international journal of critical infrastructure protection 7 (2014) 257–269

Available online at www.sciencedirect.com

www.elsevier.com/locate/ijcip

Analysis of pipeline accidents in the United States

from 1968 to 2009

Kyle Siler-Evansa,n, Alex Hansonb, Cecily Sundayb,

Nathan Leonardb, Michele Tumminelloc,d
a
Department of Engineering and Public Policy, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh,
Pennsylvania 15213, USA
b
Department of Mechanical Engineering and Department of Engineering and Public Policy, Carnegie Mellon University,
5000 Forbes Avenue, Pittsburgh, Pennsylvania 15213, USA
c
Department of Social and Decision Sciences, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, Pennsylvania
15213, USA
d
Dipartimento di Scienze Economiche, Aziendali e Statistiche, Università degli Studi di Palermo, Viale delle Scienze,
Edificio 13, I-90128 Palermo, Italy

art i cle i nfo ab st rac t

Article history: Pipelines are responsible for the transportation of a significant portion of the U.S. energy
Received 16 January 2013 supply. Unfortunately, pipeline failures are common and the consequences can be
Accepted 18 September 2014 catastrophic. Drawing on data from the Pipeline and Hazardous Materials Safety Admin-
Available online 11 November 2014 istration (PHMSA) that covers approximately 40,000 incidents from 1968 to 2009, this paper

Keywords: explores the trends, causes and consequences of natural gas and hazardous liquid pipeline

U.S. pipeline accidents accidents. The analysis indicates that fatalities and injuries from pipeline accidents are

Accident trends generally decreasing over time, while property damage and, in some cases, the numbers of

Power-law distribution incidents are increasing over time. In five of the ten cases considered in this paper, the
damage from pipeline accidents – in terms of injuries, fatalities and volume of product
spilled – are well characterized by a power-law distribution, indicating that catastrophic
pipeline accidents are more likely than would be predicted by more common “thin-tailed”
distributions. The results also indicate that relatively few accidents account for a large
share of total property damage, while smaller, single-fatality and single-injury incidents
account for a large share of total fatalities and injuries (43% versus 32%, respectively).
& 2014 Elsevier B.V. All rights reserved.

1. Introduction
Natural gas accounts for 23% of the primary energy con-
sumption in the United States [15]. This vast supply of energy
is delivered to customers through a network of transmission
and distribution (T&D) pipelines that totals more than
3.8 million km (2.4 million miles) [10]. Pipelines also account
for a large – and growing – share of liquid fuel transport.
Hazardous liquid (HL) pipelines carry crude oil and liquid
fuels such as diesel, gasoline, jet fuel and kerosene. In 2008,
282,000 km (175,000 miles) of hazardous liquid pipelines
accounted for 83% of crude oil transport and 62% of petro-
leum transport (measured in ton-miles) [1].
Unfortunately, pipeline failures are common and the con-
sequences can be catastrophic. For example, one million barrels
of oil were spilled when a hazardous liquid pipeline ruptured in

n
Corresponding author.
E-mail address: ksilerevans@cmu.edu (K. Siler-Evans).

http://dx.doi.org/10.1016/j.ijcip.2014.09.002
1874-5482/& 2014 Elsevier B.V. All rights reserved.
258 international journal of critical infrastructure protection 7 (2014) 257 –269
Kalamazoo, Michigan in July 2010. In September 2010, a trans-
mission pipeline explosion in San Bruno, California, resulted in
eight fatalities; another explosion in Allentown, Pennsylvania, in
February 2011 killed five people. Accidents such as these have
raised concerns about the safety of the nation's pipeline system.
This paper analyzes historic accident data from natural gas
and hazardous liquid pipelines with the goal of informing
future safety measures. Drawing on a database of U.S. pipeline
incidents, which includes approximately 40,000 incidents from
1968 to 2009, three important hypotheses are tested:

Pipeline accident frequency and damage are decreasing
over time.

The damage caused by pipeline accidents follows a power-
law distribution.

