EVALUATION OF THE OIL SPILL
RISK ANALYSIS AS PRESENTED IN
ST. GEORGE BASIN SALE 89 EIS
JONES & STOKES ASSOCIATES, INC. 12321 P STREET I SACRAMENTO, CA. 95816
-------
EVALUATION OF THE OIL SPILL
RISK ANALYSIS AS PRESENTED IN
ST. GEORGE BASIN SALE 89 EIS
Submitted to:
U. S. Environmental Protection Agency
Region 10
Prepared by:
Jones & Stokes Associates, Inc.
1802 136th Place NE
Bellevue, Washington 98005
31 May 1985
-------
TABLE OF CONTENTS
Page
EXECUTIVE SUMMARY i
CHAPTER 1 - INTRODUCTION 1
Purpose and Objectives 1
Overview of MMS Model Approach 2
Summary of St. George Basin EIS 3
Spill Rates for Proposed Action 3
Oil Spill Trajectories 4
Overall (Combined) Spill Risk Probability 4
Cumulative Spill Risk 5
CHAPTER 2 - OIL SPILL RISKS 7
Summary 7
General Discussion 7
Size of Spill 8
Approach to Estimating Spill Frequency 9
The Applicability of Past Experience 10
Independent Events 10
Spills Rates and Volume of Production 11
Mean Case Estimates 14
Reasonableness of the Approach 14
MMS-Assumed Frequencies of Oil Spills 14
Platforms 14
Pipelines 15
Tankers 18
Single Buoy Moorings 21
Summary 21
Verification of Spill Risk Estimates 21
Alternative Approaches 26
The North Sea: A High Latitude Oil 26
Field Model
Alternative Exposure Indices 26
Implications for Impact Assessment 28
CHAPTER 3 - OIL SPILL TRAJECTORIES 31
Summary 31
Overview 31
Trajectory Evaluation 31
General Climatology of the Bering Sea 32
Simulation of Bering Sea Winds 33
Simulation of Bering Sea Circulation 35
Oceanography of the Bering Sea 35
The Numerical Model 36
Verification 38
-------
Page
Ice 38
Influence of Numerical Model on Trajectory 39
Analyses
Batedinic Currents 39
Freshwater Runoff 39
Interpretation of the Trajectory Studies 40
Number of Trajectories 44
CHAPTER 4 - CONCLUSIONS AND RECOMMENDATIONS 47
Overview 47
Sources of Uncertainty 47
Interpretation of the Oil Spill Risk Analysis 48
Recommendations 49
REFERENCES 51
Literature Cited 51
Personal Communications 53
APPENDIX A - EVALUATION OF OIL SPILL RISK
APPENDIX B - DEPICTION OF OIL SPILL STATISTICS
LIST OF PREPARERS
-------
LIST OF FIGURES
Figure Page
3-1 February and August Wind Roses for Designated 34
Locations in the Bering Sea
3-2 MMS-Calculated Risk to Pribilof Islands 41
Resource Areas
3-3 Launch Points Used in MMS Oil Spill Risk 42
Analysis
3-4 Biological Resource Areas 43
-------
LIST OF TABLES
Table Page
2-1 Confidence Intervals for the Expected Value 12
of the Poisson Distribution
2-2 Platform Spills (>1,000 bbl) in U. S. Waters, 16
1955-1980
2-3 Pipeline Spills (>1,000 bbl) in U. S. Waters 17
2-4 Summary of Data on Oil Spills from Vessels 19
Carrying Petroleum as Cargo
2-5 Crude Oil Spills of ^.1,000 bbl from Tankers 20
Worldwide
2-6 MMS-Calculated Expected Number of Spills per 22
Bbbl
2-7 Cook Inlet Spill Data 24
2-8 Reported Impact Assessment for Marine Birds 29
and Marine Mammals
3-1 Monte Carlo Error as a Function of Number of 45
Trials and Estimated Probability
3-2 Percent Error for a Biased and Unbiased Coin 45
-------
EXECUTIVE SUMMARY
An EIS must fully disclose the information and analytical
procedures used in assessing impacts. The oil spill risk
analysis is a key component of EISs prepared for OCS oil and gas
lease sales. EPA Region 10 became concerned about the
reasonableness and adequacy of the oil spill risk analysis used
in EISs for oil and gas lease sales in the Bering Sea. This
report describes and evaluates the current approach to the oil
spill risk analysis as conducted for St. George Basin Sale 89.
Overview of Approach to Risk Analysis
An oil spill trajectory analysis (OSTA) model was developed
for MMS to calculate the risk of oil spills damaging
environmentally sensitive resources. The model has three
distinct parts. These parts produce a combined probability that
one or more large (J>1,000 bbl) oil spills will occur during the
life of the field and contact a sensitive resource area.
The first part of the OSTA model uses historical spill data
from all U. S. OCS areas to estimate the expected number of
spills for different types of activities (e.g./ platform spills,
modes of transportation). The second part of the model predicts
trajectories of spilled oil given that a spill occurs at selected
points. The third part of the model combines the probability
that a spill will occur with the probabilities associated with
trajectory simulations. Each launch point is weighted by the
volume of oil handled. The total risk to a resource area is the
probability that spills will occur at each launch point combined
with the probability that those spills will reach the target.
Oil Spill Risks
As might be expected, the oil spill risk evaluations are
inherently uncertain. Uncertainty arises from a number of
sources: the estimates of mean-case resource levels, the
statistics derived from historical spill records, and the
scenarios chosen for development of the field. Because the risks
are based solely on the mean-case resource estimate, the
probability of a spill is derived from a single exposure index
tied to production.
The uncertainty in frequency of very large spills
0100,000 bbl) is extremely large because it is based on an
extrapolation of statistics outside of the range of observations.
-------
There is no unique way to pose a spill risk model. The
approach taken is as reasonable as can be if it is a. j3r_.ipj:i
constrained that the risks are to be based on a single exposure
index.
Oil Spill Trajectories
The ocean and atmosphere modeling studies used by MMS appear
in principle to be capable of generating a reasonable and
adequate impact assessment. The ocean model is capable of
describing oceanic circulation, and the meteorological model is
capable of simulating winds. There remains concern about the way
the two models are coupled since the coupling is done at the
expense of faithful reproduction of baroclinic currents.
However, the potential errors are not considered serious since
baroclinic currents are not a major concern with regard to
surface dispersion of oil.
The most serious concerns are .that the number of
trajectories used may be insufficient for adequate projection of
risks and that summer conditions may be under-represented in the
statistics reproduced in the EIS.
Conclusion and Recommendations
Insufficient information is presented in the EIS and its
support documents to evaluate the reasonableness and adequacy of
the oil spill risk analysis. Personal communication with the
analysts was necessary to obtain important information about some
basic features of the risk analysis.
Although the oil spill risk analysis appears to be capable
of providing a reasonable and adequate assessment of impacts, its
interpretation and use in the EIS should be modified in several
key points. The reason for this finding is based primarily on
the recognition of great uncertainties in the risk assessment and
the subsequent need for a reasonable and adequate worst-case
analysis. We recommend that:
Sufficient trajectories should be run to assess error
bounds on their probability distributions.
Greater consideration should be given to use of
conditional probabilities in assessing the
environmental consequences of the project.
The EIS should include a discussion of the impact of
season on conditional probabilities, particularly the
summer season with reference to the Bering Sea lease
sale EISs.
Little credence should'be given to the EIS's summary of
overall risk as currently presented.
11
-------
The EIS should better document what meteorological and
oceanographic features of the environment are
incorporated in the trajectory simulation model. A
separate document should be published that describes
how these features are modeled.
The EIS should clearly document the assumptions made
and their implications when appropriate. In
particular, those implications that may ultimately
compromise the worst-case analysis must be clearly
stated.
111
-------
Chapter 1
INTRODUCTION
Purpose and Objectives
The Minerals Management Service (MMS) of the U. S.
Department of the Interior conducts oil and gas leasing on the
U. S. outer continental shelf (DCS). The leasing process is
subject to the National Environmental Policy Act (NEPA), which
requires that MMS evaluate potential environmental impacts such
as an oil spill damaging environmentally sensitive resources.
NEPA also requires that the Environmental Impact Statement (EIS)
be a full disclosure document, i.e., that the assumptions and
analytical methodologies be clearly described in plain language
such that decision makers and the public can understand how the
findings of fact have been derived from the description of the
proposed action and the affected environment (40 CFR Section
1502.1 and Section 1502.8).
EPA Region 10, pursuant to NEPA and Section 309 of the Clean
Water Act, reviews draft and final EISs prepared by MMS for
proposed oil and gas lease sales. During review of the DEIS (DOI
1984) for Lease Sale 89 (St. George Basin), EPA Region 10 became
concerned about the documentation for the oil spill risk analysis
and the reasonableness and adequacy of the models used to develop
the risk analysis. With respect to biological resources in the
affected environment, the oil spill risk analysis is one of the
most important features of the impact assessment.
The purpose of this report is to review the MMS oil spill
risk analysis as presented in the Lease Sale 89 EIS (DOI 1985)
and its supporting documents. The objectives of the report are
to:
Describe the MMS approach to the risk analysis.
Determine whether the underlying assumptions and
structure of the risk analysis are reasonable.
Provide recommendations to EPA regarding the
interpretation and use of the oil spill risk analysis
in assessing the environmental consequences of the
proposed action and alternatives.
-------
Overview of MMS Model Approach
An oil spill trajectory analysis (OSTA) model was developed
for MMS to calculate the risk of oil spills contacting
environmentally sensitive resources. The OSTA model is described
by Smith et al. (1982). The primary concern in using the OSTA
model are spills of 1,000 barrels (bbl) or larger. The"
1,000 bbl cutoff was selected to limit evaluations to those
spills large enough to travel long distances on the ocean surface
and have the potential to do serious damage (Smith et al. 1982;
Lanfear and Amstutz 1983).
The model used by MMS has three distinct steps that are
taken to calculate probable risk to resources. The first step is
common to all oil spill risk analyses prepared for offshore oil
and gas lease sales. It calculates the unit risk, i.e., the
expected number of oil spills of 1,000 barrels (bbl) or greater
per unit volume of oil handled during certain types of
activities. As might be expected, there is considerable
uncertainty in forecasting whether spills will occur and, if so,
how many and how large. Thus, the model uses a probability
distribution partially based on historical data from other U. S.
OCS lease sales. Spill occurrence rates for different activities
were developed by Lanfear and Amstutz (1983) and Nakassis (1982).
Spill rates are assumed to be directly proportional to the volume
of oil produced and are reported as expected number of spills per
billion barrels (Bbbl) produced or handled.
The second part of the model addressees the conditional
probability of an oil spill hitting a specified target, i.e., it
assumes that a large spill occurred at a specified location. The
likely paths or trajectories of an oil spill are also
probabilistic because they depend on wind and current conditions
during and following the spill. Conditional probabilities
reported in the EIS are averaged out for the life of the project,
i.e., seasonal variations in trajectories are averaged. Thus,
the conditional risk reported in the EIS is likely to be greater
or less than the true conditional risk for a given season.
The output from the first and second parts of the model are
then used to estimate the combined risk to specified sensitive
resource areas. The MMS OSTA model uses matrix algebra
(specifically, matrix multiplication) to calculate the combined
probability of an oil spill occurring and making contact with a
target. One matrix lists the conditional probabilities derived
from the trajectory analysis, i.e., each element in the matrix
represents the mean probability that target "i" is hit by a spill
occurring at point "j". The second matrix represents spill
occurrence, i.e., each element in the matrix represents the
expected number of spills occurring at "j" as a result of
production of a unit volume of oil at site "k". The points "j"
and "k" are distinguished because some launch points ("j")
represent locations along a pipeline or tanker route.
-------
The resulting product matrix is then multiplied by the
volume of oil expected to be found at production site "k" to
obtain a final product matrix containing the expected number of
oil spills that occur and contact target "i". Thus, each launch
point is weighted by the amount of oil handled over the life of
the project at that point. Thus, the combined probability is
used to determine the overall combined probability, which
expresses the risk to a specified target from all the launch
points. It is this latter figure (sum of the combined
probabilities) that is instrumental in the EIS assessment of
impact to environmental resources. The mathematical derivation
of combined probability for each launch point is important to
understand because this probability is less than the conditional
probability unless the volume of oil handled at that launch point
is high enough that the expected number of spills is 21.
Before examining the details of the oil spill risk analysis,
it should be noted that large spills (>1,000 bbl) are assumed by
MMS to be rare, random, independent events. The Poisson
probability distribution is a mathematical method of describing
the probability of such rare events. The Poisson distribution is
defined by only one parameter: in this case, the expected number
of spills. Thus, the elements in the final product matrix are
inserted in the Poisson distribution formula to calculate the
probability of one or more spills occurring and hitting a
specified target ("i").
Summary of St. George Basin EIS
The following discussion briefly describes the oil spill
risk analysis as used (Samuels 1984) for the proposed action
described in the St. George Basin EIS. The proposed action calls
for pipelines connecting production platforms north of 56ฐN
Latitude with a facility on St. George Island. Oil would then be
tankered south through Unimak Pass. Production platforms south
of 56ฐN Latitude would connect to an offshore collection platform
where oil would be loaded and tankered south.
Spill Rates for__Proposed Action
Unless specified otherwise, spill rates described in this
chapter refer to spills ฃ1,000 bbl, i.e., those treated in the
oil spill risk analysis for the EIS. Spill occurrence rates were
calculated separately for the northern and southern sectors of
the lease sale with the assumption that the mean case production
scenario for the field (1.124 Bbbl) would be equally divided
between northern and southern sectors (Hale pers. comm.).
For transportation activities in the northern sector, spill
rates of 1.6/Bbbl for pipelines, 0.2/Bbbl for tankers in port,
and 0.9/Bbbl for tankers at sea were used. For tankers at sea,
MMS assumed that the rate would be 0.45 rather than 0.9/Bbbl
because of a 50 percent chance that the spill would not occur in
the lease sale area (Hale pers. comm.). Thus, for the northern
-------
sector, transportation-related spill rates were assumed to be
2.25/Bbbl (1.6 for pipelines, 0.2 for tankers in port, and 0.45
for tankers at sea), or 1.26 transportation-related spills
expected for the northern sector of the lease sale (2.25 x 0.56).
Similar calculations were made for the southern sector;
however, spill rates for tankers loading from collection
platforms were assumed to be included in spill statistics for
tankers at sea (Prentki pers. comm.; DOI 1985, p. IV-7). Thus,
for the southern sector, transportation-related spill rates were
assumed to be 2.05/Bbbl (1.6 for pipelines and 0.45 for tankers
at sea), or 1.15 spills expected (2.05 x 0.56).
Platform spills 11,000 bbl were assumed to occur at a rate
of 1.0/Bbbl, i.e., 1.12 spills for the entire lease sale. Thus,
for the St. George Basin lease sale, 3.54 spills are expected
(2.42 for all transportation plus 1.12 for platforms) for the
proposed action. Using the MMS-assumed probability distribution
(the Poisson distribution), the expected number of spills yields
a probability of 0.97 that there will be one or more spills of
^1,000 bbl.
Oil Spill Trajectories
In Atlantic, Gulf of Mexico, and California DCS lease sale
EISs, MMS used the full OSTA model to predict oil spill
trajectories. Wind and current data were used and applied to as
many as 100 launch points and 500 hypothetical spills from each
launch point for each season of the year.
In Bering and Beaufort Sea lease sale EISs, a numerical
model developed by the Rand Corporation (Liu and Leendertse 1982)
was used to predict trajectories. Oil spill trajectories were
generated from 29 launch points within the lease sale boundary
with 26 trajectories each during ice-free conditions and 36
trajectories each for winter conditions of average ice cover.
The trajectories were used by MMS with the OSTA model to evaluate
the probability of risk to targeted resources.
Overall (Combined) Spill Risk Probability
The volume of oil assumed at each launch point is critical
to the calculation of the combined probability. In the case of
Sale 89, the lease sale area can be divided into subregions based
on the resource estimates. The Pribilof Island and Unimak Pass
deferral areas are assumed to contain 10 and 5 percent of the
total resource, respectively. The remaining subregions of the
northern and southern sectors contain 40 and 45 percent of the
resource, respectively. Within each subregion, the expected
production is divided equally between the launch points (Hale
pers. comm.).
For transportation-related spills, the expected number of
spills associated with the volume produced at "k" is divided
equally among the launch points carrying that volume. The
-------
expected number of spills at a particular launch point is the sum
of these allocated fractions. This sum is then multiplied by the
volume of oil produced at "k" to give the weighted spill risk.
Thus, the spill is geometrically accumulative as oil moves
through a collection system to a storage or processing area.
