next,
the test for detecting possible low butliers is applied. The same K,,,
value is used in equation 8a to compute the low outlier threshold (0,~:
XL = x KNs
= 3.3684 2.467i.2456) = 2.7625 (125)
QL = antilog (2.7625) = 579 cfs
There are no recorded values below this threshold value. No outliers
were detected by either the high or low tests. For this example a
generalized skew of 0.6 is determined from Plate I. In actual practice
a generalized skew may be obtained from other sources or from a special
study made for the region. A weighted skew is computed by use of
Equation 5. The mean square error of the station skew can be found
within Table 1 or computed by Equation 6. Computation of meansquare
error of station skew by Eq. 6:
1o A 0 Cl oglOW~)
MSEC ci 11
c
Where:
A= 0.33 + 0.08 IGI = 0.33 + 0.08(.730) = .2716 (126)
0 = 0.94 0.26 IGI = 0.94 0.26(.730) = .7502 (127)
MSEC p' IO C se2716 7502 ~~qok4)+ 1o s.55683 5 0.277 (12R')
128
Example 1 Fitting the LogPearson Type III Distribution (continued)
The meansquare error of the generalized skew from Plate I is 0.302.
Computation of weighted skew by equation 5:
G, = MSPg (G) I MSEG(G)
MSEE + MSEG
= .302( .73;;4 + .277(.6) = 0.6678 (129)
= 0.7 (rounded to nearest tenth)
Step 4 Compute the frequency curve coordinates.
The logPearson Type III K values for a skew coefficient of 0.7 are
found in Appendix 3. An example computation for an exceedance
probability of .Ol using Equation 1 follows:
log Q = x f KS = 3.3684 + 2.82359(.2456) = 4.0619 (1210)
Q= 11500 cfs
The discharge values in this computation and those in Table 123 are
rounded to three significant figures.
129
Example 1 Fitting the LogPearson Type III Distribution (continued)
TABLE 123
COMPUTATION OF FREQUENCY CURVE COORDINATES
'Gw,P
P for Gw = 0.7 log Q Q
cfs
.99 1.80621 2.9247 841
.90 1.18347 3.0777 1200
.50 0.11578 3.3399 2190
.lO 1.33294 3.6957 4960
.05 1.81864 3.8150 6530
.02 2.40670 3.9595 9110
01 2.82359 4.0619 11500
.005 3.22281 4.1599 14500
.002 3.72957 4.2844 19200
The frequency curve is plotted in Figure 12l.
Step 5 Compute the confidence limits.
The upper and lower confidence limits for levels of significance of
.05 and .95 percent are computed by the procedures outlined in
Appendix 9. Nine exceedance probabilities (P) have been selected to
define the confidence limit curves. The computations for two points
on the curve at an exceedance probability of 0.99 are given below.
1210
Computed Frequency Curve
With Expected Probability
Confidence Limit
:CEEDANCE PR4BABiITY
Figure 121
Frequency Curves for
Fishkill Creek at
Beacon, New York
Example 1
12l 1
Example 1 Fitting the LogPearson Type III Distribution (continued)
Equations in Appendix 9 are used in computing an approximate value for
The normal deviate, zc, is found by entering Appendix 3 with a
KP,c
skew coefficient of zero. For a confidence level of 0.05, zc = 1.64485.
The Pearson Type III deviates,KG p are found in Appendix 3 based on
WP
the appropriate skew coefficient, For an exceedance probability of 0.99
and skew coefficient of 0.7, KG p = 1.80621.
W’
22
a = l,b = 1 .q$$yp* = 0.9412 (12U)
22
b = K2 + (1.80621)2 b644W224 3,14g7 (1212)
G,,” 
# KGw,P+dF l.80621)2(.9412)(3.1497)
KP,c = a .9412
(1213)
pl.80621 I.5458 = , . 33g2
.9412
The discharge value is:
Log Q = 3.3684 + (1,3392)(,2456) (1214)
= 3.0395
Q = 1100
For the lower confidence coefficient:
I?KGwaF "&,P ab
= 1.80621 .5458 = * e 4989 (1275)
P,C = a .94i*
1212
Example 1 Fitting the LogPearson Type III Distribution (continued) ).'.,
The discharge value is:
Log Q = 3.3684 + (2.4989)(.2456) (1216)
= 2.7546
Q = 568
The computations showing the derivation of the upper and lower confi
dence limits are given in Table 124. The resulting curves are shown
in Figure 12l.
TABLE 124
COMPUTATION OF CONFIDENCE LIMITS
KGw,B 0.05 UPPER LIMIT CURVE 0.05 LOWER LIPiIT CURVE
P for G, = 0.7
'b
U
SC log Q cfs
Q IbL
9C log Q cfs
Q
.99 1.80621 1.3392 3.0395 1100 2.4989 2.7546 568
.90 1.18347 0.7962 3.1728 1490 1.7187 2.9462 884
50 0.11578 0.2244 3.4235 2650 0.4704 3.2528 1790
.lO 1.33294 1.9038 3.8359 6850 0.9286 3.5964 3950
.05 1.81864 2.5149 3.9860 9680 1.3497 3.6998 5010
.02 2.40670 3.2673 4.1708 14800 1.8469 3.8220 6640
.Ol 2.82359 3.8058 4.3031 20100 2.1943 3.9073 8080
.005 3.22281 4.3239 4.4303 26900 2.5245 3.9884 9740
.002 3.72957 4.9841 4.5925 39100 2.9412 4.0907 12300
Step 6 Compute the expected probability adjustment.
The expected probability plotting positions are determined from
Table 11l based on N 1 of 23.
1213
Example 1 Fitting the LogPearson Type III Distribution (continued)
TABLE 125
EXPECTED PROBABILITY ADJUSTMENT
Expected
P Q Probability
.99 841 .9839
.90 1200 .889
.50 2190 .50
.lO 4960 .111
.05 6530 0060
.02 9110 .028*
.Ol 11500 .0161
.005 14500 e 0095”
.002 19200 * 0049”
*Interpolated values
The frequency curve adjusted for expected probability is shown
in Figure 12l.
1214
EXAMPLE 2
ADJUSTING FOR A HIGHOUTLIER
a. Station Description
Floyd River at James, Iowa
USGS Gaging Station: 066005
Lat: 42o34'30", long: 960 18'45"
Drainage Area: 882 sq. mi.
Annual Peaks Available: 19351973
b. Computational Procedures
Step 1 Compute the statistics.
The detailed computations for
have been omitted; the results
the
of
systthe
ematic
computations
record 19351973
are:
Mean Logarithm
Standard Deviation
Skew Coefficient
Years
of logs
of logs
3.5553
0.4642
0.3566
39
At this point, the analyst may wish to see the preliminary
frequency curve based on the statistics of the systematic
record. Figure 122 is the preliminary frequency curve based
on the computed mean and standard deviation and a weighted
skew of 0.1 (based on a generalized skew of 0.3 from Plate I).
Step 2 Check for Outliers.
The station skew is between + 0.4; therefore, the tests for
both high outliers and low oaliers are based on the systematic
record statistics before any adjustments are made. From
Appendix 4, the KN for a sample size of 39 is 2.671.
The high outlier threshold (QH) is computed by Equation 7:
= si;+ KNS
xH
= 3.5553 f 2.671(.4642) = 4.7952 (1217)
QH = antilog (4.7952) = 62400 cfs
1215
rti 0
Observed Annual Peaks
Preliminary Frequency Curve
(Systematic record with
‘weighted skew)
v
#XEDANCE PR~BADilTY ’
Figure 122
Preliminary
Frequency Curve for
Floyd River at James, Iowa
Example 2
1216
Example 2 Adjusting for a High Outlier (continued)
The 1953 value of 71500 exceeds this value. Information from local
residents indicates that the 1953 event is known to be the largest
event since 1892; therefore, this event will be treated as a high
outlier. If such information was not available, comparisons with
nearby stations may have been desirable.
The lowoutlier threshold (QL) is computed by Equation 8a:
xL
= x KNS
= 3.5553 2.671(.4642) = 2.3154 (1218)
QL
= antilog (2.3154) = 207 cfs
There are no values below this threshold value.
Step 3 Recompute the statistics.
The 1953
remaining
value is
systematic
deleted and
record:
the statistics recomputed from the
Mean Logarithm
Standard Deviation
Skew Coefficient
Years
of logs
of logs
3.5212
0.4177
0.0949
38
Step 4 Use historic data to modify statistics and plotting positions.
Application
statistics
of the procedures
to be adjusted by
in Appendix
incorporation
6 allows
of the
the computed
historic data.
(1) The
and
historic period
the number of
(H) is
low values
18921973
excluded
or 82 years
(L) is zero.
(2) The systematic
or 38 years.
period (N) is 19351973 (with 1953 deleted)
(3) There is one event (Z) known to be the largest in 82 years.
(4) Compute
14
weighting
= E
factor (W) by Equation 6l:
= 82l
38 + 0
= 2.13158
1217
(1219)
Example 2 Adjusting for a High Outlier (continued)
Compute adjusted mean by Equation 62b:
‘L
M
x
WNM
cxz
'L
M
=
f
=
=
=
WNM I cXz
HWL
M = 3.5212
285.2173
4.8543
290.0716
290.0716/(82O) = 3.5375 (1220)
Compute adjusted standard deviation by Equation
:2
%L2 %2
=
W(Nl)S2 + WN(MM) SC (Xz M)
HWL1
s = .4177
63b:
W(Nl)S2
%L2WN(MM)
:2c(XzPI)
=
=
=
13.7604
.0215
1.7340
15.5159
%
s =
15*515g
82Ol
.4377
= Jg,fj (1221)
Compute adjusted skew:
First compute adjusted skew on basis of record by Equation 64b:
1218
Example
2 Adjusting for a High Outlier (continued)
s H WL
G = +
(H&l)(HWL2)'
%3
f WN(MM)
G = 0.0949
W(N1)(N2)S3G= .5168
N
3W(Nl)(M;)S2 = .6729
%3
WN(MM) = a0004
= 2.2833
1 .a932
H
= *I509
(HWLl)(HWL;)33
G = .1509 (1.0932) = .1650
Next compute weighted skew:
For this example, a generalized skew of 0.3 is
Plate I. Plate I has a stated meansquare error
Interpolating in Table I, the meansquare error
based on H of 82 years, is 0.073. The weighted
use of Equation 5:
G, = .302(.1650) + +073(.3) o
00745
.302 f .073
= 0.1 (rounded to nearest tenth)
GW
1219
'L 2
3W(Nl)(MM)S
IL3
+X(X, M)
3
(1222)
determined from
of 6.302.
of the station skew,
skew is computed by
(1223)
Example 2 Adjusting for High Outlier (continued)
Step 5 Compute adjusted plotting positions for historic data.