Relatively few pipeline accidents account for the majority
of damage.
Throughout this analysis, the severity of (or damage from)
pipeline incidents are measured in terms of three metrics:
fatalities, injuries and monetary property damage. For hazar-
dous liquid pipelines, the volume of product spilled is also
used as a measure of accident severity.
Previous studies of pipeline accidents fall into two cate-
gories: those focused only on pipelines and those that
compare risks across various energy-supply chains, which
may include natural gas and hazardous liquid pipelines.
Montiel, et al. [7] have performed a pipeline-specific analysis
of 185 accidents in 95 countries. Their study finds an increas-
ing trend in the number of pipeline accidents over time and
identifies mechanical failure as the most common cause of
natural gas pipeline accidents (43%). Hirschberg, et al. [5]
conclude that approximately “21% of all natural gas accidents
involving pipelines were caused by mechanical failures and
24% by impact failures.”
Two recent studies are specific to U.S. pipelines. Restrepo,
et al. [12] have conducted a regression analysis of hazardous
liquid pipeline incidents from 2002 to 2005; their results
reveal that the cause of an accident is a strong predictor of
the resulting monetary damage. Their study also identifies
corrosion as the most common cause of hazardous liquid
pipeline accidents. Simonoff, et al. [13] have developed a
predictive model to estimate the magnitude of damage from
transmission pipeline incidents based on their causes. Their
study also presents several accident scenarios that are
intended to inform risk management efforts.
Several studies compare the accident risks from various
energy supply chains, including natural gas. Sovacool [14]
observes that, in the energy sector, natural gas systems fail
most frequently, accounting for 33% of all accidents. How-
ever, Sovacool notes that natural gas accidents account for
only 9% of the total property damage and 0.39% of the total
fatalities. Studies conducted by the Paul Scherrer Institute [5]
reveal that the world-wide fatality rates from natural gas
accidents, when normalized by energy production, are
roughly one-fifth that of oil or coal.
The findings discussed above are not fully applicable to
the U.S. pipeline system. With the exception of the work by
Restrepo, et al. [12] and Simonoff, et al. [13], the studies are
based on international databases, many of which do not
disaggregate data between OECD (Organization for Economic
Cooperation and Development) and non-OECD countries. In
addition, Felder [4] indicates that “a fundamental problem
with investigating energy accidents is [the] incompleteness
[of data sets].” For example, Sovacool [14] identifies only 279
severe accidents from 1907 to 2007, 91 of which involved
pipeline systems. Hirschberg, et al. [5] use the Energy-Related
Severe Accident Database from 1945 to 1996, which contains
only 159 natural gas accidents.
This paper provides a detailed analysis of pipeline acci-
dents and, to our knowledge, is the first to fit a distribution to
pipeline damage. The analysis is limited to U.S. pipeline
systems and draws on a database containing approximately
40,000 incidents from 1968 to 2009.
2. Data collection and filtering methods
The analysis presented in this paper is based on data
collected by the Pipeline and Hazardous Materials Safety
Administration (PHMSA) [9]. The database covers approxi-
mately 40,000 pipeline incidents from 1968 to 2009. Detailed
information provided about each incident includes fatalities,
injuries, monetary property damage, accident location, acci-
dent cause and pipeline material.
In the case of gas transmission and distribution pipelines,
the reporting requirement includes incidents: (i) with more
than $50,000 in property damage, including damage to the
operator and other entities, but excluding the value of lost
gas; (ii) with one or more fatalities; (iii) with one or more
injuries that result in hospitalization; (iv) involving a shut-
down of a liquefied natural gas (LNG) facility; or (v) judged by
operators to be significant, even if the incidents do not satisfy
the other four criteria.
Unfortunately, the definition of an “incident” has changed
over time, leading to inconsistencies in the data. Before 1984,
only incidents that resulted in fires were reported. A dramatic
drop in incident reports occurred after this criterion was
eliminated. For consistency, the transmission and distribu-
tion pipeline data were filtered to include only the incidents
that met the first three criteria listed above. The fourth
criterion pertaining to liquefied natural gas facilities is out-
side the scope of this analysis. The fifth criterion was also
excluded because operator judgment is unlikely to have been
consistently applied and may, therefore, bias the results.
In the case of hazardous liquid pipelines, the reporting
requirements have remained unchanged since 2002. Inci-
dents are reported if they result in (i) more than $50,000 in
property damage, including the value of lost product, cleanup
costs and damage to the operator and other entities; (ii) one
or more fatalities; (iii) one or more injuries that result in
hospitalization; (iv) an explosion or fire; or (v) the loss of five
gallons or, in some cases, five barrels of hazardous liquids.
Before 2002, there were numerous changes in the report-
ing criteria. For example, the threshold for property damage
was increased from $5,000 to $50,000 in 1994. In 2002, the
threshold for the volume of hazardous liquid spills was
reduced from 50 barrels to five gallons (or in some cases five
barrels). For consistency, the first four criteria listed above
 international journal of critical infrastructure protection 7 (2014) 257 –269
were applied to the complete data set. The threshold for the
fourth criterion (volume of hazardous liquid spilled) was
increased to 50 barrels. Note that certain inconsistencies
cannot be mitigated. For example, the definition of an injury
was changed in 2002, but it is unclear if the change affected
the number of injuries that were reported.
Table 1 summarizes the data that is most relevant to this
study. Numbers in parentheses correspond to the raw data while
numbers without parentheses correspond to the filtered data.
Distribution pipelines have approximately 80% more reported
incidents compared with transmission and hazardous liquid
pipelines. However, after the data was filtered, hazardous liquid
pipelines have about 75% more incidents than distribution
pipelines and approximately 340% more incidents than trans-
mission pipelines. The ratios of fatalities and injuries remain the
same after filtering; distribution pipelines have four times as
many fatalities and seven to ten times as many injuries.
Distribution pipeline accidents have roughly one-sixth of the
property damage compared with the other two pipeline types.
Note that an outlier was removed from the distribution pipeline
data. The outlier, which was caused by Hurricane Katrina,
resulted in $520 million in property damage compared with the
next highest property damage of $44 million.
Hazardous liquid and transmission pipelines are similar in
terms of total incidents, fatalities and property damage, but
transmission pipelines account for more injuries than hazardous
liquid pipelines. Two significant outliers were removed from the
hazardous liquid pipeline injury data. Both incidents occurred
near Houston, Texas on October 24, 1994. According to the
PHMSA data, one incident involved a diesel spill that resulted
in 926 injuries and the other involved a gasoline spill that
resulted in 925 injuries (the next highest value for a single
incident is 31 injuries). A report by the National Transportation
Safety Board (NTSB) [8] notes that the incidents were caused by
flooding in the area, which resulted in eight pipeline ruptures
and 547 mostly minor burn and inhalation injuries. The PHMSA
data appears to overstate the injuries from this incident and we
suspect that injuries were mistakenly double-counted for the
diesel and gasoline spills. The results presented in this paper
discount the Hurricane Katrina and Houston incidents. However,
the implications of these outliers are discussed in later sections.
3. Incident causes
The PHMSA data gives brief descriptions of incident causes. The
descriptions were filtered into six categories: (i) material failure;
Table 1 – Summary of PHMSA data before and after filtering.
Distribution
Total incidents 4353 (18,366)
Incidents with fatalities 449
Total fatalities 707
Incidents with injuries 2788
Total injuries 5384
Incidents with property damage 2732a (14,512)
Total property damage (Million 2010 USD) 680a (1260)
a
Outlier removed; see text for description.
259
(ii) corrosion; (iii) outside forces; (iv) weather and natural
disasters; (v) operator error; and (vi) others. Material failure
includes the failure of a valve, weld, body of pipe or other
component. Corrosion includes internal corrosion and external
corrosion. Outside forces include failures due to excavation
damage and pipelines struck by vehicles or machinery.
Weather and natural disasters include lightning, high winds
and earth movement. Operator error includes incidents result-
ing from incorrect operation by carrier personnel or operators.
Other includes incidents that are not covered by the other five
categories.
Fig. 1 presents summary data about the causes of pipeline
incidents. In the case of hazardous liquid pipelines, incidents
were most often caused by corrosion and material failure
(  29% each), followed by other (23%), outside forces (12%)
and operator error (6%). The data suggests that material failure
has become less problematic in hazardous liquid pipelines.
Through the 1970s, nearly 40% of incidents were attributed to
material failure, compared with only 20% of incidents between
1980 and 2009. Approximately 41% of distribution incidents
have no descriptions (all incidents from 1970 through 1983). The
remaining incidents were caused by outside forces (30%), other
(19%) and operator error (5%). Corrosion accounted for only 2%
of distribution pipeline incidents. Outside forces were the most
common accident cause for transmission pipelines (30%),
followed by a fairly even split between corrosion, material
failure and other (  20% each). Corrosion was attributed to an
increasing share of transmission pipeline accidents, accounting
for roughly 9% of incidents in the 1970s compared with 25% of
incidents between 1990 and 2009.
4. Hypothesis testing
This section presents the analysis, methods and results corre-
sponding to three hypotheses: (i) pipeline accident frequency
and damage are decreasing over time; (ii) the damage from
pipeline accidents follows a power-law distribution; and (iii)
relatively few pipeline accidents account for the majority of
damage.
4.1. Hypothesis 1
Hypothesis 1 states that pipeline accident frequency and
damage are decreasing over time. For transmission, distribu-
tion and hazardous liquid pipelines, trends are observed in
accident frequency and accident damage. The accident fre-
Transmission Hazardous liquids
2263 (10,398) 7655 (10,863)
82 75
152 154
400 222a
751 534a
2124 (7597) 6430 (9329)
2050 (2140) 2218 (2260)
260 international journal of critical infrastructure protection 7 (2014) 257 –269