However, the allocation and weighting procedure does not result
in "multiple counting" of the transported volumes for that
particular mode of transportation. In other words, the expected
number of transportation-related spills for a particular
transportation mode in the lease sale does not exceed the unit
risk for that mode of transportation multiplied by the mean-case
production estimate for the field. "Multiple counting" occurs
only in the sense that the same barrel of oil may be collected
and transported by pipeline and then shipped by tanker out of a
storage facility and, therefore, that oil is exposed to two
separate transportation-related risk calculations.
Cumulative Spill Risk
As part of its evaluation, MMS includes an analysis of oil
spill risk from the cumulative activities in the Bering Sea,
i.e., other Bering Sea lease sales and transportation of oil from
Arctic production fields. In developing the cumulative-case
scenario for the DEIS, MMS assumed that the proposed action for
Sale 89 will be replaced by the pipeline transportation
alternative because the North Aleutian Basin lease sale (Sale 92)
involves pipeline activity adjacent to St. George Lease Sale 89,
and it would be more reasonable to tie the southern sector of
Sale 89 to the Sale 92 pipeline system (Hale pers. comm.).
Expected number of spills in the cumulative case (DOI 1984,
Table IV-9) is calculated from the sum of the operations
(DOI 1984, Table IV-5) and used in the Poisson distribution
formula. In the FEIS, MMS assumed that the proposed action for
Sale 89 would also be valid for the cumulative case, and the
expected number of spills (DOI 1985, Table IV-10) is readily
obtained from data in Table IV-5 of the FEIS.
-------
-------
Chapter 2
OIL SPILL RISKS
Summary
MMS uses a log-normal distribution to estimate the
probability distribution for sizes of spills. The agreement
between calculated and observed frequencies of spill volume
appears reasonable. Although the approach provides the best
estimate that can be made, it should be noted that the approach
can result in extremely large uncertainties.
Similar conclusions are made for the frequency of large
spills. MMS assumes large spills occur as a Poisson process and
that the expected number of spills can be derived from past DCS
history and the mean-case resource estimate for the proposed
field. The approach taken by MMS is reasonable, but the expected
number of spills that is derived by the approach is characterized
by great uncertainty. An important element of the risk analysis
is the selection of the exposure index used to develop the mean
expected number of spills. The assumptions made about the
exposure index are not demonstrably better than alternative
assumptions, and alternative assumptions very likely would alter
the expected number of spills. However, the degree of
uncertainty at all steps in the calculations is so large that
only great changes in the expected number of spills are likely to
have any real meaning.
The large uncertainty inherent in the calculation of
expected number of spills is of greater significance to the
impact assessment than small modifications of the calculated
number.
General Discussion
Risk assessments are generally probabilistic in nature.
Three fundamental elements in a risk assessment model are:
1) the probability distribution function, which provides the
probabilities of events; 2) an exposure index, which sets the
probability parameters; and 3) the ability to predict future
events using the exposure indices. A sufficient number of
historical events, each with the same exposure index, allows
evaluation of the probability function. With a small database,
however, assumptions on the nature of the probability function
must be made. These assumptions are not verifiable in an
absolute sense and can only be judged by their reasonableness.
-------
To determine the exposure index requires first that the
distribution function is known and then that it is known for a
number of different exposures. Exposure indices that are good
fits to the data but not predictable into the future are of
little value. If the exposure index is not predictable, then
forecasting is not possible.
The MMS oil spill risk analysis model must be considered
within the constraints mentioned above. Because vastly different
inferences can be drawn from similar data, the burden of proof is
always on the declarer to present a thorough description of the
steps leading to conclusions.
Data on spill incidents are kept by a number of sources:
the U. S. Coast Guard, the U. S. Geological Survey, EPA, and
Lloyds of London. The recent compilation of data for the DOI by
The Futures Group (1982) represents the most- complete set of
publically available information. The Futures Group report
classifies spills by five sources: platforms, pipelines, single
buoy moorings, tankers at sea, and tankers in ports. These
categories, where possible, are further subdivided by cause of
spill. The Futures Group, however, was not able to complete the
analysis of appropriate exposure indices; therefore, their
conclusions are tentative and expressed as suggestions for
further study.
The technique adopted by MHS to highlight the
disproportionate importance of the rare, large spill is to
analyze the problem in terms of spill frequency models and spill
volume distributions. The former are used to predict the number
of spills that might occur, the latter to predict the volume
spilled given the event. The predictions are made in terms of
probabilities, i.e., "n" spills will occur with the probability
P(n) and less than "x" gallons will be spilled with probability
P(x) .
Size of Spill
MMS selects spills of volume ^1,000 bbl for its risk
analysis model based on the belief that portions of spills of
this size or larger remain on the water surface long enough
(30 days) to be transported away from the source to
environmentally sensitive areas (Lanfear and Amstutz 1983).
These authors also point out that a 1,000 bbl spill is large
enough not to go unnoticed, so reporting records tend to be
reliable.
A 10,000 bbl spill is likely to have the same environmental
impact as a 12,000 bbl spill. Thus, in terms of risk analysis,
the size of a spill need not be defined to great precision. In
the oil spill risk analysis, therefore, the primary effort is
devoted to frequencies of spillage. The frequency distribution
for spill volumes is of interest, however, from the standpoint of
worst-case risk analyses. MMS assumes that spill volumes
-------
2100,000 bbl constitute a worst-case; smaller spills are more
frequent but also tend to be more localized in their adverse
effect. To determine the probability of an event requires a
great deal of data. It is impossible to determine with a small
data set the correctness of any of the distribution functions
chosen. Unfortunately for the statistician, there have been too
few spills of 2100,000 bbl for proper study. Thus, a means is
needed to estimate the probability of these events.
Lanfear and Amstutz (1983) compare the distribution of spill
sizes (spills 21*000 bbl) for both platform and pipeline spills
with a log-normal distribution. The agreement is reasonable and
the log-normal distribution is probably the best assumption that
can be made. It should be noted that the log-normal distribution
is a generic distribution used to fit widely scattered data.
Deriving rates for this distribution, particularly outside of the
range of data used to evaluate the distribution, can result in
extremely large uncertainties.
In the FEIS for Lease Sale 89 (DOI 1985), spills of
21,000 bbl and 2100,000 bbl were considered. Because there have
been no platform spills of 2100,000 bbl, the log-normal
distribution function was used to evaluate the probability of
platform spills 2100,000 bbl. The extrapolated rate constant is
0.036/Bbbl. Of eight pipeline spills in the DCS records, one was
>100,000 bbl. Using the same distribution function as for
platform spills, MMS evaluated a spill rate of 0.065/Bbbl for
pipeline spills 2100,000 bbl (DOI 1985, p. IV-7).
.Approach to Estimating Spill Frequency
The following critical assumptions about spill frequency are
used by MMS in its risk analysis:
Future spill frequencies can be based on past DCS
experience (DOI 1985, p. IV-6).
Spills occur independently of each other (DOI 1985,
p. IV-6).
The spill rate is dependent on the volume of oil
produced or transported (DOI 1985, p. IV-6).
The mean-case estimate of the oil resource can be used
to estimate the volume of oil produced (DOI 1985,
p. IV-6).
The reasons for and implications of these assumptions are useful
in interpreting the reasonableness of the oil spill risk
analysis.
-------
The Applicability of Past Experience
It would be unreasonable to presume that the physical
environment (or working conditions) of the harsh, stormy Bering
Sea is comparable to existing U. S. DCS conditions (i.e., the
Gulf of Mexico and southern California). Even Cook Inlet does
not experience the storm and wave conditions typically found in
the Bering Sea.
It is reasonable, however, to presume that production
platforms will be engineered to meet the environmental rigors of
the lease sale area. This implies a second assumption, i.e.,
that the engineering of platforms has met the environmental
rigors of existing oil fields. The implied assumption has not
been strictly satisfied, but improvements have been made in
technology as a resu-lt of previous accidents. Since oil recovery
technology is not starting from a new basis in each new
development, it is permissible to take into account trends in
reduction of spill rates in the risk evaluation. MMS currently
evaluates risk based on a model (Nakassis 1982) which maintains
the critical assumptions in the risk analysis but allows for
industry improvement. Analysis shows that platforms and tankers
have shown improvements in spill rates over time; pipelines have
not.
Independent Events
A basic assumption of the approach taken by MMS is that
spills occur as a Poisson process, with volume of oil produced or
handled as the exposure variable. One of the key supporting
documents for the MMS oil spill risk analysis is the Nakassis
(1982) study of platform spills, discussed in detail in
Appendix A. Nakassis was unable to conclude that large oil
spills occur as a Poisson process, but he assumed a Poisson
process in order to test the hypothesis that platform spill rates
have decreased over time. Subsequent work (Lanfear and Amstutz
1983) continues to assume large oil spills occur as a Poisson
process.
One way of understanding the Poisson distribution is to
understand how it may be derived from games of chance. The
outcome of such games is a win or lose situation with a
probability of winning or losing. How does one fare after "n"
trials in a game with a probability "p" of winning and a
probability "q" of losing? The answer is given by the binomial
distribution function, but this function is complicated. If the
gambler seldom wins and wins are random and independent events,
success is approximated by a Poisson process, which is a
considerably simpler mathematical expression to deal with. In a
number of problems, rare and random events are counted and the
average rate (expected value) of some process is thereby
estimated using the hypothesis that they are Poisson distributed.
Oil spills are approximated by this Poisson distribution on
the assumption that they are rare events. Several oil spills in
10
-------
the Gulf of Mexico occurred in October 1964 as a result of a
single hurricane. Consistent with the Poisson assumption, MMS
treats these as one event. In accident work, the fact that
people are more careful after a series of accidents causes
deviations from the Poisson law. MMS modifies the Poisson law
parameters to include industry improvement in its rate
predictions (Nakassis 1982).
Table 2-1 shows how well the mean expected value of a
Poisson process is estimated as a function of the number of
observed events. The table illustrates the difficulty of
estimating a true mean even if it is known that the information
comes from a Poisson process. Verifying that a process is
Poisson would be even more difficult.
Occasionally it will be possible to test directly the
Poisson assumption in its entirety. If there are numerous
observations, each with the same exposure, then the associated
numbers of spills represent independent observations. from_ a^ J,',",
single Poisson distribution, and the standard statistical te^sts
for goodness-of-fit can be employed. A possible case is tanker
spills, where a contemplated exposure index is tanker-years.
Every tanker which has been in service for the same period will
have the same exposure. Stewart and Kennedy (1978, p. 24)
performed goodness-of-fit tests in this situation and concluded
the Poisson model was acceptable.
Spill Rates and Volume of Production
A risk assessment model cannot be formulated without an
exposure index. Criteria for choosing an exposure index (Smith
et al. 1982) are:
It should be simple.
It should not intuitively violate to any significant
extent technical assumptions made in the analysis (this
refers primarily to the Poisson assumption).
ซr It should be a quantity that is predictable in the
future.
MMS has chosen to characterize the risk associated with
production of oil from a lease sale area with a single number:
the expected number of spills of 21/000 bbl during the life of
the field. This number is related to experience by the
assumption that the number of spills is proportional to the total
expected production from the oil field. Thus, if the anticipated
production is a tenth of the production of the "past experience
production," then the risk is a tenth of that in the "past
experience data."
The approach scales all risks according to anticipated
volume of production. This is a simplification of oil recovery
operations that avoids consideration of the many factors that
11
-------
Table 2-1,
OBSERVED
COUNT
0
2
4
6
8
10
20
Confidence Intervals for the Expected Value of
the Poisson Distribution
LEVEL OF CONFIDENCE
LOWER
LIMIT
90%
UPPER
LIMIT
LOWER
LIMIT
98%
UPPER
LIMIT
0.0
0.355
1.37
2.61
3.98
5.43
13.25
3.0
6.3
9.15
11.84
14.43
16.96
29.06
0.0
0.149
0.823
,79
,91
4.13
11.08
1,
2,
4.61
8.41
11.60
14.57
17.40
20.14
33.10
SOURCE: Wilson 1952
12
-------
bear on possible causes of spills. There is some justification
for this approach. At the time of leasing, it is not known if
oil will be discovered and, if so, where in the lease sale area
the oil would be found. Because the risk statistics are based on
interpretation of past experience independent of causal factors,
scenarios depicting the number of wells, tanker traffic, or
lengths of pipelines are used only to allocate "share of the total
risk. A consequence of this choice of exposure index is the
important assertion that all activities involved in the
production of oil over the lifetime of the field are directly
related to total production. Thus, even spills associated with
transportation of oil are directly related to volume of oil
produced.
It may not be practical to find an alternative exposure
index to total production. It can never be known if oil is
present in commercially marketable quantities until a discovery
has been found by drilling. It is also reasonable to assume that
environmental regulation and environmental awareness of the
industry will result in technology meeting the environmental
rigors. However, adoption of the above assumptions precludes the
inclusion into the risk analysis those factors which individually
constitute real elements in the risk. Such factors are:
rapid changes in production as a result of world oil
demand, which might result in the use of developing
technology;
production with a few very large platforms rather than
many smaller platforms;
rapid or slow development based on economic conditions
of the industry;
large local tank storage because the marine climate
shortens "windows" during which tankers can be safely
loaded;
tectonics of the region;
use of the lease area in activities other than oil
production, e.g., fishing in the Bering Sea;
comparison of alternate modes of oil transportation in
terms of net risk; and
important variations in transportation modes, e.g.,
long and large, pipelines instead of many short and
small pipelines, and/or the use of a few large tankers
instead of many small tankers.
13
-------
Mean Case Estimates
Anticipated resource levels and volume of production
during the lifetime of an oil field are derived from geological
information and information from neighboring regions. It is
difficult to place a strict confidence limit on this number. The
resource estimates are based on primary production methods, thus
volume of production and the expected number of spills are also
based on primary production methods. Limitation of the oil spill
risk analysis to primary production is perhaps reasonable since
it is not certain that oil will be found let alone whether
secondary production is economically practical. It is important,
however, to recognize that this limitation may have a bearing on
the interpretation of the mean case and maximum case resource
estimates and production life of the field.
Reasonableness of the Approach
It is clear from the foregoing discussion of the assumptions
of the spill risk analysis that estimations of expected number of
spills are characterized by great uncertainty. The degree of
confidence in the calculated expected number of spills decreases
as one examines in detail its derivation. This is not to say
that the approach is unreasonable, rather it means that
unquestioning acceptance of the expected number of spills at face
value is unreasonable. The approach is reasonable to the extent
that the calculated values can be used to determine whether the
probability of a spill is relatively high or low. Thus, the
great deal of effort expended in refining the expected number of
spills (even by as much as a factor of two or so) may not be
warranted in view of the large confidence interval inherent in
the Poisson statistics (Table 2-1).
MMSAssumed Frequencies of Oil Spills
The MMS OSTA model uses risk estimates for different
operations in offshore oil production. There are two main
categories: platform and transportation spills. Included in
platform spills are blowout, tank, and miscellaneous spill
categories. Included in transportation are pipelines and tankers
at sea and in port. The spill rates are derived from historical
data using assumptions about the industry. In all cases
discussed below, only spills 11,000 bbl are considered unless
otherwise noted. The number of spills this large in the
historical record is not uniformly agreed upon by the various
analysts who have used the spill records.
Platforms
Before 1981, OSTA model runs used DCS platform spill rates
based on studies by Devanney and Stewart (1974) and Stewart
(1975). The database for this work was 10 spills of 21,000 bbl
in handling 5.338 Bbbl, yielding a rate of 1.87 spills/Bbbl.
14
-------
Samuels et al. (1981) used U. S. Geological Survey (USGS)
accident records (1979a, 1979b), which reported nine spills of
11,000 bbl from 1964-1975, and used a 1964-1980 federal DCS oil
production of 4.386 Bbbl to compute a rate of 2~.Q5 spills/Bbbl of
oil produced.
Nakassis (1982) examined the spill record and concluded that
a trend existed that indicated improvement in the platform spill
rate (Appendix A). Using a maximum-likelihood approach, he
estimated that the present spill rate for U. S. OCS platforms is
0.79 spills/Bbbl. Nakassis began with the assumption that spills
can be represented by a Poisson process. He did not prove that
oil spills come from a Poisson process (Appendix A).
To help update its own estimates of spill rates, DOI
contracted with The Futures Group to prepare a database of oil
spills and to perform a preliminary analysis of spill rates.