For the largest event (Equation 66):
iii,= 1
For the succeeding events (Equation 67):
i;; = W E (Wl)(Z I 0.5)
d
m2 = 2.1316(2) (2.13161111 * .5) (1224)
= 2.5658
For the Weibull Distribution a = 0; therefore, by Equation 68
p”p= L (100)
H+l
1
P"p = (100) = 1.20 (1225)
1 82+1
PT2 = y (100) = 3.09 (1226)
Exceedance probabilities are computed by dividing values obtained from
Equation 1226 by 100.
TABLE 126
COMPUTATION OF PLOTTING POSITIONS
Weibull Plottino
Position
Event Weighted Percent Exceedance
Number Order Chance Probability
Year Q W E m p"p 66
1953 71500 1 .oooo 1 1 .oooo 1.20 .0120
__I___________________I______
1962 20600 2.1316 2 2.5658 3.09 0309
1969 17300 2e1316 3 4.6974 5.66 .0566
1960 15100 2,1316 4 6.8290 8.23 .0823
1952 13900 2.1316 5 8.9606 10.80 .1080
1971 13400 2.1316 6 11.0922 13.36 .1336
1951 8320 2.1316 8' 13.2238 15.93 .1593
1965 7500 2.1316 15.3554 18.50 .1850
1944 7440 2.1316 lo" 17.4870 21.07 .2107
1966 7170 2.1316 19.6186 23.64 .2364
Only the first 10 values are shown for this example
1220
Example 2 Adjusting for a High Outlier (continued)
Step 6 Compute the frequency curve.
TABLE 127
COMPUTATION OF FREQUENCY CURVE COORDINATES
KG, A’
P
.99
.90
.50
.lO
.05
.02
.Ol
.005
.002
The final
for G, =
2.25258
1.27037
0.01662
1.29178
1.67279
2.10697
2.39961
2.66965
2.99978
frequency
0.1 log Q Q
cfs
2,5515 356
2.9815 958
3.5302 3390
4.1029 12700
4.2697 18600
4.4597 28800
4.5878 38700
4.7060 50800
4.8504 70900
curve is plotted on Figure 123.
12a
Observed Peaks with
Weighted Plotting Positions II Illltt '
Final Frequency Curve
II I I I I I II III
II I I1 I I II III
II I, I,,,
II 11 I I I II II
CEEDANCE PRO.BAllilTI’
Figure 123
Final Frequency Curve for
Floyd River at James, Iowa
Example 2
1222
EXAMPLE 3
TESTING AND ADJUSTING FOR A LOW OUTLIER
a. Station Description
Back Creek near Jones Springs, West Virginia
USGS Gaging Station: 016140
Lat: 39030'43", long: 78oO2'15"
Drainage Area: 243 sq. mi.
Annual Peaks Available: 192931, 19391973
b. Computational Procedures
Step 1  Compute the statistics of the systematic record.
The detailed computations have been omitted; the results of the
computations are :
Mean Logarithm 3,722O
Standard Deviation of logs 0.2804
Skew Coefficient of logs 0.7311
Years 38
At this point the analyst may be interested in seeing the preliminary
frequency curve based on the statistics of the systematic record.
Figure 124 is the preliminary frequency curve based on the computed
mean and standard deviation and a weighted skew of 0.2 (based on a
generalized skew of 0.5 from Plate I).
Step 2  Check for outliers.
As the computed skew coefficient is less than 0.4, the test for
detecting possible low outliers is made first. From Appendix 4,
the KN for a sample size of 38 is 2.661.
1223
Observed Annual Peaks
Preliminary Frequency Curve
(Systematic record with
ICEEDANCE ?i?~DABl;lTY~
Back
Figure 124
Preliminary
Frequency Curve for
Creek nr. Jones Springs, W. VA.
Example 3
1224
Example 3 Testing and Adjusting for a Low Outlier (continued)
The low outlier threshold is computed by Equation 8a:
XL = ii KNS
QL
=
=
3.7220
antilog
2.661
(2.9759)
(.2804) =
=
2.9759
946 cfs
(1227)
The
and
1969
will
event of
be treated
536 cfs
as a
is
low
below
outlier.
the threshold value of 946 cfs
Step 3 DelPte the low outlier(s) and
Mean Logarithm
Standard Deviation of logs
Skew Coefficient of logs
Years
recompute
3.7488
0.2296
0.6311
37
the statistics.
Step 4 Check for high outliers.
The highoutlier threshold
statistics in Step 3 and
events exceed the threshold
computations to determine
is computed to be 22,760 cfs
the sample size of 37 events.
value. (See Examples 1 and
the highoutlier threshold.)
based on the
No recorded
2 for the
Step 5 Compute and adjust conditional frequency curve.
A conditional frequency curve is computed based
in Step 3 and then modified by the conditional
on the
probability
statistics
adjustment
1225
Example 3 Testing and Adjusting for a Low Outlier (continued)
(Appendix 5). The skew coefficient has been rounded to 0.6 for ease
in computation. The adjustment ratio computed from Equation 5la is:
'L
P = N/n = 37/38 = 0.9737 (1228)
TABLE 128
COMPUTATION OF CONDITIONAL FREQUENCY CURVE COORDINATES
Kc Adjusted"d Exceed? se
for G = 0.6 log Q Q Probabr r sty
pd
cfs WPd)
.99 1.88029 3.3171 2080 .9639
.90 1.20028 3.4732 2970 .876
.50 0.09945 3.7260 5320 .487
.lO 1.32850 4.0538 11300 ,097
.05 1.79701 4.1614 14500 .049
.02 2.35931 4.2905 19500 .0195
.Ol 2.75514 4.3814 24100 * 0097
.005 3.13232 4.4680 29400 s 0049
.002 3.60872 4.5774 37800 .0019
The conditional frequency curve% along with the adjusted frequency
curve, is plotted on Figure 125.
1226
11 I I I II III
I I II III
I’ 1 I
I I I I I II III,,,
Conditional Frequency Curve
Frequency Curve with Conditional
Probability Adjustment
CEEDANCE PRqlABlLlTY
Figure 125
Adjusted Frequency Curves for
Back Creek nr. Jones Springs, W. VA,
Example 3
1227
Example 3 Testing and Adjusting for a Low Outlier (continued)
Step 6 Compute the synthetic statistics.
The statistics of the adjusted frequency curve are unknown.
The use of synthetic statistics provides a frequency curve
with a logPearson Type III shape. First determine the Q~,,,Q~,,
and Q.5, discharges from the adjusted curve on Figure 125.
,
4.01
= 23880
4.10
= 11210
4.50 =
5230
Next, compute the
cfs
cfs
cfs
synthetic skew coefficient by Equation 53.
GS
= 2.50 + 3.12 1w~4~,,/4~,,)
'"g(Q.10'Q.50)
= 2.50 + 3.12 H
(1229)
= 2.50 + 3.12 :;;;;;
= 0.5948
1228
Example 3 Testing and Adjusting for a Low Outlier (continued)
Compute the synthetic standard deviation by Equation 54.
sS = ‘o~(Q~o,/Q~50~/(K,olK~50>
= log (23880/5230)/~.75514(.09945)] (W30)
s, = .6595/2.8546 = 0.2310
Compute the synthetic mean by Equation 55.
% = log (Qs5,) I(s50(Ss)
Step 7 
= log (5230) (.09945)(.2310)
5
= 3.7185 + .0230 = 3.7415
Compute the weighted skew coefficient.
The meansquare error of the station
based on n = 38 and using Gs for G
Gw = .302(0.5948) + .183(.5) =
.302 + .183
skew, from
o 55go
.
Table 1,
(1231)
is 0.183
(1232)
GW
= 0.6 (rounded to nearest tenth)
1229
Example 3 Testing and Adjusting for a Low Outlier (continued)
Step 8  Compute the final frequency curve.
TABLE 129
COMPUTATION OF FREQUENCY CURVE COORDINATES
P for G, = 0.6 log Q Q
cfs
.99 1.88029 3.3072 2030
.90 1.20028 3.4642 2910
.50 0.09945 3.7185 5230
.lO 1.32850 4.0484 11200
.05 1.79701 4.1566 14300
.02 2.35931 4.2865 19300
"01 2.75514 4.3780 23900
.005 3.13232 4.4651 29200
,002 3.60872 4.5751 37600
The final frequency curve is plotted on Figure 126
Note: A value of 22,000 cfs was estimated for 1936 on the basis of data
from another site. This flow value could be treated as historic
data and analyzed by the producers described in Appendix 6. As
these computations are for illustrative purposes only, the remaining
analysis was not made.
1230
Observed Annual Peaks
Final Frequency Curve
9
:CEEDANCL PROBABi,TI
Figure 126
Final Frequency Curve for
Back Creek nr. Jones Springs, W. VA,
Example 3
1231
EXAMPLE 4
ADJUSTING FOR ZERO FLOOD YEARS
a. Station Description
Orestimba Creek near Newman, California
USGS Gaging,Station: 112745
Lat: 37O19 Ol", long: 121°07'39"
Drainage Area: 134 sq. mi.
Annual Peaks Available: 19321973
b. Computational Procedures
Step 1 Eliminate zero flood years.
There are 6 years with zero flood events, leaving 36 nonzero events.
Step 2 Compute the statistics of the nonzero events.
Mean Logarithm 3.0786
Standard Deviation of logs 0.6443
Skew Coefficient of logs 0.8360
Years (NonZero Events) 36
Step 3 Check the conditional frequency curve for outliers.
Because the computed skew coefficient is less than 0.4, the test for
detecting possible low outliers is made first. Based on 36 years, the
lowoutlier threshold is 23.9 cfs. (See Example 3 for lowoutlier
threshold computational procedure.) The 1955 event of 16 cfs is
below the threshold value; therefore, the event will be treated as a
lowoutlier and the statistics recomputed.