100%

Percentage of Incidents
80%
Weather/Nat. Disaster
Outside Forces
60% Other
Operator Error
40% No Data
Material Failure
Corrosion
20%

0%
HL. Dist. Trans.

Fig. 1 – Causes of pipeline incidents.

Incidents
500
Haz. Liquid
Distribution
Number of Incidents

400
Transmission
300

200

100

0
1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Fatalities Injuries
50 600

40 500

400
Fatalities

30
Injuries

300
20
200
10 100

0 0
1970 1980 1990 2000 2010 1970 1980 1990 2000 2010

Property Damage Product Spilled

500

400 0.5
Million barrels
Million 2010$

0.4
300
0.3
200
0.2
100 0.1

0 0
1970 1980 1990 2000 2010 1970 1980 1990 2000 2010
Year Year

Fig. 2 – Trends in pipeline accident frequency and severity between 1968 and 2009.

quency is measured in terms of incidents per year. Accident pipeline incidents. The trends are considered in terms of
damage is measured in terms of fatalities, injuries and absolute values as well as values normalized by: (i) total
property damage. The volume of liquid spilled is also natural gas consumption; and (ii) kilometers (km) of pipeline
included as a measure of severity for hazardous liquid installed.
 international journal of critical infrastructure protection 7 (2014) 257 –269 261