Completed in September 1982, the database contains detailed
records of platform, pipeline, and tanker spills. The Futures
Group database contains records of 462 platform accidents
worldwide from 1955-1980, including 15 spills of 11,000 bbl in
U. S. waters.
Lanfear and Amstutz (1983) used Nakassis1 methodology on the
database in Table 2-2 (i.e., excluding three spills that were
included by The Futures Group: two spills in 1964 [2,559 bbl and
6,387 bbl] and one spill in 1969 [18,363 bbl]) and computed a
spill rate of 1.0 spills/Bbbl of production. This is the figure
used in the EIS for the St. George Basin lease sale. Lanfear and
Amstutz (1983) modified The Futures Group data somewhat for their
analysis. There were several spills in October 1964 that
occurred during a single hurricane. Since these spills occurred
as a result of the same event, Lanfear and Amstutz chose to make
these spills a single event because, if spills are not
independent events, they will not be modeled by a Poisson
process. Because these spills came early in the study period
(1964-1980), the analysis would have indicated greater industry
improvement had they not been grouped together. Thus, the
analysis is conservative in the way it treated this information.
Pipelines
Most U. S. OCS oil produced is transported via pipelines.
The MMS database for U. S. OCS pipeline spills is in Table 2-3.
Because anchor dragging is the prime cause of large pipeline
spills (The Futures Group 1982), there is reason to expect that
pipeline length is the important indicator of risk. The Futures
Group tentatively concluded that pipeline length was a more
accurate predictor of failure rate than volume transported. The
Futures Group considered all failures regardless of spill size:
a total of 235 accidents and a mean spill size of 190 bbl. The
Futures Group concluded that corrosion-related failures were on
the increase; however, most corrosion-related failures result in
small spills.
15
-------
Table 2-2.
Platform Spills OlfOOO bbl) in U.
1955-1980
S. Waters,
DATE LOCATION
4/8/64 Eugene Island 208
10/3/64 (7 Platforms)
7/19/65 Ship Shoal 29
1/28/69 Santa Barbara
3/16/69 Ship Shoal 72
8/17/69 Main Pass 41
2/10/70 Main Pass 41
12/1/70 South Timbalier 26
7/20/72 (Unspecified, Gulf
of Mexico)
1/9/73 West Delta 79
11/23/79 Main Pass 51
11/17/80 Galveston
SIZE (bbl)
5,108
17,500
1,688
77,000+
2,500
16,000
30,500
53,000
4,300
9,935
1,500
1,500
CAUSE
Collision
Hurricane
Blowout
Blowout
Blowout
(weather)
Tank Spill
(weather)
Blowout
Blowout
Unspecified
Tank Spill
Tank Spill
Tank Spill
+ = Estimates vary.
SOURCE: Lanfear and Amstutz 1983.
16
-------
Table 2-3. Pipeline Spills (>1,000 bbl) in U. S. Waters
DATE LOCATION SIZE CAUSE
10/17/67 West Delta 73 160,638 Anchor Dragging
3/12/68 South Timbalier 131 6,000 Anchor Dragging
2/11/69 Main Pass 299 7,532 Anchor Dragging
5/12/73 Grand Island 73 5,000 Corrosion
4/18/74 Eugene Island 317 19,833 Anchor Dragging
9/11/74 Main Pass 73 3,500 Environmental
12/18/76 Eugene Island 297 4,000 "Damaged"
7/17/78 Eugene Island 215 1,000 Anchor Dragging
SOURCE: Lanfear and Amstutz 1983.
17
-------
Samuels et al. (1981), using USGS accident data from 1964 to
1979, computed a rate of 1.82 spills/Bbbl produced for spills of
2lrOOO bbl. For large spills (21,000 bbl), Lanfear and Amstutz
(1983) stated, "On a likelihood basis, volume of oil is better
than km-yr in explaining the spill record. The length of
pipelines has increased more than threefold since 1969, with no
corresponding increase in spill occurrences. Perhaps km-yr,
adjusted for some experience factor, may yet prove to be a
superior exposure variable. However, such an adjustment would
cost a statistical analysis at least two degrees of freedom (for
shape and parameter value), making its superiority very difficult
to demonstrate with only eight spill occurrences." On the basis
of data through 1980, Lanfear and Amstutz suggested that the
expected spill rate should be 1.6 spills/Bbbl produced.
Tankers
Tankers present a problem in assigning risks. The
difficulties of finding exposure indices for tankers are
considerable. For-example, 1983 had the lowest rate of tanker
accidents in 16 years. The world tanker fleet shrank in 1983,
and more restrictive operating rules were in effect for this
year; nevertheless, recent figures in the 1983 Oil Spill
Intelligence Report (New York Times, Oct. 7, 1984, p. 34) show
a 930 percent increase in spill volume in 1983 vs. 1982.
Approximately 241.8 million gal of oil were lost by spillage,
fire, or sinking in 1983. This is the largest amount of oil lost
since 1979. Of the total, 80 million gal were associated with
the Middle East conflict and an additional 78.5 million gal were
lost when a tanker burned and sank near South Africa.
The DOI did not maintain a database of tanker accidents as
it did for platforms and pipelines. All tanker spill rates were
derived from published world-wide spill data. Devanney and
Stewart (1974), examining spills on major trade routes, reported
99 spills of 2.1,000 bbl in transporting 29.326 Bbbl of oil.
Stewart (1976) reported 178 spills in transporting 45.941 Bbbl of
oil, for a rate of 3.87 spills/Bbbl; all of these spills occurred
before 1976.
The Futures Group (1982) database provided the DOI with the
first opportunity since 1976 to review and update the tanker
spill rates. Because of the difficulty and expense of collecting
spill data, primary emphasis was placed on collecting data on
spills of 21,000 bbl since 1974, although spills of all dates and
sizes were included. The data summarized in Table 2-4 contain
855 records of accidents involving vessels engaged in
transporting oil as a product.
Spills of crude oil of 21,000 bbl, from tankers worldwide
are shown in Table 2-5. That at least 31 percent of the spills
occurred in harbors or at piers is particularly important for
evaluating environmental impacts, as these spills would not be
subject to the same advective and weathering effects of winds and
currents as spills in open water on the DCS. Earlier analyses
18
-------
Table 2-4. Summary of Data on Oil Spills from Vessels
Carrying Petroleum as a Cargo
NUMBER OF SPILLS
YEAR ANY SIZE >1.000 bbl
Pre-1969 49 33
1969 20 13
1970 40 22
1971 47 19
1972 89 44
1973 78 49
1974 82 30
1975 67 27
1976 57 26
1977 88 34
1978 81 27
1979 111 43
1980 76 27
TOTAL 885 394
SOURCE: Lanfear and Amstutz 1983.
19
-------
Table 2-5. Crude Oil Spills of 21fOOO bbl from Tankers
Worldwide, by Location
YEAR AT SEA IN PORT UNSPECIFIED TOTALS
1974 10 82
1975 9 43
1976 16 41
1977 12 40
1978 8 12
1979 11 91
1980 _ฃ _5_
TOTAL 69 35 10 114
SOURCE: Lanfear and Amstutz 1983.
20
-------
did not make this important distinction. The Futures Group could
not find a statistically significant exposure index. The MMS
analysis relies on the Lanfear and Amstutz (1983) paper on spills
21,000 bbl. Using an exposure of approximately 88 Bbbl of oil
transported between 1974 and 1980, the calculated spill rate
becomes 0.90 spills/Bbbl for spills at sea (open, restricted, or
unknown waters) and 0.40 spills/Bbbl for spills in port (harbors
and piers), for a total of 1.3 spills/Bbbl. Spills in port must
be assumed to be divided evenly between the inbound and outbound
portions of the voyage, as the database does not make this
distinction.
The tanker spill rate since 1974 appears to be only a third
of that before 1973. Stewart (1976) reports more spills before
1976 than are contained in The Futures Group database, but this
could be due to the latter group's incomplete collection of data
from the earlier years (emphasis was on years 1974 and later).
Goldberg et al. (1981) also report more incidents for the years
before 1972 than does The Futures Group but about the same number
for later years. (Their classification scheme, however, is not
exactly the same; individual records are not available, so the
comparison is only approximate.) Unless the databases are very
much in error, it appears that the tanker spill rate for spills
of 2.1,000 bbl dropped significantly sometime between 1972 and
1974.
Single Buoy Moorings
The St. George Basin Sale 89 FEIS (DOI 1985) assumes that
the risk for single buoy moorings is included in the risk for
tankers at sea. The Futures Group considered it a separate
category. There have not been any spills of J>1000 bbl
associated with si-ngle buoy moorings.
The Futures Group noted a significant increase (above
port-related spills) in spill volume associated with loading at
single buoy moorings, although the spill rates were about the
same as for ports. Ship calls were deemed the best exposure
index. Spill rates were three times larger at unloading than at
loading; however, volumes lost during unloading were considerably
smaller than during loading.
Summary
The unit risk values used by MMS in the oil spill risk
analysis are summarized in Table 2-6.
Verification of Spill Risk Estimates
One of the major objectives of this study has been to
determine whether the oil spill risk analysis is reasonable and
adequate for impact assessment. Two possible ways of achieving
this are: 1) testing the projected values against the experience
record for a comparable oil field, or 2) examining the effects of
21
-------
Table 2-6. MMS-Calculated Expected Number of Spills
per Billion Barrels
21,000 bbl 2100,000 bbl
SOURCE: DOI 1985
22
Platforms 1.0 0.036
Pipelines 1.6 0.065
Tankers
At Sea 0.9 0.190
Per Port Call 0.2 0.042
-------
alternative sets of assumptions or exposure indices on the
projected values.
The FEIS for Sale 89 (DOI 1985) compares the Alaskan spill
record with the expected number of spills generated by the oil
spill risk analysis. Oil production began in Cook Inlet in 1964.
By 1980 about 0.7 Bbbl of oil had been produced. Spill
statistics for 1965-1980 are shown in Table 2-7. In Cook Inlet,
there were 3 years with total spillage in excess of1,000 bbl:
1966f 1967, and 1968. These spill volumes resulted from 9 spills
in 1966, 9 spills in 1967, and 32 spills in 1968 (DOI 1984).
Included in data for these 3 high-spill years were two tanker
spills of >1,000 bbl and one pipeline spill of >1,000 bbl. These
numbers correspond to pipeline spill rates of 1.42/Bbbl and total
transportation spill rates of 2.84/Bbbl if the data are not
adjusted for industry improvements.
The following paragraphs are quoted from the Lease Sale 89
FEIS (DOI 1985, p. IV-8). The projected values in the quoted
material apparently have been adjusted for improvement in spill
rates.
"Because DCS statistics are compiled as 'number of spills
per volume produced,1 the only comparisons of DCS statistics with
Alaskan data are for the state-leased offshore Cook Inlet and
Prudhoe Bay/Kuparuk fields. Based on OCS spill statistics, and
assuming that Alaska also experienced the post-1974 improvement
in platform and tanker (not pipeline) performance seen in OCS
statistics, the number of spills which would be projected for the
Cook Inlet and Prudhoe Bay/Kuparuk fields are shown below:
Number of 1,000-Barrel or Greater Spills
(through August 1983)
Cook Inlet Prudhoe Bay/Kuparuk
Projected Observed Projected Observed
Platforms 1.79 0 3.0 1- 3(*)
Pipelines 1.28 2(**) 4.8 6
Tankers 2.06 2 3.9 1
Total 5.13 4 11.7 10
(*) The 3 includes two airfield spills.
(**) From Gulf Research and Development Company 1982.
[Note: The information cited at this point in the FEIS includes
spill statistics through 1983. Table 2-7, also from the
FEIS, includes data only through 1980.]
SOURCE: MMS, Alaska OCS Region, 1984.
23
-------
Table 2-7. Cook Inlet Spill Data
YEAR
PRODUCTION
(Mbbl)
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
0.03
2.65
15.9
52.5
60.9
70.1
66.2
63.7
61.7
59.9
60.0
54.5
49.8
45.0
38.4
32.3
SPILLAGE3
(bbl)
87
2467
1982
2278
246
28
75
22
131
150
23
76
10
14
4
3
a Spills of known volume.
SOURCE: DOI 1984
NO. OF
SPILLSa
1
9
9
32
12
9
10
11
12
25
13
15
12
9
5
3
24
-------
"For Cook Inlet, the probability (0) of observing platform
spills is 17 percent. The above calculations indicate that we
would have projected 5.1 spills to occur as a result of
production and transportation of oil in Cook Inlet. In fact,
4 spills were observed. The probability of observing only
0 to 4 spills overall is 42 percent, almost an even chance.
Thus, the OCS oil-spill-occurrence statistics applied to Cook
Inlet production shows a reasonable agreement with the observed
number of spills.
"The OCS statistics project 11.7 spills for Prudhoe
Bay/Kuparuk production and transportation; we observed 10. OCS
statistics projected 3 platform spills, and 1 to 3 spills
(depending upon inclusion of airfield spills) were observed. We
projected 4.8 pipeline spills; we observed 6. We projected
3.9 tanker spills; we observed 1. The probability of observing
0 to 1 platform spills is 20 percent, and the probability of
observing 1 to 3 platform spills is 60 percent. The probability
of observing 5 to 7 pipeline spills is 41 percent. The
probability of observing 0 to 2 tanker spills is 25 percent. Th(
probability of observing 0 to 10 spills overall is 38 percent.
"In conclusion, Alaska has not produced enough oil to ,
statistically demonstrate that it has a different spill rate than
the rest of the OCS."
The comparison between OCS statistics and the Prudhoe
Bay/Kuparuk fields illustrates one of the difficulties in
extrapolating data. Of eight pipeline spills (21,000 bbl) in the
Gulf of Mexico, five were the result of anchor dragging. Anchor
dragging is not a significant problem for the trans-Alaska
Prudhoe Bay/Kuparuk pipelines. Since anchor dragging is not a
possible cause of pipeline spills in the Prudhoe Bay/Kuparuk
fields, it is misleading to use OCS statistics as was done above.
If the projected pipeline spill rate is adjusted downwards by
5/8, then 1.8 spills would have been predicted instead of 4.8 as
listed in the FEIS. This would suggest that the observed
pipeline spill rate is much more than the projected rate.
However, even this manipulation is challengeable because the
Prudhoe Bay and Kuparuk fields are not offshore environments.
Cook Inlet, although a harsh environment for ice and tides,
is not an open ocean environment. It does not have the same risk
of high wave conditions as the Bering Sea OCS nor does it have
the shipping and fishing traffic of the Gulf of Mexico. Apart
from these two important differences, use of Cook Inlet data
would be reasonable as a rough approximation of the adequacy of
the projected values.
Given these considerations, use of Alaskan data to test the
applicability of U. S. OCS data should be done with care.
25
-------
Alternative Approaches
Apart from using different models for U. S. DCS production
(e.g., the recent North Sea spill record), probably the only
alternative approach to the current oil spill risk analysis is to
re-examine the selection of exposure indices. As noted earlier
in this chapter, this would entail a more critical examination of
the causes of oil spills and less emphasis on the volume of oil
produced. The way to achieve this is to identify historical data
with regard to cause, i.e., to determine if engineering
improvements have lessened! the probabilities of certain types of
accidents, to determine relationships between certain types of
accidents and the environment, and to relate these findings to
developments in the proposed lease area. Accomplishing this
analysis might require use of the data for spills of all sizes,
although the final focus of the subsequent impact assessment
would be only on the larger spills. Appendix B has been prepared
for the purpose of allowing the reader to visually examine the
spill record and relate spill data to different factors.
The primary difficulty with applying a single concept to a
risk analysis situation is that confidence limits cannot be
derived. One way to achieve a measure of confidence is to
consider different models based on different assumptions and then
compare risks. If alternative approaches have no major effects
on the expected number of spills, then it can be concluded that
the spill risk estimates are reasonably accurate. This section
is prepared as a preliminary effort to recommend different models
or approaches. It is not presented to prove or disprove a risk
model.
The North Sea; A High Latitude Oil Field Model
The North Sea oilfield is particularly attractive as an
alternative model because the marine environment is similarly
harsh, and it is a new production field that underwent rapid
development. Use of the model assumes that industry technology
and practices are not likely to be significantly different from
U. S. DCS practices. Unfortunately, the data were not available
for this review. The Futures Group obtained 3 years of North Sea
data, which proved to be insufficient for analysis. The main
deterrent for further examination of North Sea data appears to be
its high cost from Lloyds (Prentki pers. comm.).
Alternative Exposure Indices
One could also develop models that use more than one
exposure index as the predictor of spill frequency. This
approach would allow for more flexibility in defining senarios
for oil field development. The following is a brief list of the
types of variables that may be important in determining the
exposure index.
Production. This could be further refined to define the
number and types of tankers, production platforms, and pipelines.
26
-------
Changes in Production. This could be an important factor in
determining risk, particularly in a region where rapid changes in
production are anticipated. Spill data have been graphically
related to rate of change in Appendix B. These figures suggest
rate of change may have some influence on spill rate, but the
necessary statistical evaluation has not yet been done.