Mean Logarithm 3.1321
Standard Deviation of logs 0.5665
Skew Coefficient of logs 0.4396
Years (Zero and low
outliers deleted) 35
1232
Example 4 Adjusting for Zero Flood Years (continued)
Step 4 Check for high outliers
The high outlier threshold is computed to be 41,770 cfs based on the
statistics in Step 3 and the sample size of 35 events. No recorded
events exceed the threshold value. (See examples 1 and 2 for the
computations to determine the highoutlier threshold.)
Step 5 Compute and adjust the conditional frequency curve.
A conditional frequency curve is computed based on the statistics
in step 3 and then adjusted by the conditional probability adjustment
(Appendix 5). The skew coefficient has been rounded to 0.4 for ease
in computation. The adjustment ratio is 35/42 = 0.83333.
TABLE 1210
COMPUTATION OF CONDITIONAL FREQUENCY CURVE COORDINATES
Adjusted
KG,P Exceedance
'd for G = 0.4 log Q Q
cfs
Probability
(P.P,)
.99 2.61539 1.6505 44.7 .825
.90 1.31671 2.3862 243 .750
.50 0.06651 3.1698 1480 .417
.lO 1.23114 3.8295 6750 .083
.05 1.52357 3.9952 98900 .042
.02 1.83361 4.1708 14800 .017
.Ol 2.02933 4.2817 19100 .0083
.005 2.20092 4.3789 23900 .0042
.002 2.39942 4.4914 31000 .0017
Both frequency curves are plotted on Figure 127.
1233
Observed Peaks
Based on 36 Years
Conditional Frequency Curve
(Without zero and lowoutlier
events)
Frequency Curve with
Conditional Probabilitv
Adjustment
II I I I I I
I I I I I Irl
II I I I I
I I,
ICEEDANCE PRO.BABlLlTY
Figure 127
Adjusted Frequency Curves for
Orestimba Creek nr. Newman, CA
Example 4
1234
Example 4 Adjusting for Zero Flood Years (continued)
Step 6 Compute the synthetic statistics.
First determine the Q,O,,Q.,O, and Q50 discharges from the adjusted
curve on Figure 127.
= 17940 cfs
Q.01
6000 cfs
Q.10 =
1060 cfs
Q.50 =
Compute the synthetic skew coefficient by Equation 53.
G, = 2.50 + 3.12 !og(17g40/6000) = 0.5287 (1233)
log(6000/1060)
Gs = 0.5 (rounded to nearest tenth)
Compute the synthetic standard deviation by Equation 54.
ss = log(l7940/1060)/(1.95472 .08302) (1234)
= 0.6564
sS
Compute the synthetic mean by Equation 55.
= log(lO60) (.08302)(.6564) (1235)
G
= 2.9708
Y
Step 7 Compute the weighted skew coefficient by Equation 5.
A generalized skew of 0.3 is determined from Plate I From Table I,
the meansquare error of the station skew is 0.163.
Gw = .302(.529) + .163(.3) = o
.
4487
(1236).302 + .163
Gw = 0.4 (rounded to nearest tenth)
1235
Example 4 Adjusting for Zero Flood Years (continued)
Step 8 Compute the final frequency curve.
TABLE 1211
COMPUTATION OF FREQUENCY CURVE ORDINATES
KGa
wp
P for Gw = 0.4 log Q Q
cfs
.99 2.61539 1.2541 17.9
.90 1.31671 2.1065 128
.50 0.06651 3.0145 1030
.lO 1.23114 3.7789 6010
.05 1.52357 3.9709 9350
.02 1.83361 4.1744 14900
.Ol 2.02933 4.3029 20100
.005 2.20092 4.4155 26000
.002 2.39942 4.5458 35100
This frequency curve is plotted on Figure 128. The adjusted frequency
derived in Step 4 is also shown on Figure 128. As the generalized skew
may have been determined from stations with much different characteristics
from the zero flood record station, judgment is required to determine the
most reasonable frequency curve.
1236
Observed Peaks
Based on 42 Years
 Final Frequency Curve
Frequency Curve at
CEEDANCE PROIAIILITI
Figure 128
Frequency Curves for
Orestimba Creek nr. Newman, CA
Example 4
1237
Appendix 13
COMPUTER PROGRAM
+
Programs have been developed that compute a logPearson Type III
distribution from systematically recorded annual maximum streamflows at
a single station and other large known events. Special routines are
included for managing zero flows and very small flows (outliers) that would
distort the curve in the range of higher flows. An option is included to
adjust the computed curve to represent expected probability. Copies of
agency programs that incorporate procedures recommended by this Guide may
be obtained from either of the following:
Chief Hydrologist
U,S. Geological Survey,
National Center, Mail
Reston, VA 22092
WRD
Stop 437
Hydrologic Engineering
U.S, Army Corps of En609 2nd Street, Suite
Davis, CA 95616
gineers
I
Center
Phone: (703) 8606879 Phone: (916) 7561104
There is no specific recommendation to utilize these particular computer
programs. Other federal and state agencies as well as private organizations
may have developed individual programs to suit their specific needs. +
13l
Appendix 14
"FLOOD FLOW FREQUENCY TECHNIQUES"
REPORT SUMMARY
*
Following is a summary of "Flood Flow Frequency Techniques," a
report by Leo R. Beard, Technical Director, Center for Research in Water
Resources, The University of Texas at Austin, for the Office of Water
Resources Research and the Water Resources Council. Much of the text
and a majority of the exhibits are taken directly from the report,
The study was made at the Center for Research in Water Resources of
The University of Texas at Austin at the request of and under the general
guidance of the Work Group on Flood Flow Frequency, Hydrology Committee,
of the Water Resources Council through the auspices of the Office of
Water Resources Research. The purpose was to provide a basis for develop
ment by the Work Group of a guide for flood frequency analysis at locations
where gage records are available which would incorporate the best technical
methods currently known and would yield greater reliability and consistency
than has heretofore been available in flood flow frequency determinations.
The study included: (a) a review of the literature and current
practice to select candidate methods and procedures for testing, (b)
selection of longrecord station data of natural streamflows in the
United States and development of data management and analysis computer
programs for testing alternate procedures3 (c) testing eight basic
statistical methods for frequency analysis including alternate distribu
tions and fitting techniques, (d) testing of alternate criteria for
managing outliers, (e) testing of procedures for treating stations with
zero flow years, (f) testing relationships between annual maximum and
partialduration series, (g) testing of expected probability adjustment,
(h) testing to determine if flood data exh%bft consistent longterm
trends, and (i) recommendations with regard to each procedure tested and
development of background material for the gufdes being developed by the
Work Group.
14l
Data
In all, 300 stations were used in the testing. Flows were essentially
unregulated. Record length exceeded 30 years with most stations having
records longer than 40 years. The stations were selected to give the
best feasible coverage of drainage area size and geographic location and
to include a substantial number of stations with no flow for an entire
year. Table 141 lists the number of stations by size and geographic
zone.
Split Record Testing
A primary concern of the study was selection of a mathematical
function and fitting technique that best estimates flood flow frequencies
from annual peak flow data, Goodness of fit of a function to the data
used in the fitting process is not necessarily a valid criterion for
selecting a method that best estimates flood frequencies, Consequently,
split record testing was used to simulate conditions of actual application
by reserving a portion of a record from the fitting computation and
using it as "future" events that would occur in practice, Goodness of
fit can nevertheless be used, particularly to eliminate methods whose
fit is very poor.
Each record of annual maximum flows was divided into two halves,
using odd sequence numbers for one half and even for the other in order
to elimlnate the effect of any general trend that might possibly exist,
This splitting procedure should adequately simulate practical situations
as annual events were tested and found independent of each other,
Frequency estimates were made from each half of a record and tested
against what actually happened in the other half,
Development of verification criteria is complicated, because what
actually happens in the reserved record half also is subject to sampling
irregularities. Consequently, reserved data cannot be used as a silmple,
accurate target and verification criteria must be probabilistic, The
test procedure, however* simulates condltIons faced by the planner,,
designer, or operator of water resource projects, who knows neither that
past events are representative nor what future events will be.
142
The ultimate objective of any statistical estimation process is not
to estimate the most likely theoretical distribution that generated the
observed data, but rather to best forecast future events for which a
decision is formulated. Use of theoretical distribution functions and
their attendant reliability criteria is ordinarily an intermediate step
to forecasting future events. Accordingly, the split record technique
of testing used in this study should be more rigorous and direct than
alternative theoretical goodnessoffit tests.
Frequency Computation Methods
Basic methods and fitting techniques tested in this study were
selected by the author and the WRC Work Group on Flood Flow Frequency
after careful review of the literature and experience in the various
agencies represented; those that were tested are listed below. Numbering
corresponds to the identification number of the methods in the computer
programs and in the attached tables.
1. LogPearson Type III (LP3). The technique used for this is
that described in (35). The mean, standard deviation, and skew coefficients
for each data set are computed in accordance with the following equations:
(14l)
S2 = C X2 (CX)'/N (14Z)
Nl
!J = N2CX3 3NCXCX' + 2(CX)3
N(N1) (N2)9 (143)
where
X = logarithm of peak flow
N = number of items in the data set
x = mean logarithm
s= standard deviation of logarithms
9 = skew coefficient of logarithms
1143
Flow logarithms are related to these statistics by use of the
following equation:
X =X+kS (144)
Exceedance probabilities for specified values of k and values of k
for specified exceedance probabilities are calculated by use of the
normal distribution routines available in computer libraries and the
approximate transform to Pearson deviates given in reference (31).
2. Log Normal (LN). This method uses a 2parameter function
identical to the logPearson III function except that the skew coefficient
is not computed (a value of zero applies), and values of k are related
to exceedance probabilities by use of the normal distribution transform
available in computer libraries.
3. Gumbel (G). This is the FisherTippett extremevalue function,
which relates magnitude linearly with the log of the log of the recip
rocal of exceedance probability (natural logarithms). Maximum likelihood
estimates of the mode and slope (location and scale parameters) are
made by iteration using procedures described by Harter and Moore in
reference (36). The initial estimates of the location and scale statistics
are obtained as follows:
M= x 0.45005 s (145)
B= .7797 s (146)
Magnitudes are related to these statistics as follows:
X = M + B(ln(1nP)) (147)
where
M = mode (location statistic)
B = slope (scale statistic)
X = magnitude
P = exceedance probability
S = standard deviatlon of flows
144
Some of the computer routines used in this method were furnished by
the Central Technical Unit of the Soil Conservation Service.