Fig. 2 shows the pipeline accident frequency and damage
trends in terms of the absolute values; the normalized trends
are shown in Figs. 3 and 4. Note that incidents involving
hazardous liquid pipelines (dashed lines) are shown from
1968 to 2009. Incidents involving transmission pipelines (solid
lines) and distribution pipelines (dash-dot lines) are shown
from 1970 to 2009. All the trends correspond to annual totals.
Table 2 shows the results of a correlation test, where β is
the slope of a linear curve-fit for the trend data. A negative β-
coefficient is consistent with the hypothesis, indicating that
there is a decreasing trend over time. Statistical significance
is noted at the 5% level (n) and 1% level (nn).
Fig. 2(a) shows the annual number of pipeline incidents
from 1968 to 2009. The trend in accident frequency is
decreasing for hazardous liquid pipelines, increasing for
transmission pipelines and statistically insignificant for dis-
tribution pipelines.
Fatality and injury trends are shown in Fig. 2(b) and (c),
respectively. There is a decreasing trend in fatalities from
distribution and hazardous liquid pipeline accidents. While these
trends are statistically significant, the rate of change is quite
small. For example, fatalities from hazardous liquid pipeline
accidents are decreasing at the rate of approximately one per
decade. Injuries from distribution pipelines are decreasing by
approximately seven per year; transmission pipeline injuries are
also decreasing, although the rate is less than one per year.
There is no significant trend in hazardous liquid pipeline injuries.
As discussed in Section 2, an outlier was excluded from the
analysis of hazardous liquid pipeline incidents. The excluded
incident occurred near Houston, Texas in 1994 and resulted in
more than 500 injuries. Even when the Houston incident is
included, the trend in hazardous liquid pipeline injuries remains
statistically insignificant.
For the three types of pipeline systems, the annual property
damage from pipeline incidents is increasing by two to four
million dollars per year; the trend is statistically significant in
all cases. Note that property damage is measured in nominal
2010 dollars; using real dollars would strengthen the increasing

Incidents
0.4
Incidents per billion m3

0.3

0.2

0.1
Distribution
Transmission
0
1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Fatalities Injuries
0.08 1
Fatalities per billion m3

Injuries per billion m3

0.8
0.06

0.6
0.04
0.4

0.02
0.2

0 0
1970 1980 1990 2000 2010 1970 1980 1990 2000 2010

Property Damage
0.8
2010$ per thousand m3

0.6

0.4

0.2

0
1970 1980 1990 2000 2010
Year

Fig. 3 – Trends in accident frequency and accident damage normalized by total U.S. natural gas consumption.
262 international journal of critical infrastructure protection 7 (2014) 257 –269

Incidents
4
Haz. Liquid

Incidents / thousand km
Distribution
3 Transmission

0
1970 1975 1980 1985 1990 1995 2000 2005 2010
Year

Fatalities Injuries
0.08 0.8
Fatalities / thousand km

Injuries / thousand km
0.06 0.6

0.04 0.4

0.02 0.2

0 0
1970 1980 1990 2000 2010 1970 1980 1990 2000 2010

Property Damage Product Spilled

1.5 4000
Million 2010$ / thousand km

Barrels / thousand km

3000
1

2000

0.5
1000

0 0
1970 1980 1990 2000 2010 1970 1980 1990 2000 2010
Year Year

Fig. 4 – Trends in accident frequency and accident damage normalized by kilometers of pipeline installed.

trend. As discussed above, the results presented here exclude a 4.2. Hypothesis 2
distribution pipeline incident caused by Hurricane Katrina,
which resulted in $520 million in property damage. Including Hypothesis 2 states that the damage from pipeline accidents
the Hurricane Katrina incident would strengthen the already follows a power-law distribution. In a normal distribution, most
increasing trend. events are clustered around the mean and extreme events have
In summary, a decreasing trend is observed for all the a negligible probability of occurrence. For example, the average
metrics with the following exceptions: the frequency of height of an adult American male is 178 cm, and it is unusual to
transmission pipeline incidents (increasing) and distribution see someone taller than 200 cm or shorter than 155 cm. In
pipeline incidents (flat), hazardous liquid pipeline injuries contrast, a power-law distribution has “fat tails,” meaning that
(flat) and property damage for all three pipeline types extreme events are much more likely to occur. For example, city
(increasing). Therefore, the hypothesis that the severity and population follows a power-law distribution [3]. While the mean
frequency of pipeline incidents is decreasing over time population of a city in the United States is approximately 8000,
cannot be uniformly accepted. there are many cities with much larger populations. A power-
Normalizing the data by natural gas consumption or kilo- law distribution can account for the populations of cities such as
meters of pipeline does not change the trends discussed here, Los Angeles and New York, whereas more common “thin-tailed”
with two exceptions. First, when normalized, the total number distributions do not.
of incidents for distribution pipelines is decreasing. Second, the The injury and fatality data are fitted using a discrete
normalized fatalities for transmission pipelines is decreasing power-law distribution with the following probability mass
over time (see Table 2). function (pmf):
 international journal of critical infrastructure protection 7 (2014) 257 –269 263

Table 2 – Statistical tests for the hypothesis that pipeline accident frequency and damage are decreasing over time.