Size of Platforms. The database has not been examined for
the relative safety records of small vs. large platforms.
Seasons. Risks to mammals, birds, crabs, and migrating fish
may have a seasonal dependence. Smith et al. (1982) state that
such factors can be put into the MMS model using trajectory data.
It may be worthwhile to examine effects of season on spill
frequency.
Infrequent Severe Storms. No analysis has yet been done to
ascertain the influence of severe storms. If platforms are
designed for 50-year storms, what is the probability of the
100-year storm striking during the lifetime of the oil field?
Geological Features. Oil production in highly fractured
areas is more hazardous than in more stable areas.
Size of Holding Tanks. The risk of a large spill could
increase with increases in average container size.
Size and Frequency of Tanker Traffic. Spills of all sizes
from U. S. flag tankers have been shown (Stewart and Kennedy
1978) to occur at a rate of about one per 3 tanker-years. Thus,
an alternative exposure index could be compared with the current
values in the oil spill risk analysis.
Pipelines. Here there are variables such as length, size,
ability to quickly detect and respond to problems, and frequency
of anchoring along the pipeline route.
Other Activities. The interactions between other activities
(e.g., a large commercial fishery) and oil production activities
could change risk estimates. In some cases, these may be closely
linked to season.
Incorporating an alternative or even a multivariate exposure
index into a risk model would require changes to the present risk
analysis and investigation into a number of scenarios. The
assumptions operating under the current approach are not
demonstrably more reasonable than the alternatives. However, the
outcome of such analyses may serve no purpose other than to
provide analysts with better information on the causes of spills.
The inherent nature of probability functions means that
uncertainty will occur about the expected number of spills
irrespective of which assumptions are used. It is doubtful that
the use of a more accurate exposure variate, for example, will
result in greater confidence in the expected number of spills
because the confidence interval is so large.
27
-------
Implications for Impact Assessment
The foregoing discussion indicates that spill risk
predictions may be sensitive to different sets of assumptions
used to calculate the risk and that there is no compelling reason
for accepting any particular set of assumptions over another.
The use of a "superior" exposure index may change the expected
mean number of spills which, in turn, may significantly alter the
probability function. However, decision makers and the public
should recognize that the inherent uncertainties in the process
are such that changes of this scale are unlikely to substantively
reduce the uncertainty. What is more important is recognition of
how the uncertainty might affect interpretation of the impact
assessment.
Marine birds and mammals are perhaps most susceptible to the
effects of oil spills because these animals have frequent and
regular exposure to the water surface. Table 2-8 summarizes
findings of the impact assessment on these organisms as reported
in the FEIS (DOI 1985) or readily derived from the assessment.
The four whale species included in Table 2-8 have been selected
because they are known to occur regularly and in significant
numbers in the St. George lease sale (Jones & 'Stokes Associates
1984), and they are designated as endangered species. Other
endangered cetaceans are either not likely to occur or are so few
in number that the probability of contact with an oil spill is
particularly remote.
The table shows the MMS-calculated combined probability that
a spill will occur and contact the Pribilof Islands area, the
Unimak Pass area, or the shelfbreak south of the Pribilofs
(Resource Areas 10 and 11 in the EIS analysis). These numbers
can then be compared to risk criteria established by MMS for
marine birds and gray whales. (It is not clear why the criteria
are different. One could interpret "low" and "unlikely" to be
similar categories, but MMS assumes that a "medium" risk for
birds is a probability of 11-25 percent and an "unlikely" risk
for gray whales is a probability of 11-30 percent. The important
point is that the two sets of numbers for resource areas and
species can be compared.) Clearly, the risk to birds and mammals
is high in the vicinity of the Pribilof Islands and borders on
high in areas along the shelfbreak south of the Pribilofs. Given
the uncertainties involved in the estimated expected number of
spills, the combined probability may have a confidence range that
is wide enough to include more than one of the probability risk
criteria.
This conclusion is particularly important when evaluating
the summary of the FEIS and comparing it to the supporting
documentation. The impact rating categories are clearly defined
in the EIS, but it is not immediately clear that the rating in
the summary integrates impacts from all activities (e.g., noise
disturbance and oil spills). Using the definitions established
by MMS and examining only conditional probabilities as discussed
in the EIS, significantly different categories are noted even
28
-------
Table 2-8. Reported Impact Assessment for Marine Birds and Marine Mammals
PROBABILITY RISK CRITERIA*1
COMBINED MEDIUM HIGH
PROBABILITY8 . LOW (UNLIKELY) (LIKELY)
* %
NJ
VO
RESOURCE/
RESOURCE AREA
Prlbilof Islands
Unimak Pass
Shelfbreak south of Islands
Marine birds
Marine mammals6
Gray whale
Fin whale
Humpback whale
Sperm whale
33-34
13
5- 9
44-46
15
14-34
0-10
11-25
11-30
26-50
31-60
DEIS
IMPACT
RATINGC
Mod.-Kaj.
Mod.
Minor
Minor
Minor
Negli.
LARGE
SPILL
EFFECTS0
Major
Major
Major
Mod.
Mod.
Mod.
From Table E-10, probability a spill occurs from proposed action And hits target in "x" days.
Averaged over ail modeled launch (feints. Does not include cumulative case.
Qualitative scale used in discussion of marine birds (DOI 1985, p. IV-46) and gray whales
(in parens) (DOI 1985, p. IV-72) .
From Table S-l, integrates all activities (e.g., noise disturbance and oil spills) and classified
as major, moderate, minor, negligible.
Conditional probability, i.e., assumes a large spill occurs during summer in the vicinity of a
species-specific high use area in lease sale.
Noncetacean species.
-------
when it is recognized that the reported conditional probabilities
are annual means of the trajectories.
30
-------
Chapter 3
OIL SPILL TRAJECTORIES
Summary
Winds are more important than ocean currents in determining
spill trajectories. This chapter begins, therefore, with a brief
discussion of the general climatology of the Bering Sea and how
wind observations at the Pribilof Islands relate to the
climatology. The purpose of the discussion is to describe major
features of the weather that must be simulated if the trajectory
analysis is to be adequate. It is not possible to discern from
the EIS or its supporting documents what meteorological data are
used or how they are used. Personal communication with Liu and
Leendertse was necessary to determine that the information used
in the model to simulate winds could result in reasonable
trajectory simulations. The numerical model's ability to portray
ocean circulation also appears reasonable, although apparently
minor concerns remain about the model's "as-run" ability to
accomodate monthly changes in densitydriven currents.
Significant concerns remain about the adequacy of the number
of trajectories currently used in the oil spill risk analysis.
Evaluation of the risk assessment in the St. George Basin Sale 89
EIS suggests that the risk to targets during summer periods may
be under-represented. Over half (58 percent) of the trajectory
simulations are run for winter conditions, which may account for
the apparent under-representation. It is not clear, however,
whether under-representation also results from the manner in
which overall risk is calculated or from the error inherent in
the trajectory simulations.
Overview
Trajectory Evaluation
An important fact that stands out when one attempts to
predict oil spill risks for a proposed OCS lease area is that the
problem is fundamentally probabilistic. A great deal of
uncertainty exists not only with regard to the location, number,
and size of spills that will occur during the course of
development, but also with regard to wind and current conditions
that give direction to the oil at the particular times that
spills occur. While some of the uncertainty reflects incomplete
or imperfect data for which it is difficult to assign error
bounds, the trajectory should be amenable to error analysis.
31
-------
The purpose of the oil spill trajectory studies done for
EISs is to assess the probabilities that oil spills from
locations within a proposed lease area will reach specific
targets. To accomplish this goal, a simulation is made of a
number of trajectories of oil spills and then these trajectories
are treated statistically. Oil spill trajectories in the
Beaufort and Bering Sea regions are computed using a numerical
model developed by Liu and Leendertse (1982) for the Rand
Corporation. The model consists of two major features: one
dealing with meteorological conditions, and the other with ocean
circulation. The trajectories are then inserted into the MMS
OSTA model (which is described by Smith et al. 1982) to evaluate
the conditional probability of oil reaching the targets. The use
of the Liu and Leendertse numerical model to predict trajectories
is unique to Beaufort and Bering Sea EIS documents.
General Climatology of the Bering Sea
Movement of oil on the sea surface is directed by local wind
and wave conditions and ocean currents. Should ice be present,
it would modify the response of oil to the appl ied forces. The
most important element in determining the trajectory of an oil
spill, in both cases, is the wind.
The Bering Sea is affected by arctic, continental, and
maritime air masses. In summer, the entire region is normally
under the influence of maritime air from the Pacific. The
southern-portion of the Bering Sea is most frequently under the
influence of maritime air, except during January and February
(Grubbs and McCollum 1968) when normally a strong flow of air
from the north and east brings in continental and arctic air.
For the remainder of the year, the movement of low-pressure
centers and associated winds dominate the a-tmospheric
circulation in the southern Bering Sea.
A major influence on the general atmospheric circulation in
the area is the region of low pressure normally located in the
vicinity of the Aleutian Chain, referred to as the Aleutian Low.
On monthly mean-pressure charts, this appears as a low-pressure
cell normally oriented with the major axis in an east-west
direction. This is a statistical low, indicating only that
pressures are generally lower along the major axis as a result of
the passage of low-pressure centers or storms. Storms are most
frequent and more intense in this area than in adjacent regions.
The most frequent track or trajectory of movement of these storms
is along the Aleutian Islands and into the Gulf of Alaska in
winter, and along the same general path in the west, but curving
northward into the Bering Sea in summer (Overland 1981). The
monthly frequency of low-pressure centers in the southern Bering
Sea is slightly higher in winter (generally four-five) than in
summer (three-four). Winter storms are much more intense than
summer stoxms.
In winter, the most frequent airflow is northeasterly around
the northern side of the low-pressure cell that is present at
32
-------
some location along the Aleutian Chain. In summer, with the
movement of lows into the Bering Seaf a more southwesterly mean
flow develops over the lower two-thirds of the region.
Climatology of the southern Bering Sea is characterized by a
progression of storms rather than fixed weather types (Overland
1981; Overland and Pease 1982). These storms produce increased
cloudiness, reduced diurnal temperature range, and winds that
rotate through the compass. During the summer in the southern
Bering Sea, frontal activity can be severe as very cold arctic or
continental air comes in contact with the warm air from the
Pacific Ocean, forming a sharp discontinuity and localized winds.
Figure 3-1 shows wind roses for selected locations and
marine regions in the Bering Sea during February and August. The
wind roses show the percentage of observations from each of eight
possible directions. (The data were from Brower et al. [1977]
and Grubbs and McCollum [1968] and compiled by Overland [1981].)
In winter ("February" in Figure 3-1), the northern stations
show a high percentage of winds >17 kn from the north and
northeast, whereas the winds over Bristol Bay (Marine Area C) are
uniformly distributed over direction with moderately high speeds.
This is indicative of a fairly continuous progression of storms
through the area. Wind speeds over the Bering Sea in summer
("August" in Figure 3-1) are generally lower than in winter
although conditions are seldom calm. Marine Area A to the north
shows little preferred direction, but the other stations show
predominance of south and southwest winds in the summer.
Simulation of Bering Sea Winds
This section concerns the procedures used by the Rand
Corporation to simulate winds for use in the oil spill trajectory
studies. Because the EIS and its support documents contain
insufficient detail to permit reconstruction of the technical
details, we met with Liu and Leendertse at Rand Corporation in
order to obtain information on the simulation procedures. The
discussion below is based on personal communications with Liu and
Leendertse but does not constitute a formal review of the Rand
programs. It should be noted that statements in the EIS and
support documents can be misleading; for example, they state that
Putnins1 (1966) study is used to model the winds. The data are
used but in a highly modified way; no mention is made of the
modification or its form in the EIS. Many of the questions
concerning the simulation of wind conditions that concerned this
review team were discussed at the meeting. Based on these
discussions, we conclude the approach taken by the Rand
Corporation could be capable of wind simulations adequate for
trajectory studies. Until a full documentation of the procedures
is provided for peer review, the simulations will remain a
concern to the scientific community.
Rand Corporation has identified 11 basic weather patterns
for the Bering Sea region. These weather patterns are similar to
33
-------
NORTHEAST CAPE
FEBRUARY
SPEED CLASSES
1-6 KN
7-16 KN ==
17- KN
o 10 10 so
-------
the more predominant patterns selected by Putnins (1966);
however, they are also based on more recent meteorological data.
The wind model uses these basic patterns and their frequencies
and patterns of occurrence to simulate the large scale
atmospheric features.
Since the Bering Sea region is one of frequent storm
activity, the Rand simulation procedure interrupts the large
scale weather patterns and inserts a traveling storm. For winter
months, the simulated storms occur at a rate of four-five per
month, and for summer months at a rate of three-four per month.
In addition, Rand has carried out a study of the intensity and
statistics of storm events in the Bering Sea region. These
statistics are incorporated in the simulation procedures.
The simulation procedure has been shown by Rand Corporation
(unpublished) to reproduce the observed wind roses at selected
locations in the Bering Sea region. In the DEIS for the St.
George Basin lease sale, it is stated that the wind model is
verified because it reproduces the average wind speed and
direction at Nome, Alaska. This is a gross over-simplification
of the work that has been done at Rand Corporation. The
simulation procedure contains more reality than is referenced in
the EISs for the Bering Sea lease areas. However, the simulation
procedure has not been documented nor has it been demonstrated
that the analysis reproduces both the spatial and time scales
which occur in actual Bering Sea winds.'
The question concerning spatial and time scales of simulated
winds is important to the trajectory analysis. When a trajectory
evaluation is made, trajectories from a number of launch points
are evaluated using the same simulated winds. If these winds
have larger spatial patterns than the launch point separations,
then adjacent trajectories will not be statistically independent.
Consequently, it is important that the scales of atmospheric
forcing be simulated correctly.
Simulation of Bering Sea Circulation
Oceanorah of the Berin
The reason that a numerical model of ocean circulation was
used for Bering Sea leases is that the region has strong tidal
currents, is subjected to strong wind events, and has domains in
which horizontal density gradients are large. Furthermore, the
shelf circulation has regimes which are distinctly different as
to the influences of bottom friction, thermohaline processes, and
oceanic influence. The inclusion of a dynamic model in the risk
analysis study is intended to increase the reliability of oil
spill trajectory estimates.
The Bering Sea continental shelf is broad (500 km) and the
bottom grades smoothly offshore to a relatively deep (170 m)
shelf break, with the 50 and 100 m isobaths dividing the shelf
35
-------
into three zones with distinguishable water column
characteristics. The 50 m contour lies 80-150-km from shore, and
the 100 m contour is 100-150 km landward of the shelf break,
leaving a central region over 200 km wide with intermediate
depths.
The dominant water motion on the shelf is by tidal currents,
which are relatively strong (20-50 cm/sec) and account for
60-90 percent of the horizontal kinetic energy. Turbulent energy
for mixing the water column comes from only two sources: tidal
currents (up from the bottom) and wind (down from the surface).
During winter and spring, the wind-mixed layer is 10-70 m thick,
with an average thickness of about 50 m. In summer, this layer
is 5-20 m thick, with an average value of about 10 m. The
tidally mixed bottom layer is 30-50 m thick. Over the inner
shelf (depths <50 m), these wind and tidal boundaries merge,
creating a well-mixed water column. Over the middle shelf
(depths 50-100 m), the boundary layers are separated. The layer
between has no significant source of turbulent mixing energy. It
is within this central part of the water column that shelfwater
is working its way seaward. Over the outer shelf, oceanic water
penetrates inward beneath the outflowing water from the middle
shelf.
The circulation described above is weak with horizontal
currents generally <4 cm/sec. Tidal currents are much stronger;
however, tidal motions are elliptical. It terms of transport of
oil over time spans of 3-30 days, neither the weak mean flows
associated with the distribution of temperature and salinity over
the Bering Sea shelf nor the tidal currents are of predominant
importance. The wind-driven currents dominate in the transport
of oil on the sea surface. The wind-driven currents fluctuate on
periods of 2-10 days, with magnitudes up to 20 cm/sec (Schumacher
pers. comm.).
In addition to the wind-driven circulation over the Bering
Sea shelf, there are strong currents along the continental shelf
between Unimak Pass and the Pribilof Islands. These currents
fluctuate at periods longer than 10 days, and the fluctuations
are not apparently related to local wind events. The average
speed of the currents is about 6 cm/sec; however, currents as
high as 20 cm/sec and as small as 2 cm/sec may be found
(Schumacher pers. comm.). The general direction of flow is along
the shelfbreak to the northwest. During the summer of 1977, six
satellite-tracked drifters (drogued at 17 m) were deployed over
the shelf/slope break. Vector mean speeds were 5-15 cm/sec over
deeper water and 1-3 cm/sec in the shallower shelf waters. The
buoys drifted towards the northwest.