4. Log Gumbel (LG). This technique is identical to the Gumbel
technique except that logarithms (base 10) of the flows are used.
5. Twoparameter Gamma (62). This is identical to the 3parameter
Gamma method described below, except that the location parameter is set
to zero. The shape parameter is determined directly by solution of
NHrlund's (37) expansion of the maximum likelihood equation whfch gives
the following as an approximate estimate of ~1:
CL = 1 + /l + $ (lnq i ElnQ) (148)
Aa
4 (In V $.ZlnQ)
where
8= average annual peak flow
N = number of items in the data set
Q = peak flow
Aa = correction factor
f3 is estimated as follows:
(149)
6. Threeparameter Gamma (632. Computation of maximum likelihood
statistics for the 3parameter Gamma distribution is accomplished using
procedures described in reference (38). If the minimum flow is zero, or
if the calculated lower bound is less than zero, the statistics are identical
to those for the 2parameter Gamma distribution. Otherwise, the lower
bound, yp is initialized at a value slfghtly smaller than the .lowest value
of record, and the maximum likelihood value of the lower bound is derived
by iteration using criteria in reference (38). Then the parateters a and S
are solved for directly using the equations above replacing Q with Qy,
Probabilities corresponding to specified magnitudes are computed directly
by use of a library gamma routine. Magnitudes corresponding to specified
145
probabilities are computed by iteration using the inverse solution.
7, Regional LogPearson Type III (LPR). This method is identical
to the logPearson Type III method, except that the skew coefficient is
taken from Figure 14l instead of using the computed skew coefficient.
Regionalized skew coefficients were furnished by the U.S. Geological
Survey.
8. Best Linear Invariant Gumbel (BLI). This method is the same as
for the Gumbel method, except that best linear invariant estimates
(BLIE) are used for the function statistics instead of the maximum
likelihood estimates (MLE). An automatl"c censoring routine is used
for this method only, so there are no altenative outlier techniques
testad for this method. Statistics are computed as follows:
M =. C(X(I)U(N,J,I)) (1410)
B = c(X(I).V(N,J,I)) (14U)
where
U = coefficient UMANN described in reference (39)
V= coefficient BMANN described in qeference (39)
J = number of outliers deleted plus I
I = order number of flows arranged %"n ascendingmagnitude
order
N= sample size as censored,
Since weighting coefficients U and V were made available in this study
only for sample sizes ranging from 10 to 25, &year samples are not
treated by this method, and records (or half records) of more than 25
years are divided into chronological groups and weighted average coeffi
cients used in lieu of coefficients that might otherwise be obtained if
more complete sets of weighting coefficients were avallable. Up to two
outliers are censored at the upper end of the flow array. Each one is
removed If sequential tests show that a value that extreme would occur
by chance less than 1 time IO on the basis of the BLIE statistics.
Details of this censoring technjque are contained in refer
ence (40). Weighting coefficients and most of the routines used in this
method were furnished by the Central Technical Untt of the Soil Conserva
tion Service.
Outliers
Outliers were defined for purpose of this study as extreme values
whose ratio to the next most extreme value in the same (positive or
negative) direction is more extreme than the ratio of the next most
extreme value to the eighth most extreme value.
The techniques tested for handling outliers consisted of
a. keeping the value as is,
b. reducing the value to the product of the second largest event
and the ratio of the second largest to eighth largest event,
C. reducing the value to the product of the second largest event
and the square root of that ratio, and
d. discarding the value.
In the cases of outliers at the low end, the words largest in (b) and
(c) should be changed to smallest.
Zero Flow
Two techniques were tested for handling stations with some complete
years of no flow as follows:
(a) Adding 1 percent of the mean magnitude to all values for
computation purposes and subtracting that amount from subsequent
estimates, and
(b) removing all zeros and multdplying estimated exceedance frequen
cies of the remaining by the ratlo of the number of nonzero values to
the total number of values, This is the procedure of combining probabil
Ities described in reference (27).
PartialDuration Series
A secondary concern of the study was the relationship between
annual maximum flow frequencies and partfalduration flow frequencies,
Because a partialduration series consists of all events above a
specified magnitude# it is necessary to define separate events. The
definition normally depends on the application of the frequency study as
147
well as the hydrologic characteristics of the stream. For this study
separate events were arbitrarily defined as events separated by at least
as many days as five plus the natural logarithm of the square miles of
drainage area, with the requirement that intermediate flows must drop
below 75 percent of the lower of the two separate maximum daily flows.
This is considered representative of separation criteria appropriate for
many applications.
Maximum daily flows were used for this part of the study, because
there were insufficient readily available data on instantaneous peak
flows for events smaller than the annual maximum. There is no reason to
believe that the frequency relationship would be different for peak
flows than for daily flows.
The relationship between the maximum annual and partialduration
series was expressed as a ratio of partialduration to annual event
frequencies at selected annual event frequencies. In order to develop
partialduration relationships independent of any assumptions as to
frequency functions, magnitudes corresponding to annualmaximum event
exceedance probabilities of 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, and 0.7 are
established for complete records at each station by linear interpolation
between expected probability plotting positions (M/(n+l)) for the annual
maximum events. Corresponding frequencies of partialduration flows are
established simply by counting the total number of independent maximum
daily flows at each station above each magnitude and dividing by the
total number of years at that station. Ratios of partialduration to
annual event frequencies were averaged for all stations in each USGS
zone and compared with ratios derived for certain theoretical conditions
by Langbein (9).
Expected Probability Estimation
The expected probability is defined as the average of the true
probabilities of all magnitude estimates for any specified flood frequency
that might be made from successive samples of a specified site. For any
specified flow magnitude, it is considered to be the most appropriate
estimate of probability or frequency of future flows for water resources .
planning and management use.
It Is also a probability estimate that is theoretically easy to
148
verify, because the observed frequencies in reserved data at a large
number of stations should approach the computed probability or frequency
estimates as the number of stations increases. Accordingly, it was
considered that expected probability estimates should be used in the
split record tests.
A method of computing expected probabilities has been developed for
samples drawn from a Gaussian normal distribution as described in (21).
Similar techniques are not available for the other threoretical
distribution functions. Consequently, an empirical transform is derived
for each distribution. To do this a calibration constant was determined
which, when multiplied by the theoretical normal transform adjustment,
removed the observed average bias in estimating probabilities for the
300 stations used in this study. This empirical transform was used in
making the accuracy tests that are the main basis for judging the relative
adequacy of the various methods tests.
Trends and Cycles
There is some question as to whether longterm trends and cycles
(longer than 1 year) exist in nature such that knowledge of their
nature can be used to improve forecasts of flood flow frequencies for
specific times in the future. As a part of this research project, lag
1 autocorrelation coefficients of annual peak flows for all stations
were computed. If trends or cycles exist in any substantial part of the
data, there should be a net positive average autocorrelation for all
stations. A statistically significant positive average autocorrelation
was not found.
Accuracy and Consistency Tests
Criteria used in judging the adequacy of each method for fitting a
theoretical distribution were as follows:
Accuracy tests consisted of the following comparisons between
computed frequencies dn onehalf the record with frequencies of events
that occurred In the reserved data.
a. Standard deviation of observed frequencies (by count) fn
reserved data for magnitude estimates corresponding to exceedance
probabilities of 0.001, 0.01, 0.1, and 0.5 computed from the part of the
record used. This is the standard error of a frequency estimate at
individual stations that would occur if a correction is made for the
average observed bias in each group of stations for each selected frequency
and method.
b. Rootmeansquare difference between expected probability
plotting position (M/(n+l)) of the largest, upper decile and median
event in a half record and the computed expected probability exceedance
frequency of that respective event in the other half. This is the
standard error of a frequency estimate at individual stations without
any bias adjustment for each method and for the frequency of each selected
event.
c. Rootmeansquare difference between 1.0 and the ratio of the
computed probability of flow in the opposite half of a record to the
plotting position of the largest, upper decile and median event (in
turn) in a half record. This criterion is similar to that of the preceding
paragraph except that methods that are biased toward predicting small
frequencies are not favored.
Consistency tests involved the following comparisons between
computed frequencies in each half of the record with the total record.
a. Rootmeansquare difference between computed probabilities from
the two record halves for full record extreme, largest, upper decile and
median events, in turn. This is an indicator of the relative unliformity
of estimates that would be made with various random samples for the same
location.
b. Rootmeansquare value of 1.0 minus the ratio of the smaller to
the larger computed probabilities from the two record halves for full
record extreme, largest, upper decile and median events, in turn. This
is essentially the same as the preceding criterion, except that methods
that are biased toward predicting small frequencies are not favored.
The extreme event used in the consistency tests is an arbitrary
value equal to the largest multiplied by the square root of the ratio of
the largest to the medfan event for the full record,
It should be recognized that sampling errors in the reserved data
are as large or larger for the same sample size as are sampling errors
1410
of computed values. Similarly, sampling errors are comparable for
estimates based on opposite record halves used for consistency tests.
Consequently, a great number of tests is necessary in order to reduce the
uncertainty due to sampling errors in the reserved data, Further, a
method that is biased toward estimating frequencies too low may have a
small standard error of estimating frequencfes in comparison with a
method that is biased toward high frequencies, if the bias is not removed.
The latter may have smaller percentage errors. Accordingly, consl"der
ation of the average frequency estimate for each of the eight methods
must be a component of the analyses. :
As a further means of evaluating alternate procedures the complete
record results, computed curve without any expected probability adjustment,
and the plotted data point were printed out,
Evaluation of Distributions
Table 142 shows for each method and each USGS zone the number of
stations where an observed discharge exeeeded the computed l,OOOyear
discharge. With 14,200 stationyears of record, it might be expected
that about 14 observed events would exceed true 1,000"year magnitudes,
This comparison indicates that the logPearson Type III (method l), log
normal (method 2), and logPearson Type III with generalized skew (method
7), are the most accurate,
Table 143 shows average observed frequencies (by count) in the
reserved portions of half records for computed probabilities of 0.001,
0.01, 0.1, and 0,5 and the standard deviations (accuracy test a) of the
observed frequencies from their averages for each computed frequency.