Distribution Transmission Hazardous liquids

β R2 β R2 β R2

Incidents
Total  0.66 0.09 2.28nn 0.65  5.18nn 0.66
Per trillion m3 of gas consumption  1.84nn 0.22 3.32nn 0.59
Per million km of pipeline  3.56nn 0.66 4.06nn 0.54  38.76nn 0.66
Fatalities
Total  0.34n 0.14  0.12 0.1  0.08n 0.11
Per trillion m3 of gas consumption  0.66nn 0.18  0.23n 0.11
Per million km of pipeline  0.78nn 0.41  0.43n 0.16  0.68nn 0.28
Injuries
Total  7.42nn 0.63  0.81nn 0.41  0.24n 0.1
Per trillion m3 of gas consumption  13.15nn 0.64  1.49nn 0.4
Per million km of pipeline  10.92nn 0.68  2.83nn 0.48  1.81nn 0.22
Property damage (Million 2010 USD)
Total 1.08nn 0.56 3.49nn 0.19 3.56nn 0.45
Per trillion m3 of gas consumption 1.66nn 0.51 5.31nn 0.18
Per million km of pipeline 0.62nn 0.41 7.02nn 0.17 12.69nn 0.43
Product spilled (1000 Barrels)
Total  7.178nn 0.67
Per million km of pipeline  48.69nn 0.74

n
Statistically significant at 5% level.
nn
Statistically significant at 1% level.

xα1 If the data follows a power-law distribution, log–log plots

pðxÞ ¼ ð1Þ
ζðα þ 1; xmin Þ
where ζðβ; x0 Þ, the generalized zeta function, is given by
1
ζðβ; x0 Þ ¼ ∑ ðn þ x0 Þ  β : ð2Þ
n¼0
Data for property damage and barrels spilled are fit using the
continuous form of the power-law distribution with the
following probability density function (pdf):
αxαmin
pðxÞ ¼ ð3Þ
xαþ1
where α is the scaling factor, xmin is the minimum value for
which the distribution is valid and x is a real number greater
than or equal to xmin. Note that x refers to the magnitude of
damage from pipeline accidents. The resulting distributions
estimate the probability of an accident with damage x given
that the damage is at least xmin.
The scaling parameter α is estimated using the maximum
likelihood estimator (see [3]). For continuous distributions, it
is necessary to find the value of xmin that minimizes the re-
weighted Kolmogorov–Smirnov distance between the sample
distribution and the power-law fit [11]. When fitting injury
and fatality data, which have relatively small sample sizes,
xmin must be kept as small as possible; specifically, we set
xmin ¼ 2 for hazardous liquid pipeline injuries and xmin ¼ 1 for
the remaining discrete distributions.
Fig. 5(a)–(j) shows the power-law fits for fatalities, injuries,
property damage and barrels spilled. Also shown are the
results for thin-tailed Weibull distributions, which demon-
strate that exponential distributions consistently underesti-
mate the probabilities of extreme events (i.e., large accidents).
of Pðx ZXÞ would appear linear with a slope of  α. This
behavior is questionable for hazardous liquid pipeline inju-
ries (Fig. 5(c)). In the case of property damage from transmis-
sion pipelines (Fig. 5(g)), it may appear that the Weibull and
the power-law distributions equally underestimate or over-
estimate extreme events, which is not the case. The worst-
case property damage is roughly $100 million; the power-law
overestimates the probability of such an event by an order of
magnitude, while the Weibull distribution underestimates
the probability by six orders of magnitude.
For each power-law fit, Table 3 gives the scaling coefficient
with the standard error, the xmin value and the expected value
of damage (given that the accident damage is greater than or
equal to xmin). For example, for a transmission pipeline
incident with at least one fatality, the expected number of
fatalities is 2.4. In addition, Table 3 gives the p-values for the
power-law fits, which are calculated using 10,000 simulations
according to the method presented by Clauset, et al. [3]. A
p-value greater than 0.1 indicates that a “power law is a
plausible hypothesis for the data, otherwise it is rejected” [3].
The power-law fit for five of the ten cases is rejected: fatalities
and property damage from hazardous liquid pipelines, inju-
ries and property damage from transmission pipelines and
property damage from distribution pipelines.
With a power-law distribution, the mean is undefined if αo1
and the variance is undefined if αo2. As shown in Table 3, α is
less than one for property damage from hazardous liquid and
transmission pipeline incidents; α is less than two in all cases,
indicating that a reliable upper bound for the damage from a
pipeline incident cannot be estimated. In other words, large
incidents are to be expected.
264 international journal of critical infrastructure protection 7 (2014) 257 –269

Table 3 – Coefficients for power-law fits.

Distribution xmin α σðαÞ EðxjxZ xmin Þ p-value

Hazardous liquids
Fatalities Discrete 1 1.10 0.12 6.78 0.022
Injuries Discrete 2 1.45 0.07 5.06 0.586
Property damagea Continuous 502 0.90 0.04 – 0.013
Barrels spilledb Continuous 5.4 1.69 0.10 13.2 0.401

Transmission
Fatalities Discrete 1 1.36 0.08 2.41 0.376
Injuries Discrete 1 1.30 0.06 2.75 0.012
Property damagea Continuous 150 0.79 0.02 – 0.021

Distribution
Fatalities Discrete 1 1.60 0.08 1.75 0.106
Injuries Discrete 1 1.31 0.01 2.68 0.390
Property damagea Continuous 345 1.37 0.07 1.28  103 0.026

a
Thousands of 2010 USD.
b
Thousands of barrels.