The Numerical Model
The basic 3-dimensional model developed by Liu and
Leende-rtse solves the equations of motion for water and ice,
continuity, mass, heat, salt, pollutant, and turbulent energy
balance. Implicit numerical solution methods are used in the
36
-------
vertical so that cross-layer transfer of momentum, energy, and
constituents can be computed accurately without any numerical
stability problems. The model was first described by Leendertse
et al. (1973) and again by Leendertse and Liu (1977).
Improvements to the model which have been made since these
publications include:
The horizontal grid structure includes the ellipsoidal
curvature of the earth.
Ice dynamics (including melting, salt rejection, and
ice-ice interactions) are included in the formulation.
A parameterization of oil movement under ice is in the
formulation.
The model incorporates a closed form for the generation
and decay of turbulence. This form no longer depends
on the Richardson number and related parameters.
The model includes the kinetic energy content and
dissipation associated with short wind waves.
No substantial changes in the model have been made since 1981.
Consequently, all studies of oil spill trajectories have been
carried out with the same formalism.
The 3-dimensional model can be used directly for simulating
oil movements for a duration of several days. For longer periods
(such as several months), a much more economical method is
needed. To accomplish this, Rand uses a method called the unit
response function method. Response functions, after being
generated by the 3-dimensional model under four wind directions
(N,E,S,W), are used to synthesize, through a convolution process,
the drift currents due to winds from various weather scenarios.
Response functions are generated by the difference in currents in
the 3-dimensional field with and without the wind stress under
identical tide conditions. The model is run for 5 days; the
first 3 days of information are discarded because they are
contaminated by start-up transients in the model. The last
2 days of data are used to evaluate the response functions and
average flows.
Boundary conditions used in the simulations are based on
field observations. For the shelfbreak currents between Unimak
Pass and the Pribilof Islands, historical hydrographic data have
been used to determine the density field and geostrophic
currents. The model uses a shelfbreak current of 6 cm/sec
flowing towards the northwest. Over the shelf proper, the
initial data for the model's density field are based primarily on
observations gathered in 1976 by the NOAA research ships Mo ana
Wave and Miller Freeman. Tide data at the open boundaries, which
are based on observations, complete the required suite of
boundary conditions.
37
-------
Verification
Mofjeld (1984) has compared observations of tides with
predictions made by the 3-dimensional Liu and Leendertse model
and by a vertically integrated hydrodynamic model (Sundermann
1977). Mofjeld concludes that the quantitative agreement with
observations is better for the 3-dimensional model. He states,
"This may be due to the more complete dynamics in the
3-dimensional model as well as the tuning of this [3-dimensional]
model to a larger set of tide observations than was available to
Sundermann (1977)." The simulations were for summer conditions.
The comparison employed predictions from a preliminary
calculation using the 3-dimensional model; more recent work may
demonstrate further improvements. Even without further
improvements, the 3-dimensional model has been shown to predict a
reasonably accurate tide picture for the Bering Sea region. This
could not result from the calculations if the model was
incorrectly evaluating tidal dissipation.
Ice
The Rand model allows for ice cover in the oil spill
trajectory studies. Ice, particularly near the ice edge, adds
several complications to the prediction problem. Muench (1983)
and Muench and Schumacher (1985) discuss the results of recent
experiments. One important process is the melting of the ice,
which creates a fres.hwater lens near the limit of ice flow. The
density contrast between this fresh water and the more saline
waters of the ice-free regions of the continental shelf generates
a northward-flowing baroelinic current parallel to the edge of
the ice zone. This current, coupled with reduced mobility of sea
ice (relative to open water), diminishes the importance of local
winds relative to baroclinc flows in transporting oil.,
Trajectories of oil spills generated by the Rand model show this
effect.
An average ice cover year is assumed for the model runs.
The actual marginal ice zone is not fixed in space but varies
with wind conditions. It does not appear that the Rand model can
include this variability in the trajectory calculations. This is
because the calculations use response functions that are fixed by
the ocean state at the beginning of the simulation. Thus, the
variability in trajectory paths related to the position of the
marginal ice zone may be underestimated in the calculations.
The EIS and its supporting documents inadequately describe
the model run with ice cover. The documents state only that the
model was run with ice cover. In fact, the model 'with ice
cover1 refers to a model formulation capable of incorporating ice
cover. For the trajectory studies, winter conditions were
simulated by using six trajectory simulations for each of
6 winter months. For each simulation, the marginal ice zone was
located at its multi-year average location for the month.
38
-------
Influence of Numerical Model on Trajectory Analyses
A review of the trajectory analysis is severely hampered by
a lack of documentation of the details of the calculations. When
tests and checks have been made, they should be cited in the
references in the EIS. Because our major goal has been to
consider reasonableness of the approach, we have looked at the
results as the output of a 'black box1 and tried to understand
its significance in predicting conditional risk percentages. It
is not necessary that details of the calculations or formulations
be described in the EIS. It is necessary, however, that the EIS
provide enough information about what data are used in model
formulation and an overview of how they are used so that it is
possible to determine whether important factors in trajectory
simulations are accounted for. Our review suggest that the model
and resulting trajectory analyses are reasonable, although two
concerns remain: first, regarding the model's ability to address
short-term (monthly) changes in baroclinic currents; and second,
regarding the model's ability to address effects of freshwater
runoff.
Baroclinic Currents
The evaluation of trajectories of oil spills relies heavily
on the use of response functions derived from the continental
shelf circulation model. These response functions are derived
from model runs a few days long (after allowing for startup
transients in the calculations). The use of these response
functions implies that they contain all of the relevant physics
and dynamics of the ocean circulation.
The purpose of a multi-layered numerical model is to include
baroclinic processes in the calculations. Since the response
functions are fixed in space and derived from the density field
imposed when the model run is begun, any effects due to short-
term advection of the density field are lost. If wind forcing
moves the density field or alters horizontal density gradients,
the resulting changes in circulation cannot be evaluated. New
response functions would need to be derived based on the new
density field, and this step would negate the numerical economy
achieved by using response functions.
Since baroclinc currents are weak ซ10 cm/sec) over most of
the Bering Sea shelf, the errors made in their evaluation may not
seriously hamper the trajectory evaluations. If the model were
extended to oil in water calculations, however, errors in
evaluating baroclinc flows could become more important.
Freshwater Runoff
If fresh water were to enter the system, e.g., from major
rivers such as the Yukon, the influence of this lens of water is
not accounted for unless it is present in the response functions.
Runoff can vary appreciably over periods of 30 days, particularly
in spring time. This temporal variation of input of fresh water
39
-------
would not be included in the response functions, and the
resulting calculations could be in error. Since effects of
runoff are localized, the errors need not have a major impact on
the shelf as a whole, but they could appreciably alter local
calculations in areas such as Norton Sound or inner Bristol Bay.
These comments are not to be interpreted as a statement that
the Rand model is incapable of evaluating baroclinic flows or
freshwater runoff. We note that, as used in the trajectory
analysis, the model cannot properly reproduce short-term changes
in baroclinic flows. Regions where such flows are important
include: the marginal ice zone, regions with high inputs of
fresh water, regions with time-variable input of saline water
(the most predominant is Unimak Pass), and the region near the
50 m contour.
Interpretation of the Trajectory Studies
Figure 3-2 was prepared from the risk tables in the FEIS for
the St. George Basin lease sale (DOI 1985, Table E-3). The
launch points in the lease sale area are labeled in Figure 3-3
using the notation in Table E-3. Figure 3-2 shows the overall
combined risk probabilities for the resource areas around St.
George and St. Paul Islands (Figure 3-4). The resource area is
defined by a circle of radius 50 km about each island.
Figure 3-2 shows that the majority of the risk comes from within
the quadrant to the east of the islands. Although risk to St.
George Island is minimal from the south and west, similar
findings are not made for St. Paul Island. Overall, the patterns
of probability are not inconsistent with wind rose data for the
winter (Figure 3-1). In summer, however, the wind rose for St.
Paul Island shows more persistent winds from the south,
southwest, and west (Figure 3-1). These pictures raise a
question concerning the true risk to the Pribilofs because it
seems that summer weather is not adequately represented in the
statistics.
One way to interpret Figure 3-2 is to assume that
differences in the probabilities of contact to a target from
nearby locations represent the overall statistical accuracy in the
calculations. For example, in Figure 3-2 there are two nearby
launch points to the southeast of St. George Island: one has a
probability of 42 percent of contacting the St. George resource
area and the other a 30 percent chance. Since the locations are
nearly the same, it could be assumed that the internal
consistency in the calculations is about 12 percent. Similarly,
north of St. Paul Island, there is a launch point with a
14 percent probability. A launch point nearly an equivalent
distance north of St. George Island has a 0 percent chance of
resulting in a spill that reaches the St. George resource area.
This suggests an internal consistency in the simulation procedure
of about 10-15 percent.
40
-------
NOTE: VALUES INDICATE PERCENT PROBABILITY THAT A LARGE OIL SPILL OCCURRING AT THAT POINT WILL REACH
PRIBILOF ISLANDS WITHIN 30 DAYS.
FIGURE 3-2, MMS-CALCULATED RISK TO PRIBILOF ISLANDS RESOURCE AREAS
SOURCE: DO I 1985, TABLE E-3.
-------
A ST. GEORGE BASIN DEIS
P NAVARIN BASIN DEIS
FIGURE 3-3, LAUNCH POINTS USED IN MMS OIL SPILL RISK
ANALYSIS
SOURCE: DOI 1984, GRAPHIC 5,
42
-------
CO
BIOLOGICAL RESOURCE AREA
TRACTS WITH ACCEPTED BIDS (SALE 70)
FIGURE 3-4, BIOLOGICAL RESOURCE AREAS FOR ST, GEORGE BASIN OIL SPILL RISK ANALYSIS
SOURCE: DOI 1984, FIGURE E-2,
-------
For the St. George Basin EIS, simulated oil trajectories
were generated from 24 launch points in or near the lease sale
area. These were selected to be representative of platform
locations, pipeline routes, and tanker routes. In addition,
26 launch points used for the Navarin Basin EIS were included in
the statistics. Twenty-six trajectories were launched from each
point for the ice-free season. Thirty-six trajectories were
launched from each point during periods when there could be ice
cover. It is possible that the use of so few trajectories from
each launch point can explain the differences in probabilities
noted in the preceding paragraph.
Because the winds in winter do not tend to drive oil from
the south towards the Pribilof Islands, there is little risk to
these islands during this time. Thus, 58 percent of the
tractories (36 out of 62} have a negligible chance of reaching
the Islands from the south. This leaves a maximum chance left of
42 percent for the summer tractories. With weaker summer winds
and variability in these winds, not all summer trajectories
should move northward. If a quarter of them reached the Island
resource areas, the risk percentages would be about 10 percent.
If the inherent statisical variation in the trajectory
evaluations is on the order of 10-15 percent, then a 0 percent
chance is the same as a 10 percent chance. Proper decisions can
be made only when it is recognized that the analysis as currently
done has this level of uncertainty. Some of this uncertainty may
be eliminated if this question of the number of trajectories
needed is resolved.
Number of Trajectories
The risk anlaysis calculations for the Bering Sea lease
areas use the spill trajectories provided by Rand Corporation to
evaluate the probability that a target area is hit by a spill
from a given source. This probability is defined as the number
of hits divided by the number of spills expressed as a
percentage. Once an oil spill trajectory hits land, it is no
longer counted in the calculations. This raises the question of
how many trajectories should be calculated for a specified source
point and how many source points should be evaluated.
Samuels (1984) documents the trajectory analysis for the
St. George Basin lease sale and includes a table of errors and
number of trials in Monte Carlo simulations of a binomial
process. A binomial process is the equivalent of a biased coin
toss problem. Table 3-1 summarizes some statistics from Samuels'
table. The table gives the error estimates in estimating the
true probabilities given the number of trials. The numbers in
Table 3-1 are re-examined in Table 3-2 in terms of percentage of
error and relates it to a biased/unbiased coin toss problem.
When the table is expressed in this way, it is apparent that many
experiments are needed to determine the bias in an extremely
biased coin, while relatively many fewer are needed to determine
the probabilities in an unbiased coin.
44
-------
Table 3-1. Monte Carlo Error as a Function of Number of
Trials and Estimated Probability
Number of Trials
Prob
0.02
0.50
SOURCE :
0
0
UL
.07
.26
Samuels
Afl.
0.04
0.13
1984,
0
0
Tabl
5JI
.03
.12
e 2.
100
0.02
0.08
500
0.01
0.04
1000
0
0
.01
.03
2000
0
0
.01
.02
Table 3-2. Percent Error for a Biased (P=0.02) and
Unbiased (P=0.50) Coin.
Number of Trials
Prob 10_ 4J1 5_ฃ 100 500 1000 2000
0.02 350 200 150 100 50 50 50
0.50 52 26 24 16 8 6 4
Note: Derived from Table 3-1.
45
-------
Applying this model to the true probability that oil spills
might strike a specific land segment vs. the calculated
probabilities that oil strikes the land segment, we observe that
many trials are needed to adequately model low risk areas and few
trials are needed to model high risk areas. We note, however,
that the percentage of error is large even with 2,000 samples of
a coin with extreme bias. Table 3-2 suggests that for the
trajectory calculations in which 62 trials are used, there should
be inherent errors between about 24 and 150 percent.
An alternative method for viewing trajectory experiments
can be expressed by the following thought experiment, which is
geometrically similar to the trajectory problem. Suppose we
randomly generate 36 spokes radiating from the center of a circle
that has been divided into 36 arcs each of 10ฐ width. For any
arc, the mean expected number of spokes hitting the arc will be
one; however, the standard deviation in the number of spokes in a
given arc is four. Thus, in a specific experiment, it would not
be unusual to have several arcs with zero spokes and a few arcs
with four-five spokes. In terms of risk analysis made utilizing
a single sample, one arc would have four-five times the risk of
others, even though its true risk is identical to other arcs. In
the models of trajectories, the oceanographic and meteorologic
influences will tend to focus trajectories. Thus, statistics
from thought experiments, such as this, are" conservative.
Our concern with regard to the risk analysis presented for
Bering Sea lease areas is that the analysis of the trajectories
not introduce a bias and that they are sufficient in number to
eliminate statistical variations in the assessment of risk. The
documents supporting the EIS have failed to address this
question.
During the meeting with Rand Corporation, Liu presented an
overhead slide showing that independent calculations of
trajectories by Overland (PMEL unpublished), which were based on
actual weather sequences, had the same qualitative character as
Rand predicitions. The predictions were for an open ocean region
south of the Pribilofs. Oneway to look at this result is that
the Rand model is not predicting significantly different results
from what a 3.5 percent of surface wind-speed model would give.
In the Rand model, a 1.6 percent of wind-speed rule is used to
account for Stokes drift due to surface waves. The remaining
difference of 2.9 percent is included in the air-sea interaction
simulations.
The approach to the trajectory analysis appears to be
reasonable. Although ambiguities in the risk analysis procedure
appear to be of little importance as the model is now used, we do
not believe that these ambiguities should continue. The most
serious concern is that the risk to specified resource areas is
not adequately represented because the number of trajectories is
too few and perhaps biased toward winter conditions. Whether
this is the case cannot be determined until additional
trajectories are run so that error bounds can be evaluated.
46
-------
Chapter 4
CONCLUSIONS AND RECOMMENDATIONS
Overview
Our review and evaluation of the oil spill risk analysis as
presented by the St. George Lease Sale 89 EIS (DOI 1984, 1985)
and its support documents lead us to the following major
conclusions:
There is considerable uncertainty in the spill risk
estimate (expected number of spills) .
Given the inherent uncertainty in the estimate, the
approach and the reported estimates appear to be
reasonable approximations.
The wind simulation technique appears to be reasonable
and adequate for offshore lease sales in the Bering
Sea but has not been documented.
It is not clear that the number of simulated
trajectories is adequate to portray the risk to
designated targets, especially once the probability
distribution for trajectories is adjusted by mean-case
production data to generate combined probabilities.
The presentation of the oil spill risk analysis in the
EIS can and should be improved in several significant
ways if it is to meet NEPA requirements.
Sources of Uncertainty
There are four main sources of uncertainty in the risk
evaluations. These are:
1) Uncertainties in the resource estimates.
2) Uncertainties in the expected number of spills because
of the small database from which means have been
derived.