It is difficult to draw conclusions from these data. Figure 142 shows
a plotting of the results for the 0.01 probabiljty estimates which aids
in comparison. This comparison indicates that the log normal and log
Pearson Type III methods with generalized skew have observed frequencies
closest to those computed and the smallest standard deviations except
for method 4,
Table 144 shows the average results for all stations of accuracy
tests b and c. Results are not definitive, but again the log normal
(method 2) and logPearson Type III with generalized skew (method 7)
show results as favorable as any other method as illustrated for test b
in figure 143.
Table 145 shows the results of the consistency tests. Figure 144
displays the results graphically for test a. The consistency test results
are not substantially different from or more definitive than the accu
racy results. From Figure 144 it appears that the logPearson Type III
method with generalized skew yields considerably more consistent results
than the log normal.
Results of Outlier Testing
Table 146 shows results for all stations of the accuracy and
consistency tests for the four different outlier techniques. Results of
these tests show that for the.favorable methods [log normal (method 2)
and logPearson Type III with generalized skew (method 7)1, outlier
techniques a and b are most favorable. Unfortunately, no discrimination
was made in the verification tests between treatment of outliers at the
upper and lower ends of the frequency arrays. Outliers at the lower end
can greatly increase computed frequencies at the upperend. Average
computed frequencies for all half records having outliers at the upper
or lower end are generally high for the first three outlier techniques
and low for the fourth.
It is considered that this is caused primarily by outliers at the
lower end. Values observed are as follows:
Average plotting position of maximum flow 0.042
Average computed probability, method a 0.059
Average computed probability, method b 0,050
Average computed probability, method c 0.045
Average computed probability, method d 0.038
Until more discriminatory outlier studies are made, method a
appears to be the most logical and justifiable to use.
Results of Zero Flow Testings
Table 147 shows the average for all stations of the results of
accuracy and consistency tests for the two different zero flow techniques.
1412
These test comparisons indicate that for the favorable methods [log
normal (method 2) and logPearson Type III with generalized skew (method
7)1, technique b is slightly better than a.
Results of PartialDuration Studies
Results of partialduration studies are shown in Table 148. 'It
can be seen that there is some variation in values obtained for different
zones and that the average of all zones is somewhat greater than the
theoretical values developed by Langbein. The theoretical values were
based on the assumption that a large number of independent (random)
events occur each year. If the number of events per year is small, the
average values in Table 148 would be expected to be smaller than the
theoretical values. If the events are not independent such that large
events tend to cluster in some years and small events tend to cluster in
other years, the average values in Table 148 would be expected to be
larger than the theoretical values.
It was concluded that values computed for any given region (not
necessarily zones as used in this study) should be used for stations in
that region after smoothing the values such that they have a constant
relation to the Langbein theoretical function,
Expected Probability Adjustment Results
The ratios by which the normal expected probability theoretical
adjustment must be multiplied in order to compute average probabilities
equal to those observed for each zone are shown in Tables 149, 1410,
and 1411. It will be noted that these vary considerably from zone to
zone and for different exceedance intervals. Much of this variation,
however, is believed due to vagaries of sampling. Average ratios for
the looyear flood shown on the last line in Table 1410 were adopted
for each distribution for the purpose of comparing accuracy and the
various methods. These are as follows:
1. LogPearson Type III 2.1
2. Log Normal 0.9
3. Gumbel, MLE 3.4
4, Log Gumbel 1.2
5. 2parameter gamma 3.4
1413
6. 3parameter gamma 2.3
7. Regional logPearson Type III 1.1
8. Gumbel, BLIE 5.7
Results of this portion of the study indicate that only the log
normal (method 2) and logPearson Type III with regional skew (method 7)
are free of substantial bias because zero bias should correspond approxi
mately to a coefficient of 1.0 as would be the case l"f the distribution
characteristics do not greatly influence the adjustment factor. The
following tabulation for logPearson Type III method wSth regional skew
indicates that the theoretical expected probability adjustment for the
normal distribution applies approximately for this method, Coefficients
shown range around the theoretical value of 1.0 and, with only one
exception, do not greatly depart from it in terms of standarderror
multiples. It is particularly significant that the most reliable data
(the looyear values) indicate an adjustment factor near 1.0.
Expected Probability Adjustment Ratios for All Zones
Sample loYr lOOYr lOOOYr
Size Avg. Std. Err. Avg. Std. Err. Avg. Std. Err.
5 0.81 0*17 0.94 0.12 1.01 0.13
10 0.60 OS22 1,12 0.20 1.45 0.27
23 0.17 0.27 1.14 0.23 1.68 0.28
Results of Test for Trends and Cycles
Results of lag I autocorrelation studies to test for trends are
shown in Table 1412. It is apparent that there is a tendency toward
positive autocorrelation, indicating a tendency for flood years to
cluster more than would occur in a completely random process. The
t values shown are multiples of the standard error of the lag I correla
tion coefficient, and it is obvious that extreme correlation coefficients
observed are not seriously different from variations that would occur by
chance. It is considered that annual peak flows approximate a random
process in streams used in this study.
1414
Conclusions
Although split record results were not as definitive as anticipated,
there are sufficient clearcut results to support definite recommendations,
Conclusions that can be drawn are as follows:
a. Only method 2 (log normal) and method 7 (logPearson Type III
with regional skew) are not greatly biased in estimating future frequencies.
b. Method 7 gives somewhat more consistent results than method 2.
C. For methods 2 and 7, outlier technique "a" (retaining the
outlier as recorded) is more accurate in terms of ratio of computed to
observed frequencies than methods that give less weight to outliers.
d. For methods 2 and 7, zero flow technique "b" (discarding zero
flows and adjusting computed frequencies) is slightly superior to zero
flow technique "a."
e. Streamflows as represented by the 300 stations selected for
this study are not substantially autocorrelated; thus, records need not
be continuous for use in frequency analysis.
f. Partialduration frequencies are related to annual event
frequencies differently in different regions; thus, empirical regional
relationships should be used rather than a single theoretical relationship.
Of particular significance is the conclusion that frequencies
computed from theoretical functions in the classical manner must be
adjusted to reflect more frequent extreme events if frequencies computed
in a great number of cases are to average the same as observed frequencies.
For the recommended method, adjustment equal to the theoretical adjustment
for estimates made from samples drawn from a normal population is approxi
mately correct.
Of interest from a research standpoint is the finding that split
record techniques require more than 300 records of about 50 events each
to be definitive. This study showed that random variations in the
reserved data obscure the results to greater degree than would be the
case if curvefitting functions could reduce uncertainty to a greater
degree than has been possible,
In essence, then, regardless of the methodology employed, substan
tial uncertainty in frequency estimates from station data will exist,
but the logPearson type III method with regional skew coefficients will
produce unbiased estimates when the adjustment to expected probability
is employed, and will reduce uncertainty as much as or more than other
methods tested.
Recommendations for Future Study
It is considered that this study is an initial phase of a more
comprehensive study that should include
a, Differentiation in the treatment of outliers at the upper and
lower ends of a frequency curve;
b. Treatment of sequences composed of different types of events
such as flood flows resulting from rainfall and those from snowmelt, or
hurricane and nonhurricane floods;
c. Physical explanation for great differences in frequency character
istics among streams in a given region;
d. Development of systematic procedures for regional coordination
of flood flow frequency estimates and applications to locations with
recorded data as well as to locations without recorded data;
e. Development of procedures for deriving frequency curves for
modified basin conditions, such as by urbanization;
f. Development of a stepbystep procedure for deriving frequency
curves for locations with various amounts and types of data such that
progressively reliable results can be obtained on a consistent basis as
the amount of effort expended is increased; and
clPreparation of a text on flood flow frequency determinations
for use in training and practical application,
FIGURE 14l
GENERALIZED SKEW COEFFICIENTS OF ANNUAL MAXIMUM
STREAMFLOW LOGARITHMS
14l 7
FIGURE 142
ACCURACY COMPARISON FOR 0.01 PROBABILITY ESTIMATE (TABLE 143)

8L
i
/ I
POSSIBLE COMPARISObj LINE
I
I
METHOD NUMBER
.06
.05
.04
.03
3
, %
.02
.Ol
0.00 L
0.00 .Ol .02 .03 .04 .05 .06 .07 .08
AVERAGE OBSERVED PROBABILITY IN
FOR 0.01 COMPUTED PROBABILITY
TABLE 143
1418
FIGURE 143
ACCURACY COMPARISON FOR MAXIMUM OBSERVED FLOW
(TABLE 144, TEST 8)
.09
.08
z
w .05
ot
Y
LL
Y
0
0.00 I
0.00 .Ol .02 .03 .oa .05 .06 .07 .08
AVERAGE OBSERVED PROBABILITY IN TABLE 143
FOR 0.0 1 COMPUTED PROBABILITY
14l 9
FIGURE 144
CONSISTENCY COMPARISON FOR MAXIMUM OBSERVED FLOW
(TABLE 145, TEST A)
.2
?hI
0.00 .Ol
AVERAGE
.02 .03
OBSERVED
FOR 0.01
.04
PROBABILITY
COMPUTED
.05
IN
PROBABILITY
.06
TABLE
i
.07
143
.08
1420
USGS
ZONE
1
2
3
4
5
6
7
a
9
10
11
12
13
14
15
16
*
Total
Table 141
Numbers of Verification Stations by Zones and Area Size
Drainage area category (sq. ml.) Total
o25 25200 2001000 1 ooo+
4 8 10 5 27
2 5 12 5 24
5 3 16 1 25
1 6 a 0 15
3 2 14 1 20
4 3 13 4 24
5 2 12 2 21
8 2 11 2 23
1 7 a 2 18
0 a 4 0 12
2 5 6 0 13
0 5 9 3 17
0 2 10 5 17
0 6 a 1 15
2 1 0 Q 3
12 1 0 0 13
4 7 1 1 13
53 73 142 32 300
*Zeroflow stations (zones a, 10 & 11 only)
J421
Table 142
NUMBER OF STATIONS WHERE ONE OR MORE OBSERVED FLOOD EVENTS
EXCEEDS THE lOOOYR FLOW COMPUTED FROM COMPLETE RECORD
STATION
YEARS OF
ZONE RECORD 1
i1414 0
2 1074 0
3 1223 1
4 703 1
5 990 2
6 1124 0
7 852 1
8 969 1
9 920 3
10 636 1
11 594 1
12 777 0
13 911 1
14 761 0
15 120 0
16 637 1
* 495 1
TOTAL 14,200 14
Based on the 14,200 stationyears
about 14 observed events would
*Zeroflow stations
METHOD
2 3 4 5. 6 L s
1 8 0 10 T 2 26
3 9 0 10 7 1 19
3 7 0 9 8 4 22
2 3 0 3 3 2 1%
1 7 0 4 4 0 19
2 4 0 4 4 1 18
2 5 1 3 4 3 '87
1 10 0 3 3 1 19
0 4 0 3 3 1 16
0 2 D 1 1 0 10
1 6 0 4 4 0 11
2 2 0 2 2 2 9
0 1 0 4 2 2 14
0 3 0 4 1 1 15
0 D 0 0 0 0 2
0 4 0 4 3 0 12
0 2 0 0 0 0 12
18 77 1 68 56 20 253
of record, it might be expected that
exceed the true lOOOyear magnitudes.