Note that this analysis does not account for the trends shown
in Section 4.1, which indicate that the distributions are changing
over time. For all three pipeline types, there is an increasing
trend in property damage. The year-to-year fluctuations are also
increasing, and as a result, the estimates of α for property
damage are likely to be conservative (i.e., the power-law dis-
tributions likely underestimate the probabilities of events with
extreme property damage).
There is a decreasing trend in average injuries from gas
transmission and distribution pipeline accidents. To verify
the results, the analysis of transmission and distribution data
from 1990 to 2009, where the trend is much weaker, was
repeated. In both cases, the power-law fit is still superior to
an exponential fit and the α values are less than two.
It can be concluded that a power law is a plausible fit for
five of the ten cases tested. In these cases, catastrophic
incidents are possible and are much more likely than would
be predicted by more common thin-tailed distributions. For
example, a 1994 hazardous liquid pipeline incident resulted
in more than 500 injuries. Using the power-law fit, the
probability of such an event is estimated to be roughly 1
per 10,000 incidents that result in at least one injury. An
exponential distribution would stipulate that such an event is
nearly impossible (probability of zero to the precision of most
computers). In the remaining five cases tested, the power-law
fit is rejected. Even in these cases, however, a Weibull
distribution significantly underestimates the probabilities of
severe accidents.
Although the majority of incidents have zero fatalities,
zero injuries and low property damage, large events should
still be expected. Indeed, over the past four decades, incidents
with damage as far as five standard deviations from the
mean are relatively common.
4.3. Hypothesis 3
Hypothesis 3 states that relatively few pipeline accidents
account for the majority of damage. The results of the
previous section demonstrate that pipeline incidents follow
fat-tailed distributions in many cases, indicating that large
incidents contribute significantly to the overall risk. Based on
this finding, it is reasonable to expect that relatively few
accidents account for the majority of damage – both mone-
tary and human. Hypothesis 3 is accepted if 20% of accidents
account for at least 80% of total damage.
Lorenz curves are constructed for pipeline accidents with
non-zero damage (Fig. 6). Each curve shows the percentage of
damage (y-axis) resulting from some percentage of incidents
(x-axis). A Lorenz curve with a 451 slope indicates that all
accidents are equally severe. A greater skew in the curve
would be consistent with Hypothesis 3 that the damage is
concentrated in relatively few accidents.
Fig. 6(a) shows the Lorenz curves for pipeline accidents
resulting in one or more fatalities. The Lorenz curves are
similar for each type of pipeline, indicating that the fatalities
are relatively evenly distributed across numerous accidents.
Between 40% to 50% of the fatalities were the result of single-
fatality incidents and the worst 30% of accidents resulted in
roughly 60% of all fatalities.
Fig. 6(b) shows the Lorenz curves for accidents resulting in
one or more injuries. Approximately 40% of the injuries were the
result of single-injury incidents. Similar to the Lorenz curves for
fatalities, the worst 30% of the accidents account for approxi-
mately 65% of the total injuries. As discussed in Section 2, the
hazardous liquid pipeline data excludes an outlier, which
resulted in more than 500 injuries [8]. This single incident
accounts for more than 50% of hazardous liquid pipeline
injuries; if included, this outlier would significantly change the
Lorenz curve for hazardous liquid injuries shown in Fig. 6(b).
Fig. 6(c) shows the Lorenz curves for property damage caused
by pipeline accidents. Property damage is heavily concentrated
in a small number of incidents. About 80% of the total property
damage tolls from hazardous liquid, transmission and distribu-
tion pipeline incidents were caused by the most severe 5%, 15%
and 22% of incidents, respectively. Consistent causes are not
present for the most severe incidents. For example, for the 20%
 international journal of critical infrastructure protection 7 (2014) 257 –269 265

Transmission Distribution Hazardous Liquids

1

0.1
P(x ≥ X)
0.01

0.001
1 10 100 1 10 100 1 10 100
Fatalities Fatalities Fatalities

Transmission Distribution Hazardous Liquids

1

0.1
P(x ≥ X)

0.01

0.001

0.0001
1 10 100 1 10 100 1 10 100
Injuries Injuries Injuries

Transmission Distribution Hazardous Liquids

1

0.1
P(x ≥ X)

0.01

0.001

0.0001
102 104 106 108 102 104 106 108 102 104 106 108
Property Damage (2010$) Property Damage (2010$) Property Damage (2010$)

Hazardous Liquids
1

0.1
Data
P(x ≥ X)

0.01 Discrete power–law fit

Continuous power–law fit
Weibull fit
0.001

0.0001
100 102 104 106
Volume Spilled (barrels)

Fig. 5 – Power-law fits for fatalities, injuries, property damage and barrels spilled.