3) The distribution function for spill frequencies is only
approximated by a Poisson distribution. Therefore,
estimates of the probabilities of number of spills will
be increasingly uncertain as the number differs from
47
-------
the mean estimate. Similarly, spill size forecasts
will increase in uncertainty as the forecast departs
from the mean spill size.
4) The scenarios for production introduce additional
uncertainties.
In the DEIS for Lease Sale 89 (DOI 1984), the mean resource
estimate was 0.66 Bbbl. In the FEIS for Lease Sale 89, the mean
resource estimate increased to 1.12 Bbbl. In the 1981 synthesis
meeting, the mean resource was listed as 1.12 Bbbl. Finally in
the Federal Offshore Statistics (Essertier 1984, p. 98), the mean
resource estimate for St. George Basin is listed at 1.4 Bbbl.
The final risk estimates are linearly related to these resource
estimates. Thus, the final risks have inherent in them the
uncertainties in resource estimates.
For spills ฃ.100,000 bbl, spill rates are much less accurate
and much less deterministic than for smaller spills. This is
unfortunate because these very large spills are more likely to
have major and long term environmental consequences. The
projected spill rates are based on an extrapolation of a log-
normal distribution fit to the frequency distribution for spill
volume. Extrapolation of a log-normal distribution beyond the
set of date it was fit to is usually not attempted because almost
no confidence can be placed on the extrapolated estimate. The
FEIS (DOI 1985) presents the risk rate for very large spills as
if it had the same certainty as smaller spills. This is
misleading, particularly when one considers that the spill rate
for even smaller spills are characterized by great uncertainty.
Interpretation of the Oil Spill Risk Analysis
On several points, we conclude that the EIS interpretation
of the oil spill risk analysis requires improvement. The
majority of these points focus on two major concerns: 1) the
tendency to average out findings that either should not be or
cannot be averaged, and 2) incomplete presentation of information
necessary to judge the reasonableness and adequacy of the
findings.
By averaging the impact levels from various activities
associated with the lease sale, the significance of the impact
from oil spills is often obscured. This problem is most
noticable in the EIS summary. For example, although the EIS may
determine that impacts from seismic exploration activity may be
negligible and impacts from oil spills may be moderate for a
certain species, the EIS summary concludes that the "overall
risk" to the species is minor. This approach violates a
fundamental principle of ecology which states that the presence
and success of an organism or group of organisms depends on that
one condition that most closely approaches or exceeds the limits
of tolerance. Thus, the "overall impact" of a project is no less
than the most severe impact likely from any particular aspect of
48
-------
the lease sale activity. In a similar vein, "overall risk" to a
higher taxon (e.g., crustaceans or marine mammals) cannot be the
average of the risk to each species.
The EIS also averages out seasonal differences in trajectory
simulations. Averaging may be a reasonable approach in
calculating combined probabilities as long as this point is
clearly stated and seasonal differences are given weight
proportional to their occurrence. This latter point is
particularly important because of significant summer use of the
area by marine birds and mammals. Since 58 percent of the
trajectory simulations for Sale 89 are run for winter conditions,
a bias away from summer trajectories is incorporated in the
analysis. Thus, the conditional probabilities and combined
probabilities reported in the EIS are averaged over the annual
meteorological and oceanographic conditions and weighted more
toward winter conditions.
References to incomplete documentation have been noted
several times in previous chapters. In several cases, the
missing information is important to interpreting the
reasonableness and adequacy of the impact assessment. The most
important omissions are:
the general features of the wind simulation model;
a clear statement of the important assumptions
underlying the estimated expected number of spills;
an explanation of how the launch points have been
weighted for volume of oil handled; and
clear statements regarding the implications of the way
certain information is treated.
This latter point is particularly important. CEQ regulations
require a worst-case analysis when great uncertainty is inherent
in the analysis (40 CFR 1502.22). We believe it is reasonable to
present conditional probabilities that have been averaged out for
all seasons, for example, but the decision maker and the public
must recognize that the conditional probabilities in reality will
be higher or lower for certain seasons. It is the responsibility
of the EIS to clearly state implications such as these.
Recommendat ions
We recommend that EPA give greater emphasis to conditional
probabilities when determining the environmental consequences of
a proposed action. Greater reliance on the conditional
probabilities is, of course, predicated on: 1) a larger number
of trajectories per launch point, 2) proper balance between
seasonal conditions, and 3) clear understanding that these values
are averaged for the life of the project. The combined
probabilities are of value in that they provide a rough estimate
49
-------
of the likelihood of a particular spill event/impact, but they
should not be used in a deterministic manner because of the large
uncertainties in the expected number of spills. We believe our
recommended approach is more appropriate because it more closely
approximates a reasonable worst-case analysis, which is required
by CEQ regulations when great uncertainty is involved.
An adequate worst-case analysis should include a discussion
of conditional probabilities for the summer season. This
represents the period of time when most birds and mammals are at
greatest risk. Basing the analysis on conditional probabilities
that have been averaged out over the year would not result in a
reasonable worst-case analysis.
MMS should not continue to report "overall risk" to a
species as an average of the risks posed by different aspects of
the lease sale action. The practice has no biological value and
is misleading.
Sufficient numbers of trajectories should be simulated so
that error bounds on the probabilities can be assessed. We also
recommend that the number of trajectories be proportionately
distributed among the seasons if annual averages are going to be
used.
The EIS must do a better job of describing the Rand model.
We do not recommend that MMS describe the mathematical details of
the trajectory simulation models, rather, it is recommended that
a brief explanation be given of the factors considered in model
runs. For example, the reader must know: traveling storms are
simulated at a frequency that resembles actual conditions in the
Bering Sea; how the number of trajectories are allocated to
different months for the winter simulation; where the marginal
ice zone was for each winter simulation. This level of
information is necessary to judge the adequacy and reasonableness
of the risk assessment. We suggest that Rand publish a document
that concisely states the model's adaptation to the trajectory
studies, its meteorological inputs, and the tests that have been
made of the model. We also suggest a careful analysis be
conducted of the error bounds associated with the trajectories.
We recommend that the critical assumptions of the oil spill
risk analysis be clearly stated. In most cases, MMS has stated
the assumptions at various points in the EIS. However, we
believe the presentation can be significantly improved if these
are brought together in one location to clearly reveal to the
reader the information necessary to judge the credibility of the
analysis. As an example, we believe Table S-2 (DOI 1985) is a
commendable example of how important assumptions can be presented
to the reader. In this context, the reader can judge for himself
whether the terms as applied to endangered species (for example)
are appropriate and should be equivalent to those applied to
nonendangered cetaceans.
50
-------
REFERENCES
Literature Cited
Brower, W. A. Jr, H. F. Diaz, A. S. Prechtel, H. W. Searby, and
J. L. Wise. 1977. Climatic atlas of the outer continental
shelf waters and coastal regions of Alaska. AEIDC/NOAA,
Anchorage, AK.
Department of Interior. 1984. Draft environmental impact
statement. Proposed outer continental shelf oil and gas lease
sale, St. George Basin (Sale 89). MMS, Anchorage, AK.
1985. Final environmental impact statement. St.
George Basin Sale 89. MMS, Anchorage, AK.
Devanney, J. W. Ill, and R. Stewart. 1974. Analysis of oilspill
statistics, April 1974. Report No. MITSG-74-20. Prepared for
the Council on Environmental Quality. MIT, Cambridge, MA.
Essertier, E. P. 1984. Federal offshore statistics. OCS Report
MMS 84-0071. DOI, Washington, D.C. 123 pp.
The Futures Group and World Information System. 1982. Final
technical report: Outer continental shelf oilspill probability
assessment. Vol 1: Data collection report. 69 pp. Vol 2:
Data analysis report. Prepared for DOI/BLM under contract
No. AA851-CTO-69. The Futures Group, Glastonbury, CN.
Goldberg, N. N., V. F. Keith, R. F. Willis, N. F. Meade, and R.
C. Anderson. 1981. An analysis of tanker casualties for the
10-year period 1969-1978. Proceedings of the 1981 Oil Spill
Conference. Am. Petroleum Inst., Washington, D.C.
Grubbs, B. E., and R. D. McCollum Jr. 1968. A climatology guide
to Alaskan weather. Scientific Services Section, llth Weather
Squadron, Elmendorf AFB, Alaska.
Jones & Stokes Associates, Inc. 1984. Preliminary draft Ocean
Discharge Criteria Evaluation OCS Lease Sale 89 St. George
Basin. Prepared for U. S. EPA Region 10, Seattle, WA.
Lanf ear, K. J., and D. E. Amstutz. 1983. A re-examination of
occurrence rates for accidental oil spills on the U. S. outer
continental shelf. Pp. 355-359 In Proceedings of the 1983
Oilspill Conference. Am. Petroleum Inst., Washington, D.C.
51
-------
Leendertse, J. J.f R. C. Alexander, and S. K. Liu. 1973. A
three-dimensional model for estuaries and coastal seas. Vol.
1. Principles of computation. R-1417-OWRR. The Rand
Corp., Santa Monica, CA.
Leendertse, J. J., and S. K. Liu. 1977. A three-dimensional
model for estuaries and coastal seas. Vol IV: Turbulent
energy computation. R-2187-OWRT. The Rand Corp., Santa
Monica, CA. 59 pp.
Liu, S. K., and J. J. Leendertse. 1982. A three-dimensional
shelf model of the Bering and Chukchi Seas. Coastal
Engineering J. 18:598-616.
Mofjeld, H. 0. 1984. Recent observations of tides and tidal
currents from the northeastern Bering Sea shelf. NOAA Tech.
Memo. ERL PMEL-57. Seattle, WA. 36 pp.
Muench, R. D. 1983. Mesoscale oceanographic features associated
with the central Bering Sea ice edge, Feb-Mar 1981. J.
Geophys. Res. 80:4467-4476.
Muench, R. D., and J. D. Schumacher. 1985. On the Bering Sea
ice edge front. J. Geophys. Res. 90:3185-3198.
Nakassis, A. 1982. Has offshore oil production become safer?
USGS Open-File Report 82-232. 26 pp.
Overland, J. E. 1981. Marine climatology of the Bering Sea.
Pp. 15-22 In D. W. Hood and J. A. Calder, eds., The eastern
Bering Sea shelf: Oceanography and resources. NOAA/OMPA,
Seattle, WA.
Overland, J. E., and C. H. Pease. 1982. Cyclone climatology of
the Bering Sea and its relation to sea ice extent. Mon.
Weather Rev. 110:5-13.
Putnins, P. 1966. The sequences of baric weather patterns over
Alaska. In Studies on the meteorology of Alaska. DOC.
Prepared for Dept. of Army, IVO-14501-B53A-05-03.
Samuels, W. B., D. Hopkins, and K. J. Lanfear. 1981. An
oil spill risk analysis for the southern California (proposed
Sale 68) outer continental shelf lease area. USGS Open-File
Report No. 81-605. 206 pp.
Samuels, W. B. 1984. An oil spill risk analysis for the St.
George Basin (December, 1984) and North Aleutian (April, 1985)
outer continental shelf lease offerings. DCS Report 84-0004.
DOI/MMS.
Smith, R. A., J. R. Slack, T. Wyant, and K. J. Lanfear. 1982.
The oilspill risk analysis model of the U. S. Geological
Survey. Professional Paper 1227. USGS.
52
-------
Stewart, R. J. 1975. Oil spillage associated with the
development of offshore petroleum resources. In: Report to
the Organization for Economic Cooperation and Development. 49
pp.
1976. A survey and critical review of U. S. oil
spill data resources with application to the tanker/pipeline
controversy. Prepared for DOI. Martingale, Inc., Cambridge,
MA. 75 pp.
Stewart, R. J., and M. B. Kennedy. 1978. An analysis of U. S.
tanker and offshore petroleum production oil spillage through
1975. Office of Policy Analysis. DOI, Washington, D.C.
Sundermann, J. 1977. The semidiurnal principal lunar tide M2 in
the Bering Sea. Deutsche Hydro. Zeitschrift 30:91-101.
USGS. 1972. Outer continental shelf statistics 1953-1971.
Conservation Div.
. 1979a. Accidents connected with federal oil and
gas operations on the outer continental shelf, Gulf of Mexico.
Vol. I: 1959-1976. Conservation Div., Metairie, LA.
. 1979b. Accidents connected with federal oil and
gas operations on the outer continental shelf, Pacific area.
Conservation Div., Los Angeles, CA.
Wilson, E. B. 1952. An introduction to scientific research.
McGraw-Hill, NY.
Personal Communications
Hale, D. 1984. MMS, Anchorage, AK. Telephone conversation,
December 27.
Prentki, R. 1985. MMS, Anchorage, AK. Memo, March 1.
Schumacher, J. 1985. PMEL/NOAA, Seattle, WA. Telephone
conversation, May 3.
53
-------
Appendix A
EVALUATION OF OIL SPILL RISK
Spill Occurrence as a Poisson Process
The principal reference for the Poisson model used in the
EIS is a USGS Open File Report by Anastase Nakassis (Nakassis
1982). The abstract for this paper is short and to the point:
"ABSTRACT: In what follows we examine the hypothesis that
there has been no improvement in the offshore oil production
safety [versus] the hypothesis that there has been a gradual
improvement. Our analysis will show that the second
hypothesis is much better supported by the available data."
The report is referenced in the supporting documents to the EIS
to indirectly suggest that the modeling of oil spill frequencies
is correctly done with a Poisson process based on a modified oil
production variate as an exposure variate. For example, Lanfear
and Amstutz (1983) state that they treat oil spill occurrence as
a Poisson process. Their justification and methodology are based
on Nakassis (1982).
We conclude that Nakassis1 report does not show that oil
production is an appropriate exposure parameter. Rather,
Nakassis first shows that production is not a good exposure
variate; and then he attempts to transform the data to create a
better exposure variate under the assumption that oil spills are
caused by a Poisson-like process. The validity of the
transformation that Nakassis uses rests on the Poisson assumption
but does not verify the assumption.
Tests of Oil Spill Occurrence Patterns
The Nakassis report first examined the question, "Is there
any statistical evidence that we are not dealing with a Poisson
process whose parameter is proportional to the oil produced?"
Four different statistical procedures were used and, without
exception, all suggested that the data &Q not support the
proposed model, i.e., the relationships between the intervals
between spills are not like those predicted for a Poisson
process.
The underlying property exploited by Nakassis to test the
model is that the waiting time between Poisson-distributed events
is identically and independently distributed. This allows the
A-l
-------
use of distribution-free statistical tests (including tests of
runs and rank-order correlations) and tests based on parameter-
free ratios of summations; all of these were employed by
Nakassis. The database employed by Nakassis consisted of nine
spills that occurred between 1 January 1964 and the end of 1980.
The data consist of the dates of the nine spills and the amount
of oil produced in federal OCS waters between spills. There is,
however, one important exception: the amount of oil produced
prior to the first spill in the database (8 April 1964) is
presumed to be the amount of oil produced between 1 January and
8 April 1964.
Nakassis accepts, without discussion, 1 January 1964 as an
acceptable starting date for calculation of the exposure prior to
the first spill in his database. This is somewhat controversial.
The selection of 1 January 1964 results in an exposure of
31.3 million barrels (Mbbl) for the first spill, whereas the
cumulative OCS production prior to 1964 was 380 Mbbl. One could
argue that 31.3 is not a very likely draw from the population.
More significantly, the tests of runs and rank-order calculations
would be less conclusive if the exposure prior to the first spill
were changed only slightly.
We believe it would be preferable to use the date of the
first spill as the beginning of the first interval to be
analyzed. If we change the starting date from 1 January 1964 to
4 April 1964, we change the statistics calculated by Nakassis.
Most importantly, we reduce the number of inter-arriVal intervals
from nine to eight. With respect to Lanfear and Amstutz's
analysis, we reduce their number of intervals from 12 to 11.
These changes do not alter the conclusion that the data are not
independently and identically distributed.
Table A-l shows the modified form of Nakassis1 data. (We
also recalculated the production values and found some slight
discrepancies with Nakassis1 results.) The statistical tests we
applied to the data are Kendall's * and Q parameters which
compare rank ordering between two lists. The rank orderings we
use are based on the date of the spill (list 1) and the
inter-arrival prod-uction (list 2). The statistical test suggests
that the two lists are positively linked (the chance that they
are not is 0.27 percent), which implies that longer inter-arrival
intervals occur at later dates. This is not consistent with a
Poisson distribution. The same test applied to Lanfear and
Amstutz's data (Table A-2) showed that the chance of independence
of their data is 2 percent.
Thus, correcting the starting value and using two spills in
October, 1964, we still concur with the paper's conclusions
regarding the import of these tests. Specifically, Nakassis
states:
"Thus it seems that there are excellent reasons to reject
the hypothesis that the random variables Xj, X2, ... Xn are
identically distributed and independent."