1422
Table 143
STANDARD DEVIATION COMPARISONS
AVERAGE FOR ZONES 1 TO 16
COMPUTED METHOD
PROBABILITY 1 2 3 4 5 6 7 8
AVERAGE OBSERVED PROBABILITIES
.OOl .0105 .0041 .0109 .OOOl .OllO .0092 .0045 .0009
.Ol 0232 0153 .0315 .0023 .0309 .0244 .0170 .0015
.l .1088 .1007 .1219 .0707 1152 .1047 .1020 .0029
.5 .5090 .5149 .4576 .6152 .4713 .4950 .5108 .0037
STANDARD DEVIATION OF OBSERVED PROBABILITIES FOR SPECIFIED COMPUTED PROBABILITIES
001 .0290 0134 .0244 .0025 .0239 .0218 .0150 .0222
.Ol 0430 .029 ,045 .OlO ,043 .039 032 .035
.1 .086 .084 .089 .074 ,089 .084 .084 .067
.5 132 131 .142 .133 ,133 .141 .130 .123
Note: Averages and standard deviations are of observed frequencies in the reserved portion of each
record corresponding to computed mangitudes based on half records, Low standard deviations in re
lation to averages indicate more reliable estimates.
Table 144
Evaluation of Alternative Methods
Accuracy Tests b and cb Average Values, All Stations
Test bRoot mean square difference between plotting position and
computed probability in other half of record.
Method
1. 2 3 4 5 !i L 8
Maximum .062 ,060 ,067 ,056 ,070 ,069 .061 ,061
Decile .084 .080 .097 ,063 e 098 * 094 .081 ,082
Median .254 .105 ,657 193 ,518 .295 ,120 ,727
Test cRoot mean square difference bewteen 1.0 and ratio of
computed probability of flow in opposite half of record
to plotting position, A zero value would indicate a
perfect forecast.
Method
1 2, 2 Q ii 5. L !i
Maxtmum .53 .51 .56 .45 .56 .56 .51 .59
Decile *37 .34 .38 a27 .37 937 ,34 .40
Median .40 .12 065 .19 .59 .44 .14 .52
Table 145
Evaluation of Alternative Methods
Consistency Tests a and b, Average Values, All Stations
Test aRoot mean square difference between computed probabilities from
the two record halves for full record extreme, largest, upper
decile and median events. A zero value would indicate perfect
consistency.
Method
Event 1 2 3 a 5 s L s
Extreme .003 .006 .OOl ,010 e.001 ,002 .003 002
Maximum ,023 ,019 ,008 ,016 ,008 .OlO .OlO .012
Upper Decile ,072 ,047 ,043 .025 0037 .033 .025 ,048
Median ,119 ,076 ,072 .047 e 049 045 0041 .131
Test bRoot mean square value of (1.0 minus the ratio of the smaller
to the larger computed probabilities from the two record halves)
for full record extremeb largest, upper decile and median
events. A zero value would indicate perfect consistency.
Method
Event 1 2 a I 5 !i a 6
Extreme a87 .54 ,46 026 039 .35 e29 975
Maximum .74 *45 .$I 421 .34 ,30 ,24 .72
Upper Decile .50 .32 .31 .16 .24 .21 .17 .58
Median .21 .14 .12 010 ,08 .08 .oa ,24
Table 146
Evaluation of Outlier Techniques
Average Values, All Stations
Method
Accuracy Test b
Outlier
Technique 1 2 3 9 2 a 1.
a ,061 .062 ,071 ,057 .074 .073 ,062
b .056 ,055 .060 ,053 .063 .062 .055
C
,052 ,050 .054 ,048 ,057 ,055 .051
d ,047 .045 ,048 ,044 ,051 .050 .045
Accuracy Test c
Outlier
Technique 1 2 2 !?. 5. 6 2"
a .53 .55 .57 .47 .58 .58 .54
b .57 .5g .59 .49 .62 60 .58
C .58 .61 .60 452 .64 .63 .60
d .65 .65 .64 .38 .68 .65 .64
Consistency Test a
Outlier
Technique 1 2 2 4 2. a 2
a .002 .005 ,001 .009 . 000 .002 .002
b ,002 ,004 ,001 008 . 000 .002 ,002
C .003 .003 .ooo ,007 * 000 002 .002
d ,003 .003 .ooo .007 D 000 002 .OOl
Consistency Test b
Outlier
Techniques 1 2 P !I 5 f! I
a .87 056 .46 027 .39 .36 .30
b .86 $56 .45 .28 .38 .35 .30
C .85 056 .45 .29 .38 .35 .30
d .88 .59 .45 .31 .38 .35 .32
A zero value would indicate perfect consistency.
Method 8 includes its unique technique for outliers and was, therefore,
not included in these tests,
1426
Table 147
Evaluation of Zero Flow Techniques
Average Values, All Stations
Accuracy Test b
Method
Technique 1 2 s. 9 5 6 z
a ,057 .057 *OS9 .057 .062 ,055 *OS9
b .064 .060 .070 .057 ,068 .061 .061
Accuracy Test c
Method
Technique 1 2 9 4 5 6 2.
a 646 .32 .59 .32 40 .340 .32
b .51 .30 .59 .30 .40 .4f .31
Consistency Test a
Method
Technique L 2 z 9 5 6 7
a .007 .012 .ooo .014 001 .ooo .006
b .007 .008 e 000 .012 .ooo .OOl .004
Consistency Test b
Method
Technique 1 2 3 4 5 6 1
a .89 .e3 44 .21 .39 .34 .24
b .86 .43 .44 .19 .40 .38 .23
Method 8 was not tested because logarithms are not used in its
fitting computations and therefore zero flows are not a problem.
1427
Table 148
Summary of PartialDuration Ratios
Partialduration frequencies
for annualevent frequencies of
Zone 1L 2L 2m.w..3 A 4 A 5 .6 .7
1 (21 sta) .094 .203 328 .475 ,641 .844 1.10
2 (17 sta) ,093 .209 .353 .517 c 759 1.001 1.30
3 (19 sta) .094 .206 ,368 .507 .664 .862 1.18
4 (8 sta) ,095 .218 .341 .535 .702 D 903 1.21
5 (17 sta) .093 .213 .355 ,510 .702 ,928 1.34
6 (16 sta) .134 .267 .393 ,575 .774 1.008 1.33
7 (9 sta) .099 .248 ,412 ,598 "826 1.077 1.42
8 (12 sta) ,082 ,211 .343 .525 .803 1.083 1.52
9 (15 sta) ,106 .234 .385 .553 .765 o 982 1.26
10 (12 sta) .108 .248 .410 ,588 .776 1.022 1.34
11 (12 sta) .094 .230 .389 ,577 .836 1.138 1.50
12 (12 sta) .103 .228 352 .500 ,710 .943 1.21
13 (16 sta) ,095 .224 0372 .562 ,768 0 986 1.30
14 (14 sta) .lOO .226 ,371 .532 .709 * 929 1.22
15 (3 sta) ,099 .194 .301 .410 l 609 ,845 1.05
'16 (13 sta) .106 ,232 .355 .522 .696 ,912 1.27
Average ,099 .243 ,366 .532 .733 e 964 1.28
Langbein ,105 .223 356 .510 ,693 .917 11.20
Note: Data limited to 226 stations originally selected for the study.