of transmission pipeline accidents resulting in the greatest
property damage, the most common causes were outside forces
(28%), other (23%), corrosion (20%), material failure (16%) and
weather and natural disasters (12%).
Fig. 6(d) shows the Lorenz curve for the volume of product
spilled, which is only available for hazardous liquid pipelines.
The worst 20% of accidents are responsible for 80% of the total
volume spilled, while the upper 10% of accidents are responsible
for 65% of the total. For the 20% of hazardous liquid pipeline
accidents with the largest spills, the most common causes were
material failure (39%), other (24%), outside forces (15%) and
corrosion (14%).
Hypothesis 3, which states that the majority of damage is
concentrated in relatively few accidents, cannot be uniformly
accepted. Single-fatality and single-injury incidents account for a
significant share of total fatalities and injuries, suggesting that
day-to-day safety measures are important to reduce the risks
from pipeline accidents. In contrast, monetary property damage
and the magnitude of hazardous liquid spills are heavily con-
centrated in a few major events.
266 international journal of critical infrastructure protection 7 (2014) 257 –269

Fatalities Injuries
100 100
Eq. Line
HL

Percent of Fatalities
80 80

Percent of Injuries
Dist.
Trans.
60 60

40 40

20 20

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Percent of Accidents Percent of Accidents

Property Damage Product Spilled

100 100
Percent of Property Damages

Percent of Product Spilled

80 80

60 60

40 40

20 20

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Percent of Accidents Percent of Accidents

Fig. 6 – Lorenz curves for damage from pipeline accidents.

distributions. When assuming a power-law distribution, the 1%

5. Power-law distribution implications
Value at risk (VaR) is a measure that is widely used in risk
management (see, e.g., [6]). This section uses the value at risk
metric to demonstrate the dramatic difference between
power-law and Weibull distributions of loss from the per-
spective of risk assessment. Value at risk gives the value of
loss such that the probability of an annual loss that exceeds
the value is equal to a given statistical threshold. For
example, a 1% VaR of $10,000 indicates that the probability
of a loss larger than $10,000 is 0.01.
All the estimates of value at risk reported in this section are
based on 106 independent simulations. Simulations were per-
formed by assuming a Poisson distribution of the number of
incidents per year, where the parameter of the distribution was
estimated from real data and eventually corrected for the trend.
To simulate data with a power-law distribution, it was necessary
to account for the effects of incidents with losses lower than xmin.
This was done by (i) evaluating the frequency f of incidents with
loss lower than xmin from the data; and (ii) using the parameter f
to simulate the number of incidents with loss lower than xmin
according to a binomial distribution with parameters f and N,
where N is the simulated annual count of incidents. After the
incidents with losses lower than xmin were simulated, the specific
loss of each incident was obtained by bootstrapping real data
(see [3]), while the losses of the remaining incidents were
simulated according to the power-law distribution.
Fig. 7 shows the value at risk metric for injuries associated
with hazardous liquid pipeline incidents, where the computa-
tions were performed using the power-law and Weibull
VaR is 86 injuries compared with 36 injuries when using the
Weibull distribution. The gap widens for increasingly severe
incidents: the 0.1% VaR is 371 injuries and 46 injuries for the
power-law and Weibull distributions, respectively.
Table 4 shows results for value at risk thresholds ranging
from 5% to 0.1%, with the results calculated using the power-
law and Weibull distributions. From a risk management
perspective, the results give the expected losses from
extreme events. As shown in Table 4, the results vary
significantly for the power-law and Weibull distributions.
For a 1% VaR, the power-law distributions predict property
damage tolls that are at least one order of magnitude greater
than those predicted by the Weibull distributions. This result
is a consequence of the fact that events with extreme losses
are essentially ignored by the Weibull distribution.
Policies and practices focused on preventing incidents may
be insufficient to reduce the risks associated with pipelines to a
significant degree. Reducing the annual number of incidents
does not affect the shape of the loss distribution associated with
a given incident, which remains a power law. Therefore, even
with a reduced frequency of incidents, the risk of an extreme
annual loss remains sizable.
These results have important implications for risk manage-
ment. Call-before-you-dig programs, for example, are aimed
at reducing the frequency of pipeline accidents. The results
presented in the preceding sections suggest that such a safety
measure will not significantly reduce the property damage risk
due to the dominance of extreme events. In contrast, automatic
shut-off valves can reduce the magnitude of damage when
accidents do occur. Thus, while they do not affect accident
 international journal of critical infrastructure protection 7 (2014) 257 –269 267

Power-law
Weibull
1e-01

Weibull: one-year 1% VaR = 36 Power-law: one-year 1% VaR = 86

P(x ≥ X)

1e-02

Power-law:
Weibull: one-year 0.1% VaR = 46 one-year 0.1% VaR = 371
1e-03

1e-04

1 10 100 1000 10000

X (# of injuries per year)

Fig. 7 – Annual risk curve for injuries associated with hazardous liquid pipeline incidents.

Table 4 – Comparison of annual value at risk for power-law (P) and Weibull (W) distributions.