A-2
-------
REF
NUM
1
2
3
4
5
6
7
8
9
Table A-l. Test of Poisson
SPILL PRODUCT TON
DATE NAKASSIS
08-Apr-64
03-Oct-64
09-JU1-65
23-Jan-69
16-Mar-69
10-Feb-70
Ol-Dec-70
09-Jan-73
23-Nov-79
55
103
712
43
270
272
804
2054
0
.9
.1
.9
.5
.5
.5
.9
.3
Fit to Arrival
INTERVAL TIME
RECALCULATED
55.9
98
716
42
272
271
811
2064
.8
.9
.0
.6
.5
.2
Times Between Spills: Modified Nakassis Data
RUNS RANK STATISTICS
2 Kendall:
3
+ 6
1
4
5
+ 7
+ 8
T ซ
I =
S =
Tau = 0
Chance
25
3
22
.785
of
being
Poisson = 0.275%
Mean = 541.6
First spill noted by Nakassis is assumed to be the beginning point of the
modified data set, i.e., the first spill noted by Nakassis is not included
in determining run sequence and not ranked.
-------
Table A-2. Test of Poisson
REF
NUM
1
2
3
4
5
6
7
8
9
10
11
12
SPILL
DATE
08-Apr-64
03-Oct-64
19-Jul-65a
28-Jan-69b
16-Mar-69
17-Aug-69
10-Feb-70
Ol-Dec-70
20-Jul-72
09-Jan-73
23-Nov-79
17-Nov-SO
PRODUCTION
NAKASSIS
0
55.9
103.1
712.9
43.5
270.5
272.5
804.9
2054.3
Fit to Arrival Times Between Spills: Lanfear and Amstutz Data
INTERVAL TIME
RECALCULATED RUNS
55.9
102.6
717.3 +
38.0
124.6
147.9
271.5 +
632.7 +
178.5
2064.0 +
252.5 +
RANK
2
3
10
1
4
5
8
9
6
11
7
STATISTICS
Kendall:
T = 41
I = 14
S = 27
Tau = 0 .490
Chance of
being
Poisson = 2.027%
Mean
416.9
Nakassis reports 9 July 1965 (see Table A-l).
Nakassis reports 23 January 1969 (see Table A-l).
-------
Although it is not explicitly stated as such, a necessary
corollary to this finding is that the spill process is not well
modeled using a Poisson process with production as an exposure
parameter. This is because some of the properties that
distinguish Poisson processes are 1) that they have no memory and
2> that intervals between events are therefore identically and
independently distributed. Any finding that suggests that the
data are not independently and identically distributed also
rejects any hypothesis of Poisson behavior.
The Idea of Time-Varying Rate Parameters
Nakassis next asks the question: "Can a better model be
constructed using the idea that the rate parameter has declined
with time?" The idea for this approach is reasonable, and it is
strongly suggested by the analysis. As a result of the increased
concern over oil spillage, it is presumed that the industry has
improved its practices and consequently its safety record over
time and that the frequency of spills has decreased. Rather than
create a time-varying rate parameter, Nakassis finds a
transformation of the original production data that more closely
approximates a Poisson variate when considered from the
standpoint of the OCS spill history. This is a subtle but
important distinction that was not discussed in Nakassis1 report.
The approach taken by Nakassis is to formulate three
distinct, single-parameter transformations of the original
production date. Each of these transformed variate-families are
then treated as candidates for a new exposure variate for
modeling the spill generation process. Nakassis uses a maximum
likelihood method to select an optimal value for the fitted
parameter in each family, and the magnitude of the likelihood
function is used to determine if one of the transformation
families is a better choice than the others.
The maximum likelihood calculation is based on the
distribution of the number of spills that can occur in an
interval of arbitrary length of the exposure variate. At its
most rudimentary form, the maximum likelihood idea can be thought
of as a way of fitting annual exposure to the original spill
incidence data. The probability of having no spills over a small
interval is large (say close to 1.0), while the probability of
having many spills is small. Each transformation proposed by
Nakassis will assign its own unique pattern of large and small
intervals to the original data. Inevitably, one of the
transformations will do a better job of recasting the original
production data in terms of the magnitude of the maximum
likelihood function. If we have some years with many spills, the
"best" transformation will cause these years to have large
increments of exposure. In years with no spills, the "best"
transformation will have small increments in exposure.
Nakassis selects his intervals to correspond to the
increments of the transformed variates that occur over the
A-5
-------
calendar dates 1 January to 31 December in each of the years 1964
through 1980. (It is this procedure that results in the
controversial first spill interval.) Since it is assumed that
the spill generation process is Poisson, the distribution of the
number of events in an interval of arbitrary length is given by
the Erlang distribution, and Nakassis1 likelihood function is
based on this functional form. The procedure used by Nakassis to
find the maximum likelihood solution exploits the properties of
this likelihood function under the assumption that the
transformation is linear over any given 1-year interval (Nakassis
pers. comm.); this assumption is not explicitly stated in the
open-file report. The procedure adopted by Nakassis is
mathematically rigorous except for this assumption which
oversimplifies the exposure variate behavior.
With each transformation that Nakassis postulates, the
observed spill history creates a unique set of observations that
consist of the number of events that occur over intervals that
correspond to a calculable (and variable) exposure. This
exposure is determined by the change in the transformed variate
over the annual interval. The transformations proposed by
Nakassis are one-parameter functions, and his optimization
process is directed at finding the optimal value of this
parameter for each family.
For expository purposes, it is easiest to consider a
discrete version of Nakassis1 problem. Imagine that we are given
three different transformations with fixed parameters, and that
we have calculated the increments in these exposure variables for
each of the 17 years in our sample. Our data will then consist
of three arrays of numbers. Each array will have 2 columns and
17 rows. In any given row, the first column will contain the
number of spills observed, and the second column will contain the
change in the transformed variate from 1 January to 31 December
in the year corresponding to the row. For each of the
transformations of the original oil production data, there will
be a best estimate of the Poisson rate parameter; this estimate
is simply the number of spills divided by the total exposure over
the 17-year period. Given the exposure, the number of events,
and the rate parameter, the Erlang distribution provides us with
a probability for the outcome of each row. Assigning such a
probability to each row, the arrays can then be assigned
cumulative likelihoods that are the product of the row-wise
probabilities. Among all three, one array will have the largest
likelihood product. It will be the maximum likelihood
transformation. The method employed by Nakassis simply expands
the range of possible transformations to include the possibility
of continuous variation in the transformation parameter.
In its general formulation, Nakassis1 method is correct.
But in some details, methodology is approximate and prone to
error. For example, the assumption that the exposure of the
transformed variate over a year can be approximated by the
product of the average production rate and the value of the
transformation function at the beginning of the year is not
A-6
-------
warranted in the early period of the analysis. As we show in the
section below, this assumption predicts that the probability of
having no spills over a year is about 0.6, whereas a more exact
solution places this probability at about 0.45. Nakassis1
approach leads to a simpler estimation problem but at the expense
of accuracy.
There is another feature of Nakassis' analysis that bears
discussion. There is no guarantee that his fitting procedure
would reject a false fit to the data. The maximum likelihood
method simply ensures that of the possibilities considered, the
one selected has the closet fit to the pattern implied by the
Erlang distribution. The maximum likelihood procedure does not
distinguish between good and bad choices except to the extent
that the analyst had included a sufficiently broad selection to
begin with.
The only real test of the model comes from the subsequent
tests of the runs and correlations of the fitted data, but the
eight or nine data points have already been used to reject three
parameters and to estimate two more. It is likely that the fit
is only apparent.
The Model Implied by a Time-Varying Rate Parameter
The rate of DCS oil production varied greatly with time over
the period 1964-1980 (Figure A-l). The functional form proposed
by Nakassis (1982) for transforming production also involves
large changes from one year to the next, particularly in the
1964-1968 period (Figure A-2). In this section, exact
distributions are derived for a process like that prescribed by
Nakassis. Three generic production histories are considered:
steady production, increasing production, and decreasing
production. This analysis may be used to examine the degree of
error introduced by Nakassis through his assumption that the
production rate was constant -through the year, and that the
transformation variate could be approximated with the product of
the production rate and the transformation function evaluated at
the beginning of the year.
Assume that we have a Poisson process that generates spills
according to the quantity of oil produced. Specifically, assume
that the probability of N spills in an infinitesmal exposure
interval is given by the following equation:
1 - AAq : N=0
P[N,Aq] = AAq : N=l ^)
0 : N>1
This states that the probability of observing one event is AAq,
that the probability of two or more spills is identically zero,
and that the probability of no spills is therefore 1 minus the
probability of one spill. The exposure is denoted here by the
A-7
-------
400
I
oo
o
3
T3
O
CO
3
C
C
OIL (NO CONDENSATE) Mbbl
Annual Production
Date
FIGURE A-l, USGS PRODUCTION DATA
SOURCE: NAKASSIS 1982; USGS 1972,
-------
r
VD
v>
3
.0
CO
-------
variate "q". It is presumed to be production volume in keeping
with Nakassis1 paper. Assume that the spill rate constant
changes with time, as is first proposed by Nakassis.
Provided the inf initesmal spill probabilities are as given
above, then it follows that the probability of observing zero
spills over an interval (Ofq) must obey the following ordinary,
first order differential equation:
l + APlO,q] = 0 (2)
We may now incorporate Nakassis1 presumption that the rate
constant " A" is a function of time. Nakassis estimates that the
best fit is obtained with the form:
But time and oil production are linked via the production history
of the system that is under observation. That is, if we had one
common functional form for the rate parameter but two oil fields,
the field that had the greatest production during the high rate
constant times would have more spills on the average (all other
things being equal). More significantly, they would not share
the same distribution of the inter-arrival times.
While there may be characteristic field production curves,
it serves our purpose to consider the simplest of cases: a field
producing at a rate R that is undergoing a linear increase/
decrease of production rate over some characteristic time. To
make this idea clearer, Figure A-3 shows the inferred rates and
the rate of changes of oil production rates that would be
required to exactly model the annual OCS oil production data from
1952 through 1980. (In its present form, the approximation is
obviously unstable. The rate of change of production rate is
alternately over-estimated and then under-estimated, i.e., it
oscillates from one year to the next. This suggests that 1 year
is too long a period to approximate OCS oil production with the
simple model presented here, but for shorter periods [e.g.,
0.5 yr], it is acceptable.)
dq = (R + (6R)x t/Tc) dt
and t = T - TOf where To is the time at the beginning of the
interval.
(4)
A-10
-------
*
CM
i_
>%
^
J3
a
A
h.
>
.
JQ
.O
2
400 -i
350 -
300 -
250 -
200 -
150 -
100 -
50 -
0 -
-50 -
-100 -
.
Inferred Rate ' x%x>.
^} A t ^ A f ^^ (% a n *^ ^ '
....... i-(Qje OT onange . .
/xv/ \/ \
/ \^~^
/ x-^
/
f
/^*J
1
/
/ -/
f ซ
,r .
^ .
*** . .
'' " ' ' *. * ". . *. '.
ป .ซ*.
. . . /
;
i i i i 1 1 1 1 1 1 1
Apr-49 Oct-54 Mar-60 Sep-65 Mar-71 Aug-76 Feb-82
Year
FIGURE A-3, LINEAR PRODUCTION MODEL
-------
Given this model, we then have the following relationships
between time and production:
^o + ^ + ^T (5)
t =
where, (q-qQ)
q' = ฃ
and qo is the cumulative production at the beginning of the
interval. The 6 parameter is a nondimensional fractional value
equal to the normalized change in the production rate over the
time interval Tc.
Since the production rate model is only accurate over short
periods, we need a suitably small characteristic time interval.
A quantity with the units of time that springs naturally from
parameters at hand is:
T = l/(A R) (7)
This quantity is also small. Nakassis has chosen XQ to ke
16.4/Bbbls, and a typical DCS production rate is 0.25 Bbbl/yr, so
the characteristic time is on the order of 0.25 yr.
Defining a - i + kT0 (8a)
b = K/R (8b)
c = -XQ6b/2 (8c)
the solution to equation (2) is:
~ A
o
(9)
P {o,q} =
1 + (b + /b2 - 4ac)q/2a
1 + (b - /b2 - 4ac)q/2a
/b2 - 4ac
There are several features of this solution that bear
discussion.
If the production rate is steady, c = 0, then the
solution becomes:
A
- o
This equation does not have the form of a Poisson
distribution, but it is a close relative; and it
furthermore has the property that events are
independently and identically distributed when the
parameter q/a is used as the exposure variate. The
parameter q/a is the transformed production variate
identified by Nakas"sis.
If the rate of change of the rate parameter, Mt), is
large and if the production rate is changing rapidly
A-12
-------
(the case in 1964 in the problem at hand), then the
exact distribution is very different from a Poisson
distribution that uses the approximation employed
by Nakassis. Figure A-4 compares Nakassis1 Poisson
process with the exact form. Note that the exact form
has a much lower probability that any given quantity of
oil can be produced without incurring one or more oil
spills.
Conversely, if the rate of change parameter is small,
as it is after about 1968 for the form proposed by
Nakassis, then the exact solution is well approximated
by the Poisson equation. Figure A-5 compares the exact
solution with the Poisson solution for 1970. It can be
seen that there is some difference between the two, but
this is probably not significant.
Conclusions
Out of the infinitely numerous possibilities, Nakassis1
"better model" is a priori constrained to be a Poisson process
with a time dependent rate parameter and with exposure measured
in oil production. Considerable ingenuity has gone into tuning
this model to fit the observations as closely as possible, but
the fundamental issue of the model's applicability has not been
addressed. If we consider that: 1) the database consists of
nine (or better yet eight) inter-arrival times for larger spills,
2) the rate parameter is estimated from the data, and 3) one of
three one-parameter functional forms are also selected based on
their ability to fit the observations, then it is not too
surprising that the various tests do not reject the (resultant
four-parameter) model.
It is important to recall the purpose of the Nakassis study.
The study was undertaken to show that an hypothesis based on an
improvement in DCS oil spill safety is better supported than the
hypothesis that no improvement has taken place. To some degree,
these problems can be addressed qual itatively without a full
knowledge of the spill generation process; and based on the data,
we have no quarrel with the conclusion that spills are less
likely now than in 1964. But Nakassis1 estimate that this
increase in safety amounts to a factor of about three is
dependent on the model that underlies the analysis. As we
discussed above, the model has not been verified. A point of
particular concern is that the spill generation process may have
to do with the break-in of a new development, and the apparent
improvement in OCS spill rates might really be due to the aging
and maturing of the existing facilities. Under this interpre-
tation, each new field may be subject to a pronounced "learning
curve" history. In this case, we might postulate spill rates as
high as those seen in the mid to late 1960s (say 5/Bbbl) for any
new development. In any event, it is neither fair nor accurate
to cite this paper as proof that a Poisson model driven by
production volume is valid for the simulation of OCS oil spiels.
A-13
-------
U)
W
O
z
J3
O
Exact Distribution
-4- Exponential
0.00
0.28
Production (Bbbls)
FIGURE A-4, PROBABILITY OF NO SPILLS OCCURRING, 1964 COMPARISON
-------
I-1
en
J2
'5.
to
o
z
v^
JQ
O
1.0
0.9 -
0.8 -
0.7 -
0.6 -
0.5 -
0.4 -
0.3 ~
0.2 -
0.1 -
0
0.00
Exact
Exponential
I I
0.04
I
0.08
I
I
0.12
Production (Bbbls)
I I
0.16
0.20
i r T
0.24 0.28
FIGURE A-5, PROBABILITY OF NO SPILLS OCCURRING, 1970 COMPARISON
-------
Appendix B
DEPICTION OF OIL SPILL STATISTICS
Summarv
This appendix contains graphical presentations of two sets
of oil spill data: spill statistics for Cook Inlet, taken from
the DEIS for St. George Basin (DOI 1984), and oil spill
statistics for all U. S. DCS platforms, compiled by The Futures
Group (1982). The purpose of this appendix is to portray oil
spill statistics graphically, rather than by mathematical
formulae, so the reader can visualize the spill record and
associated trends.
Although the Cook Inlet observations and The Futures Group
compilation are independent, they have sufficient similarities so
that conclusions drawn from and supported by both data sets may
be assumed to. reasonably represent the risks of offshore oil
production. The Cook Inlet data set includes three spills of
volume >1,000 bbl. Two were tanker related and one was a
pipeline spill. The data for the smaller spills include spills
from all sources: platforms, pipelines, and tankers. Although
The Futures Group report included pipeline, tanker, and single
buoy mooring spill statistics as well as platform statistics,
only platform statistics are used in this discussion. The format
in which the data are presented in The Futures Group report is
not convenient for combining information from different risk
sources.