1428
TABLE 149
ADJUSTMEMT RATIOS FOR loYEAR FLOOD
SAMPLE
SIZE ZONE 1 27 STATIONS AVG l/2 RECORD = 26 YRS
METHOD 1 2 3 4 5 6 7 8
5YR .54 .38 076 .29 .a2 .57 .28 1 .a5
loYR .75 l 45 1.02 .27 .95 .37 .34 4.56
l/2REC 1.21 1.11 2.21 1.04 2.01 1.01 1.03 4.09
ZONE 2 24 STATIONS AVG l/2 RECORD= 22 YRS
METHOD 1 2 3 4 5 6 7 a
5YR .4a .42 1.06 .64 1.03 093 .41 1.85
loYR 1.01 .94 1.91 .68 1.60 1.3i .a0 5.70
l/PREC 1.33 1.33 2.76 1.58 1.90 .49 .54 7.14
ZONE 3 25 STATIONS AVG l/2 RECORD= 24 YRS
METHOD 1 2 3 4 5 6 7 a
5YR 1.41 1.32 1.92 1.02 1.95 1.79 1.4D 1.85
loYR 1.41 .a1 1.80 .oo 1.87 096 1.01 5.39
l/2REC 98 .14 1.65 1.88 1.17 .21 .39 4,ao
ZONE 4 15 STATIONS AVG l/2 RECORD= 23 YRS
METHOD 1 2 3 4 5 6 7 a
5YR 1.05 .94 1.20 .a5 1.29 1.15 .94 1.85
IOYR .52 .50 .12 .a5 .Ol .54 .45 3.68
l/EREC .45 .02 1.63 3.07 1.63 .46 .25 5.57
ZONE 5 20 STATIONS AVG l/2 RECORD= 25 YRS
METHOD 1 2 3 4 5 6 7 a
5YR .55 .35 1.03 .15 .98 .a8 .47 1.85
10"YR .40 .03 1.40 .96 .61 .42 .19 7*37
l/2REC .a1 ,40 2.91 3.61 1.42 699 .67 6023
ZONE 6 24 STATIONS AVG l/2 RECORD= 23 YRS
METHOD 1 2 3 4 5 6 7 a
5YR .80 .36 1.19 .15 1.11 .95 .45 9 .a5
loYR 1.43 .la 2.26 .98 1.78 .96 .33 5.64
l/PREC 1.08 .45 .2.94 3.93 1.94 07 .04 6.14
ZONE 7 21 STATIONS AVG l/2 RECORD = 20 YRS
METHOD 1 2 3 4 5 6 7 a
5YR 1.15 1.19 1.69 1.29 1.62 1.59 1.29 1.85
loYR 1.58 1.36 2.34 .12 1.99 1.62 1.57 5.78
l/2REC 1.97 1.00 2.45 .74 2.87 .92 1.17 7.11
ZONE 8 23 STATIONS AVG l/2 RECORD = 21 YRS
METHOD 1 2 3 4 5 6 7 a
5YR .a9 .a9 9.71 .79 1.41 1.36 .79 1.85
IOYR .66 1.02 .29 2.04 .35 .43 1.02 4.52
l/LREC .13 .a7 2.28 3.08 .74 .66 .a7 7.88
1429
TABLE 149 CONTINUED
ZONE 9 18 STATIONS AVG l/2 RECORD = 25 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.38 1.02 2.05 .96 1.96 1.78 1.10 1.85
loYR 1.95 1.54 2.54 .75 2.49 2.22 1.69 6.76
l/LREC .45 .36 .97 3.36 .45 .07 027 4.07
ZONE 10 12 STATIONS AVG l/2 RECORD = 26 YRS
METHOD 1 2 3 4 5 6 7 8
5YR .79 .80 .41 .83 .43 .43 .77 1.85
IOYR .03 .42 .90 1.16 .71 .35 .22 4.24
l/LREC .08 1.27 1.24 5.10 .5B .27 1.27 2.97
ZONE 11 13 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.29 1.21 1.89 1.20 1.93 1.75 1.11 1.85
loYR 1.11 1.03 2.21 .04 1.87 1.25 1,03 6.78
l/2REC .04 ,23 1.99 2.93 1.20 1.20 .23 5.32
ZONE 12 17 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.34 .73 1.34 .57 1.51 1.03 .80 1.85
loYR .79 .41 .86 .45 .92 .44 .57 4.06
l/PREC .19 .31 .54 2.94 .92 .35 .19 2.81
ZONE 13 17 STATIONS AVG l/2 RECORD = 26 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.27 1.16 1.65 .96 1.77 1.52 1.19 1.85
IOYR .26 .22 .88 .83 .67 .42 038 4.60
l/2REC .31 1.52 .21 4.89 .I7 .97 1.12 2,88
ZONE 14 15 STATIONS AVG 1;/2 RECORD = 25 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.72 1.65 2.12 1.61 2.19 2,oo 1.65 1.85
loYR 2.60 2.50 3.17 1.88 2.82 1.87 2.56 6.80
l/2REC 51 .61 1.83 1.47 1.30 .29 075 5.22
ZONE 15 3 STATIONS AVG l/2 RECORD = 20 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 2.47 2.47 2.74 2.55 2.66 2.28 2.28 1.85
IOYR 1.27 1*27 1.58 1.27 1.58 1058 1,27 2,65
l/IREC 3.29 3.29 3.29 2.79 3.29 1.90 3.29 6.33
ZONE 16 13 STATIONS AVG l/2 RECORD = 24 YRS
METHOD 1 2 3 4 5 6 a 8
5YR 069 .76 1.03 .66 1.09 1.05 .75 1.85
loYR .58 42 .83 .21 .76 .07 .42 4.24
l/PREC 1.41 07 1.68 3.43 1.25 .64 .07 5.29
ALL ZONES 287 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR .94 .79 1.38 .71 1.37 1.21 .81 1.85
IOYR .87 .52 1.52 .29 1.26 .72 .60 5.27
l/2REC .77 .04 1.93 2.66 1.34 .40 .17 5.36
Values shown are ratios by which the theoretical adJustment for Gausslan
distribution samples must be multiplied In order to convert from the com
puted 0.1 probability to average observed probabilities in the reserved
data. See note table 1411.
u30
TABLE 1410
ADJUSTMENT RATIOS FOR lDOYEAR FLOOD
SAMPLE
SIZE ZONE 1 27 STATIONS AVG l/2 RECORD= 26 YRS
METHOD 1 2 3 4 5 6 7 B
5YR 1.35 1.11 1.27 .39 1.61 1.12 .BB .25
10.YR 1.50 1.10 2.05 a.25 2.42 1.73 .73 3.42
l/2REC 2.83 2.84 3.90 1.06 4.89 3.67 1.66 5.28
ZONE 2 24 STATIONS AVG l/2 RECORD = 22 YRS
METHOD 1 2 3 4 5 6 7 8
5YR .91 ,79 1.05 .31 1.27 1.13 .63 .25
loYR 1.44 1.40 2.48 .63 2.41 2.07 1.37 5.40
l/PREC 1.00 1.08 3.69 ,B2 2e97 2.46 ,14 7.16
ZONE 3 25 STATIONS AVG l/2 RECORD= 24 YRS
METHOD 1 2 3 4 5 6 7 B
SYR 1.80 1.18 1.76 .41 2.05 1.86 1.29 0.25
loYR 2.42 1.15 2.43 .04 2.84 1.62 1.32 4.79
l/EREC 2.90 1.41 3.36 1.12 3.71 2.76 2.30 5.53
ZONE 4 15 STATIONS AVG l/2 RECORD * 23 YRS
METHOD 1 2 3 4 5 6 7 B
5YR 1.67 1.48 1.45 .59 2.27 2.02 1.64 .25
loYR ,57 .35 .56 .4B 1.07 .46 .42 , 1.60
l/PREC 1.86 .4B 1.54 1.15 2.83 .BB 1.03 3.81
ZONE 5 20 STATIONS AVG l/2 RECORD  25 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.03 .64 1.37 e24 1.19 1.12 .B2 025
loYR 1.22 .57 1.42 .29 lo27 1.09 .80 5,65
l/LREC 2.97 .21 4.38 1.24 2.97 2.39 1.68 7.25
ZONE 6 24 STATIONS AVG l/2 RECORD  23 YRS
METHOD 1 2 3 4 5 6 7 B
5YR 1.15 ,67 1.02 *04 1.17 .8B .76 .25
loYR 2.30 ,55 1.67 ,27 1.78 1.10 .66 4.43
l/LREC 1.20 ,23 3.22 1.24 2.45 .79 046 5,09
ZONE 7 21 STATIONS AVG l/2 RECORD= 20 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.04 1.07 2.23 .28 2020 2016 1620 025
10.YR 1.18 1.09 2.66 0.19 2.54 2.20 1.53 5.40
l/2REC 3.10 .47 3.92 1.80 2,99 2.29 1074 8.33
ZONE 8 23 STATIONS AVG 112 RECORD  21 YRS
METHOD 1 2 3 4 5 6 7 B
5YR .57 ,27 2.08 eo1 1.66 1,52 a27 .25
101YR 1.30 a14 1.59 .35 l"15 .93 .I4 4.17
l/2REC ,82 a32 4.36 1,13 2.16 2.16 .32 8.49
1431
TABLE 1410 CONTINUED
ZONE 9 18 STATIONS AVG l/2 RECORD = 25 YRS
METHOO 1 2 3 4 5 6 7 8
5YR 1.07 1.33 1.90 .72 2.11 2.11 1.50 .25
loYR 2.45 2.23 3.21 .90 3.75 3.55 2.57 4.39
l/2REC 1.07 .39 2.90 1.72 3.78 2.38 .66 4.49
ZONE 10 12 STATIONS AVG l/2 RECORD = 26 YRS
METHOO 1 2 3 4 5 6 7 8
5YR .lO .lO .27 .25 .29 .29 .06 .25
loYR .21 .15 .96 .59 1.06 .75 .15 2.55
l/2REC 3.29 .27 1.63 1.79 2.42 1.32 .27 4.40
ZONE 11 13 STATIONS AVG l/2 RECORD = 23 YRS
METHOO 1 2 3 4 5 6 7 8
5YR .68 .7p 1.79 .ll 1.58 1.54 .66 025
IOYR 2,41 T.51 4.14 .17 3.76 3.43 1.28 6.64
l/PREC .30 .79 5040 1.08 3.05 2.43 050 9,77
ZONE 12 17 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.81 1010 1.16 .44 1.56 1.19 1.19 625
loYR 1.99 1.93 1055 613 2.27 l,D4 2.11 2060
l/2REC 3.77 1.65 2.12 1.33 4.39 2.57 1.86 1.82
ZONE 13 17 STATIONS AVG l/2 RECORD = 26 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.63 .87 1.12 .50 1.63 1.26 1.04 .25
loYR .58 .37 1.27 .28 1.41 1.25 .60 3.28
l/PREC 1.01 .07 2.20 1.81 2.57 1.61 .81 2.69
ZONE 14 15 STATIONS AVG l/2 RECORD = 25 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.54 1.44 1.79 065 2.43 2.21 1.44 .25
loYR 2.92 2.22 2.58 .23 3.53 1.98 2.32 5.16
l/PREC 2.11 2.80 3.76 1.52 4.40 3.10 2.80 5.37
ZONE 15 3 STATIONS AVG l/2 RECORD = 20 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 2*09 2.24 2.24 1.24 2.76 1.98 1.50 .25
IOYR .26 .26 .26 069 4.84 1.84 .26 1.72
l/PREC 1.80 1.80 .93 1.31 4.37 3,16 .93 .93
ZONE 16 13 STATIONS AVG l/2 RECORD  24 YRS
METHOD 1 *2 3 4 5 6 7 8
5YR .61 .55 $90 .18 1.30 1.22 .62 .25
IOYR 1.87 1.23 1.63 .59 1.83 .99 1.33 3.64
l/LREC 4.21 1.17 3.96 1.27 4.41 2.90 2.13 4.46
ALL ZONES 287 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.16 .90 1.45 032 1.66 1.45 .94 0.25
IOYR 1.64 1.03 2.01 007 2.20 1.62 1.12 4.25
l/PREC 2.12 .87 3.40 1.23 3.35 2.30 1.14 5.66
Values shown are ratlos by which the theoretical adjustment for Gaussian
dlstrlbutlon samples must be multlplled In order to convert from the com
puted 0.01 probablltty to average observed probabll!tles In the reserved
data. See note table 1411.