For this discussion, we do not restrict the initial
discussion to large spills; rather, we focus on all sizes of
spills. This allows a large data set to be visualized, and
occurences of large spills can be interpreted within the
framework of all spills.
Four of the fundamental assumptions in the MMS oil spill
risk model are:
1. that spill frequency is directly proportional to
production;
2. that there has been an improvement over time in the
spill rate;
3. that the frequency of occurrence of oil spills can be
approximated by a Poisson distribution; and
B-l
-------
4. that the statistics describing small spills
ซ1,000 bbl) are different from those describing large
spills because the causes of large spills are
different.
If the MMS model assumptions are correct, it follows that
the graphs depicting oil spill statistics would generally show
the following:
Observed spill rates that increase monotonically and
linearly with increases in production.
Improvements in spill rates.
Ability of the Poisson distribution to estimate spill
frequency.
Differences between statistics for all spills and
statistics for only large spills.
However, the graphs present the following general patterns:
The Futures Group data, in particular, indicate a
bimodal nature to spill frequencies with high spill
rates at both low and high values of production. Cook
Inlet data do not have as clear a bimodal nature.
There is a relationship between spill frequencies and
changes in production.
The Poisson distribution tends to underestimate the
probability of occurrence of years with few spills and
also to underestimate the probability of years with
many spills.
Properties and dependencies of all spills appear to be
similar to those of only larger spills.
Cook Inlet Data
Figure B-l shows the cumulative number of spills as a
function of the cumulative oil production for Cook Inlet. All
spills reported and all spills of known volume >1 bbl are shown.
An immediate observation from these figures is that a straight
line fit to the observations would not pass through the origin.
Care must be exercised in interpreting this figure since any two
increasing-(cumulative) functions graphed against each other will
always show a trend even if they are totally unrelated. The
important point established by this figure is that if there is a
relation between spills and production, the relation does not
predict zero spills at zero production. One interpretation of
this observation is that a small number of spills will occur
which is associated with any startup activity and is not
B-2
-------
280
CO
All Spills
D Known Volume
200 400 600
Cumulative Production (Mbbl)
800
FIGURE B-l.
COOK INLET: CUMULATIVE NUMBER OF SPILLS VS
CUMULATIVE PRODUCTION
B-3
-------
dependent on the amount of oil produced. A second interpretation
is that there has been an improvement in the spill rate,
particularly early in the production period. A lessening of the
spill rate diminished the slope of the curves in the figure.
This is seen to occur after a cumulative production of about
80 Mbbl. This level of production was reached in Cook Inlet
about 1968.
Figures B-2 and B-3 show some of the details of Cook Inlet
incidence of oil spills. Figure B-2 shows the number of spills
per year as a function of year. There are two curves, the lower
curve is the number of spills whose volume is known. The upper
curve is the total number of spills including those whose volume
is unknown. Data prior to 1972 are from EPA records; 1972 and
later data were furnished by the U. S. Coast Guard. The
incidence of spills with unreported volumes diminished after
1972, probably reflecting changes in record keeping. There
appears to be a downward trend after 1968 of the annual number of
spills. Spill volumes as a function of time are shown in
Figure B-3. There is a definite reduction in spill volume after
1969. The Sale 89 FEIS indicates that there has been an
additional large spill since 1980 in Cook Inlet. This spill is
not shown because we used data only through 1980.
Shown in Figure B-4 is the annual spill frequency as a
function of annual production. There is an indication of
association between high spill rates and high production.
However, the highest spill rates are not associated with the
highest productions. The production history of the Cook Inlet
development is shown in Figure B-5. One could ask whether spills
are related to large changes in production. Stable, high
production occurred between 1969 and 1975; during this time, the
spill rate was large but not at its peak. Figure B-6 shows the
correlation of spills with changes in production. A trend of
increasing number of spills with increasing rates of change in
production may exist, but it appears weak with Cook Inlet data.
In Figure B-7, we show a comparison of actual spills and the
Poisson distribution for spills as a cumulative function. The
x-axis plots number of spills/yr (X) and the y-axis shows the
cumulative percent of years with spills less than X. Two curves
are shown, one for the Cook Inlet observations (271 total spills)
and one for the Poisson predictor with a rate constant of
16.9 spills/yr (the observed average spill rate). The cumulative
probability graph shows that for years with few spills (e.g*,
10 spills or less), the Poisson distribution underestimates the
cumulative frequency of occurrence. Thus, the Poisson
distribution predicts that 5 percent of the years will be
characterized by 10 or fewer spills, whereas Cook Inlet had
20 percent (3 of 14 years). For years with many spills (e.g.,
30 spills), the Poisson distribution overestimates the cumulative
frequency of occurrence; however, it overestimates the cumulative
frequency of years with 30 or JLaks spills. It follows, then,
that the Poisson distribution underestimates the cumulative
B-4
-------
60
50 -
ป
All Spills
D Known Volume
1965 1967 1969
i i r
1971 1973 1975
Year
1977 1979
FIGURE B-2, COOK INLET: ANNUAL SPILL RATES
B-5
-------
2.6
1965
1967 1969
1971 1973
Year
1975
1977
1979
FIGURE B-3, COOK INLET: ANNUAL SPILL VOLUME
B-6
-------
60 t
50 -
40 -
a.
to 30 -
20 -
10 -
All Spills
D Known Volume
Pป * *
D
B
20 40
Annual Production (Mbb!)
60
80
FIGURE B-4, COOK INLET: SPILL RATES VS ANNUAL PRODUCTION
B-7
-------
80
70-
60-
50-
40-
ฃm
ฃw
u .
s o
ฃ 30 H
Q.
20-
10-
All Spills
Production
i i I I i I i i i ) r i I i
1965 1967 1969 1971 1973 1975 1977 1979
Year
FIGURE B-5, COOK INLET: PRODUCTION AND NUMBER OF SPILLS
OVER TIME
B-8
-------
ou
50-
40 -
"5.
CO
. 30 -
0
z
20 -
o -
D
D
ftflQ1
B
S
D
' S ฐD D * AM Spilte
D Known Volumes
-10
0 10 20 30
Annual Production Change (Mbbl)
40
FIGURE B-6, COOK INLET: NUMBER OF SPILLS VS CHANGE IN
PRODUCTION
B-9
-------
Observed
Predicted
20 40
No. Spills/Year
NOTE: PREDICTED CUMULATIVE FREQUENCY DERIVED FROM POISSON
DISTRIBUTION WITH EXPECTED MEAN = 16.9 SPILLS/YR.
FIGURE B-7,
COOK INLET: CUMULATIVE FREQUENCY OF OBSERVED
AND PREDICTED SPILL RATES
B-10
-------
frequency of years with 30 or more spills, i.e., underestimates
the occurrence of years with many spills.
Thus, in reviewing Cook Inlet data, it appears that:
There is an indication of improvement in spill
statistics.
Spill rates may depend on large changes in production
rate as well as production level, but the database for
Cook Inlet is relatively small and trends are not
clearly observed.
The Poisson distribution underestimates the number of
years with few or many spills.
Because the Cook Inlet database is relatively small, it is
important to determine whether similar observations are evident
with the larger database provided by The Futures Group (1982).
The Futures Group Data
The Futures Group information includes only spills larger
than 50 bbl and, furthermore, breaks the spills from platforms
into three categories: blowout, tank, and other spills. We
consider the sum of all platform spills independent of category.
Figure B-8 shows the cumulative number of spills as a
function of cumulative oil production on the U. S. DCS. Two
curves are shown: one for all spills >50 bbl and one for all
spills >1,000 bbl. A similar trend was seen in Cook Inlet data
(Figure B-l). Prior to the production of the first Bbbl, the
curves for all spills and for spills >1,000 bbl are similar.
This could be due to a number of causes, e.g., better reporting
of small spills later on in production, or improvements which
helped keep spills (when they occurred) to a small size.
Figure B-9 is also similar to Cook Inlet data (Figure B-3)
in showing a dramatic reduction in volume spilled since the early
1970s.
Figures B-10 and B-ll display some of the details of The
Futures Group platform spills. Figure B-10 shows the number of
spills per year, with spills defined by five categories: >50,
>100, >250, >500, and >1,000 bbl. Figure B-ll shows the
production history of the production platforms considered in The
Futures Group report.
Spills as a function of production are shown in Figure B-12.
They show a bimodal pattern of higher spill rates at high and low
production levels. The bimodal pattern is more pronounced than
for Cook Inlet data (Figure B-4). Thus, it appears that
production level alone is not the sole predictor of spill rates.
B-ll
-------
50
40 -
ta
'o.
w
d
z
0)
30 -
ซ 20
3
E
D
O
10 -
D > 50 bbl
> 1,000 bbl
1 2 3
Cumulative Production (Bbbl)
FIGURE B-8,
U, S, OCS: CUMULATIVE NUMBER OF PLATFORM
SPILLS VS CUMULATIVE PRODUCTION
B-12
-------
1964 1966 1968 1970
1972
Year
1974 1976 1978 1980
FIGURE B-9, U.S. DCS: ANNUAL PLATFORM SPILL VOLUME
B-13
-------
10
9 -
8-
7 -
co
W
4 -
3 -
X
X
X
O O
D O
A D
A
D
~ T I ^"
O
x
1964 1966
1968 1970 1972 1974
Year
1976 1978 1980
X > 50 bbl
O > 100 bbl
D > 250 bbl
> 500 bbl
A >1000 bbl
FIGURE B-10, U. S, OCS: ANNUAL PLATFORM SPILL RATES
B-14
-------
360
1964 1966 1968 1970
1972
Year
1974 1976 1978 1980
FIGURE B-ll, U, S, DCS: PRODUCTION OVER TIME
B-15
-------
10
9 -
8 -
7 -
50 bbl
O > 100 bbl
D > 250 bbl
> 500 bbl
A > 1000 bbl
FIGURE B-12,
U, S, DCS: PLATFORM SPILL RATES VS
ANNUAL PRODUCTION
B-16
-------
Figure B-13 shows spill rates as a function of change in
production. In this figure, we see that spill incidence is
highest for both positive and negative changes in production and
least for little change in production. Figure B-14 compares
spill rates with the magnitude of change in production. Although
there appears to be a relationship indicating increased risk with
increased change, the data would not be significantly fit with a
straight line. The important point of the figure is that highest
spill rates are more closely tied to the larger changes in
production. If there is a year of many spills, this year is more
likely to be a year of high production change than a year with
production similar to the previous year.
Figures B-15 to B-19 compare spill rates with predictions
from a Poisson distribution. For all categories of spill sizes,
the Poisson distribution underpredicts the frequency of
occurrence of years with few spills and also underpredicts the
frequency of occurrence of years with many spills. The same
pattern was observed in the Cook Inlet observations (Figure B-7).
The Kolmogorov-Smirnov test is a crude statistical method of
examining goodness of fit of observations to predicted cumulative
frequency distributions. It is a fairly simple test that has one
advantage of quickly identifying data sets that do not fit
predicted distributions. (Because of its nature, it is not
capable of demonstrating that a set of data .do_e..s. fit the expected
pattern.) The calculation on the data for large spills
(>1,000 bbl) recorded in Figure B-19 reveals that the
observations pass the test at the 85 percent confidence limit.
At this level of confidence, spill rates for >500 bbl
(Figure B-18) also pass the test, but spill rates for >50 bbl,
>100 bblf and >250 bbl do not pass the test. Thus, the pattern
of all spills is not fit by a Poisson distribution.
MMS assumes that the causes of large spills are different
from causes of small spills. For example, large spills tend to
be due to blowouts or collisions (major accidents), whereas small
spills tend to be caused by equipment failure or human error. On
the other hand, one could argue that major accidents also can
result from human error and equipment failure, and the main
difference is one of magnitude. Figures B-15 through B-19 show
that all size classes of spills have the same bias relative to
the Poisson distribution. Figure B-7, comparing Cook Inlet
spills with a Poisson model, shows the same bias. Furthermore,
Figure B-10 shows that there is a high correlation between
numbers of large spills and small spills per year; the
correlation coefficient is 0.68. The Kolmogorov-Smirnov test
results indicate that spills <500 bbl do not fit a Poisson
distribution, but it is incapable of demonstrating that spills
>500 bbl do fit a Poisson distribution. Nakassis (1982) also was
unable to fit platform spills >1,000 bbl to a Poisson process
(Appendix A). These observations suggest that the assumption
that small and large spills are different must be carefully
considered. It does not appear that occurrence of large spills
B-17
-------
iu -
9 -
8 -
7 -
= 6-
'a
w 5 _
o
2 4 -
3 -
2 -
1 -
n -
X
O
c
ฎ XX
ฉ ฉ @
S&IHE . ฎi
X
X
X
0 0
D O
D A
Q A
ฎ A XD
OR . , - ,,./&> iSi iff?) A
-40
-20 0 20
Annual Production Change (Mbb!)
40
X > 50 bbl
O > 100 bbl
D > 250 bbl
> 500 bbl
A > 1000 bbl
FIGURE B-13.
U, S, DCS: NUMBER OF PLATFORM SPILLS VS
CHANGE IN PRODUCTION
B-18
-------
10
9 -
8 -
7 -
ซ 6 -
"o.
w 5 -
o
Z 4-
3 -
2 -
1 -
D
DD
D
D D D D
D D D
r-B - , - . - 1 - 1 - rQ - 1 - BT - 1
ฑ10 -20 -30 ฑ40 ฑ50
Magnitude of Production Change
FIGURE B-M. U. S, DCS: NUMBER OF PLATFORM SPILLS >50 BBL
VS MAGNITUDE OF PRODUCTION CHANGE
B-19
-------
a
CO
X
V
ฃ
100
C
>
o
fe
E
CL
E
3
u
No. Spills/Year
FIGURE B-15,
U, S, DCS: CUMULATIVE FREQUENCY OF OBSERVED
AND PREDICTED SPILL RATES (PLATFORM SPILLS
>50 BBL)
in
x
V
100-1
c
V
u
k
t>
o.
FIGURE B-16,
No.SplllE/Year
U, S, OCS: CUMULATIVE FREQUENCY OF OBSERVED
AND PREDICTED SPILL RATES (PLATFORM SPILLS
>100 BBL)
B-20
-------
tn
X
V
ฃ
5
w
fa.
c
>
0
c
V
o
e
CL
e
c
^
E
U
FIGURE B-17,
m
"o.
CO
X
V
ฃ
i
fa.
ซ
e
>
0
c
0
fa.
0.
*
ซ
3
E
3
O
FIGURE B-18,
1 UU -
80-
80-
70-
60-
60 -
1
40-
30-
20-
rJ*J CJ ป
Observed
O Predicted
I 1 ' 1 I 1 1
02468
No. Spllls/Yr.
U, S, OCS: CUMULATIVE FREQUENCY OF OBSERVED
AND PREDICTED SPILL RATES (PLATFORM SPILLS
>250 BBL)
100 -i
80-
60-
70-
60-
60-
40 -
30-
>500 bbl ^ฐ~ _ ^/^
f
I Observed
/ O Predicted
02466
No. Spllls/Yr.
U, S, OCS: CUMULATIVE FREQUENCY OF OBSERVED
AND PREDICTED SPILL RATES (PLATFORM SPILLS
>500 BBL)
B-21
-------
Observed
D Predicted
8
No.Spills/Yr.
FIGURE B-19. U, S, DCS: CUMULATIVE FREQUENCY OF OBSERVED
AND PREDICTED SPILL RATES (PLATFORM SPILLS
>1,000 BBL)
B-22
-------
is dramatically different from occurrence of all spills (other
than in magnitude).
Conclusions
Both Cook Inlet and The Futures Group data:
Illustrate significant reductions in oil spill volumes
since about 1971.
Indicate a weak dependence of spill frequency on volume
of oil produced at high volumes of production.
Indicate a weak dependence of spill frequency on
changes in oil production levels.
Display more years of few oil spills than predicted by
a Poisson distribution.
Display more years with many oil spills than predicted
by a Poisson distribution.
B-23
-------
LIST OF PREPARERS
Jones & Stokes Associates, Inc. accepts full responsibility
for the organization and content of this report. Dr. Harvey Van
Veldhuizen was Project Manager and Dr. Charles Hazel was the
Program Director and Contract Administrator. Dr. Lawrence Larsen
reviewed the oil spill trajectory analysis, made significant
contributions to Chapter 2, and prepared Appendix B. Dr. Robert
J. Stewart, special consultant to Jones & Stokes Associates, made
significant contributions to Chapter 2 and prepared Appendix A.
We thank Dr. Nakassis for his review of an early draft of
Appendix A.
------- |