1432
TABLE 1411
ADJUSTMENT RATIOS FOR lOOOYEAR FLOOD
SAMPLE
SIZE ZONE 1 27 STATIONS AK l/2 RECORD s 26 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 2.03 1.10 1.19 .2l %a12 1.44 .85 .04
loYR 2.30 88 2.21 .14 2.98 1.87 .52 4.06
l/PREC 5.01 4.13 6.94 .56 lo,11 8.16 1.66 8.54
ZONE 2 24 STATIONS AVG l/2 RECORD = 22 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.31 .83 1.18 .15 1.57 1.35 .68 .04
10"YR 1.98 2.85 3.85 .64 4.45 3.66 2.07 7..41
l/PREC la93 2.11 4.47 .45 3.56 3.56 l58 8.81
ZONE 3 25 STATIONS AVG l/2 RECORD s 24 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 2.42 1.22 2.18 .Ol 2.54 2.08 1.24 004
IOYR 6.06 2.20 3.06 .14 3.89 1.82 2.20 7.11
l/GREG 7.41 2.44 6.77 .51 7.06 4.82 2,77 11.16
ZONE 4 15 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 6 6 7 8
5YR 1.88 1050 1.46 .30 2.48 2.05 1.63 ,04
loYR 1.24 .54 947 .14 1.13 .36 .7P 1.33
l/LREC 2.86 .80 2.11 .48 3.60 3.60 2.40 2.81
ZONE 5 20 STATIONS AVG l/2 RECORD = 25 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.84 .94 1.36 .49 l92 1,45 1,32 ,04
IOYR 2,75 656 2.90 014 2,43 2.00 .91 6.02
l/2REC 5.51 1.39 5.76 .52 5089 5.30 3.22 11.70
ZONE 6 24 STATIONS AVG l/2 RECORD = 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.91 .61 1.08 .07 1.54 1.13 .79 .04
IOYR 3.99 057 1.73 006 2.33 1.57 I"12 4.53
l/2REC 2.88 1.38 2.47 .48 2006 1.63 1.24 8.92
ZONE 7 21 SBATIONS AVG l/2 RECORD a 20 YRS
METHOD 1 2 3 4 6 6 7 8
SYR 1.19 082 9.91 .I9 2.18 1.89 11040 004
IOYR 2.33 096 3.58 0'13 3.25 2,15 '1.63 6.52
l/2REC 5.99 1.48 5.36 .I6 3.90 3.90 2.34 12.61
ZONE 8 23 STATIONS AVG l/2 RECORD * 21 YRS
METHOD 1 2 3 4 5 6 7 8
5YR .83 .09 1.28 .Ol .83 .83 .14 .04
IOYR 2.79 ,42 2068 .14 1.78 1.78 842 5.90
l/PREC 2.70 .84 7.62 .49 3.54 3.54 1.32 13.61
ZONE 9 18 STATIONS AVG l/2 RECORD = 25'YRS
1 2 3 4 6 6 7 8
5YR 090 1030 I,37 049 2.33 2.33 1.55 .04
lDYR 3.61 3059 3.22 .42 6.86 5.85 3.90 6.24
l/PREC 3.69 .59 3.97 .53 2.68 1.04 1.07 6.92
1433
TABLE 1411 CONTINUED
ZONE 10 12 STATIONS AK l/2 RECORD = 26 YRS
METHOD 1 2 3 4 5 6 7 0
5YR .02 .04 .25 .04 .22 .22 .04 9.04
loYR .44 .14 .70 .14 .67 .43 .14 3.79
l/2REC 7.21 .27 3.04 .56 1.95 1.95 .27 4.50
ZONE 11 13 STATIONS AVG l/2 RECORD 23 YRS
n
METHOD 1 2 3 4 5 6 7 8
5YR 1.13 1.01 2.15 .20 2.13 1.78 a94 .04
loYR 4.31 2.44 5.95 .72 5.06 3.58 1.90 10.41
l/2REC 1.74 .91 6.38 .46 5.01 4.24 ,91 15.65
ZONE 12 17 STATIONS AVG l/2 RECORD 23 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 2.84 1*22 1.31 045 2.03 1.51 1.27 .04
loYR 4.30 2.17 2.52 010 4027 1.40 2.17 3.37
l/LREC 8.58 .75 .75 .46 2.20 1.34 ,75 4.59
ZONE 13 17 STATIONS AVG l/2 RECORD 26 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.89 1.21 1.11 .32 1.92 1.79 1.21 .04
loYR 1.27 .36 1.39 .14 1.77 1.77 ,53 3.56
l/2REC 4.01 .57 2.83 ,57 3.65 2.43 .55 4.96
ZONE 14 15 STATIONS AVG l/2 RECORD = 25 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 1.91 1.45 1.56 .47 2.66 2.03 1.45 .04
loYR 5.41 2.35 2.81 .14 4.63 2.17 2.35 5.56
l/PREC 3.45 1.04 5.12 .53 9.90 6.99 1.04 6.69
ZONE 15 3 STATIONS AVG l/2 RECORD = 20 YRS
METHOD 1 2 3 4 5 6 7 8
5YR 2.67 3.00 2.54 ,04 3.51 1.25 1.77 ,04
10"YR .14 .14 .14 .14 1.87 1.87 .14 .14
l/2REC 2.17 2.17 .3B .3B 6,15 6.15 0.38 ;38
ZONE 16 i3 STATIONS AVG l/2 RECORD m 24 YRS
METHOD 1 2 3 4 5 6 7 8
5YR .69 062 1,15 .04 1.4D 1.18 069 ,04
loYR 4.02 1.56 3.05 .I$ 3.90 1.97 2901 4.46
l/2REC 8.74 2.37 7.24 051 8.30 602% 3.76 7.24
ALL ZONES 287 STATIONS AVG.1/2 RECORD . 23 YJ&
METHOD 1 2 3 4 5 6 7 8
5YR 1.60 .95 1.40 .21 1.89 1.54 1.01 ,04
101YR a,13 1,40 2.66 .04 3.22 2.19 1,45 6.36
l/LREC 4.66 1.49 4.81 .45 4.99 4.02 1.68 8.80
Values shown are ratios by which the theoretical adjustment for Gaursian
dlstribution samples must be multiplied In order to convert from the
computed 0.001 probability to average observed probabilities In the re
served data.
1434
Table 1411 CONTINUED
Values in table 1411 are obtained as follows:
a. Compute the magnitude corresponding to a given
exceedance probability for the bestfit function.
b. Count proportion of values in remainder of record
that exceed this magnitude,
C. Subtract the specified probability from b.
d. Compute the Gaussian deviate that would correspond
to the specified probability.
e. Compute the expected probability for the given sample
size (record length used) and the Gaussian deviate determined in
d.
f. Subtract the specified probability from e.
g* Divide f by c.
*U.S. GOVERNMENT PRINTING OFFICE:1983X91614/209
1435
GENERALIZED SKEW COEFFICIENTS OF ANNUAL
MAXIMUM STREAMFLOW LOGARITHMS*
The generalized skew map was developed for those guide users who
prefer not to develop their own generalized skew relationships. The map
was developed from readily available data. Users are ercouraged to make
detailed studies for their region of interest using the procedures
outlined in Section V,B2. It is expected that Plate I will be revised
as more data become available and more extensive studies are completed,
The map is of generalized logarithmic skew coefficients of annual
peak discharge. It is based on skew coefficients at 2,972 stream gaging
stations. These are all the stations available on USGS tape files with
drainage areas equal to or less than 3,000 square miles that had 25 or
more years of essentially unregulated annual peaks through water year
1973. Periods when the annual peak discharge likely differed from
natural flow by more than about 15 percent were not used. At 144 stations
the lowest annual peak was judged to be a low outlier by equation 5
using 6 from figure 141 and was not used in computing the skew coeffi
cient. At 28 stations where the annual peak flow for one or more years
was zero, only the remaining years were used in computing the low outlier
test and in computing the logarithmic skew coefficients. No attempt was
made to identify and treat high outiiers, to use historic flood informa
tion, or to make a detailed evaluation of each frequent!; curve.
The generalized map of skew coefficients was developed using the
averaging technique described in the guide. Preliminary attempts to
determine prediction equations relating skew coefficients to basin
characteristics indicated that such relations would not appreciably
affect the isopleth position. Averages used in defining the isopleths
were for groups of 15 or more stations in areas covering four or more
onedegree quadrangles of latitude and longitude.
The average skew coefficients for all gaging stations in each one
degree quadrangle of latitude and longitude and the number of stations
are also shown on the map. Average skew coefficients for selected groups
of onedegree quadrangles were computed by weighting averages for one
degree quadrangles according to the number of stations. The averages
for various groups of quadrangles were used to establish the maximum and
minimum values shown by the isopleths and to position the intermediate
lines.
Because the average skew for 15 or more stations with 25 or more
years of record is subject to time sampling error, especially when the
stations are closely grouped, the smoothed lines are allowed to depart a
few tenths from some group averages. The standard deviation of station
values of skew coefficient about the isopleth line is about 0.55 nation
wide.
Only enough isopleths are shown to define the variations. Linear
interpolation between isopleths is recommended.
The generalized skew coefficient of 0.05 shown for all of Hawaii
is the average for 30 stream gaging stations. The generalized skew
coefficient of 0.33 shown for southeastern Alaska is the average for the
10 stations in that part of the State. The coefficient of 0.70 shown
for the remainder of Alaska is based on skew coefficients at nine stations
in the Anchorage,Fairbanks area. The average skew of 0.85 for these
nine stations was arbitrarily reduced to the maximum generalized skew
coefficient shown for conterminous United States in view of the possi
bility that the average for the period sampled may be too large.
*This g@neraliZed skew map was originally prepared for Bulletin 17 published
in 1976. It has not been revised utilizing the techniques recommended in
Bulletin 17B.