5% VaR 1% VaR 0.5% VaR 0.1% VaR

P/W P/W P/W P/W

Hazardous liquids
Fatalities 23/12 85/16 153/18 625/22
Injuries 37/28 86/36 131/40 371/46
Property damagea 250/42 1,277/53 2,715/59 15,568/72
Barrels spilledb 239/174 376/194 486/202 993/218

Transmission
Fatalities 16/12 42/17 67/19 208/23
Injuries 33/21 86/26 142/29 466/35
Property damagea 1285/98 9017/112 21,542/117 168,351/129

Distribution
Fatalities 38/30 71/36 98/40 237/44
Injuries 178/100 416/114 650/119 1991/130
Property damagea 35/27 77/33 116/36 340/42

a
Millions of 2010 USD.
b
Thousands of barrels.

frequency, automatic shut-off valves can reduce the severity of
large events. While a complete evaluation of safety interventions
is beyond the scope of this work, it is recommended that policy
makers consider separately the implications of accident fre-
quency and accident magnitude when designing and imple-
menting safety measures.
6. Discussion
Approximately 2.5 million people die in the United States
every year; roughly 5% of the deaths are the result of
accidents [16]. Compared with fatalities from motor vehicles
(  40; 000 annually), unintentional poisoning ð  30; 000Þ and
falls ð  20; 000Þ, pipelines pose a minor risk – only about 30
fatalities per year.
While the pipeline accident fatality rate is small relative to
the overall fatality rate, pipelines account for a large share of
energy-supply accidents. A recent report from Carnegie Mellon
University [2], which presents a comparison across U.S. energy
supply chains, reveals that the natural gas industry has the
second highest fatality rate, behind coal and ahead of oil.
Regardless of their comparative size, the human, mone-
tary and environmental damage tolls arising from pipeline
incidents are not negligible. Where possible, cost-effective
safety measures should be pursued. In general, there are two
approaches: (i) safety measures that reduce the number of
incidents; and (ii) safety measures that mitigate the damage
from incidents.
As discussed in Section 4.1, there has been a dramatic
drop in the number of incidents involving hazardous liquid
pipelines; indeed, hazardous liquid pipelines are much safer
today than they were 40 years ago. In contrast, the number of
268 international journal of critical infrastructure protection 7 (2014) 257 –269
distribution pipeline incidents has been fairly constant since
1970, but more recent incidents tend to have fewer fatalities
and injuries. In other words, the severity of distribution
pipeline incidents has decreased, while the number of inci-
dents has not.
Section 4.2 shows that, in many cases, the damage tolls
from pipeline accidents have a “fat-tailed” distribution with
an undefined variance, meaning that it is not possible to
calculate a reliable upper bound for incident damage. This
finding suggests that catastrophic accidents should be
expected, which is consistent with historical data. Over the
past 40 years, single incidents resulting in more than 500
injuries, 40 fatalities or $500 million in property damage have
occurred. Mitigating the damage caused by these relatively
few catastrophic incidents could significantly reduce pipeline
system risk.
Section 4.3 shows that single-fatality incidents have
resulted in more than 400 fatalities since 1970 – accounting
for more than 40% of all pipeline fatalities. This suggests that
day-to-day safety measures are also important to reduce the
overall risk.
Further research is needed to investigate the costs and
benefits of various safety measures, which include call-before-
you-dig programs, pipeline inspections, automatic shut-off valve
installation, safety violation fines and high-risk pipeline replace-
ment (e.g., replacement of old cast iron pipelines in densely-
populated areas). When evaluating safety measures, it is impor-
tant to separately consider the implications of accident fre-
quency and accident magnitude. Interventions that reduce
accident frequency are likely to be more effective at reducing
fatalities because the overall fatalities are spread over many
incidents. Interventions that reduce accident magnitude are
likely to be more effective at reducing property damage because
the overall damage is concentrated in a relatively small number
of incidents. Generally, when damage follows a power-law
distribution, reducing the frequency of incidents may not reduce
the overall pipeline risk due to the persistent possibility of
accidents with extreme damage.
7. Conclusions
The analysis of data pertaining to U.S. pipeline incidents from
1968 through 2009 reveals that the number of hazardous
liquid pipeline accidents has dropped dramatically over the
past four decades, resulting in a four-fold reduction in the
annual volume of hazardous liquid spills. However, injuries
and fatalities from hazardous liquid pipeline incidents have
remained relatively steady while property damage has
increased. In general, distribution pipeline incidents account
for the majority of injuries and fatalities, while hazardous
liquid and transmission pipelines tend to account for a larger
share of property damage.
In several cases, the magnitudes of incidents are well
characterized by power-law distributions, indicating that
reliable upper bounds for the severity of incidents cannot
be calculated. In such cases, catastrophic pipeline accidents
should be expected. The results indicate that the power-law
distribution is a plausible fit for injuries and volumes of
product spilled from hazardous liquid pipelines, injuries
and fatalities from distribution pipelines, and fatalities from
transmission pipelines.
Relatively few accidents involving transmission, distribu-
tion and hazardous liquid pipelines account for a large share
of total property damage. However, a large share of the total
fatalities and injuries are the result of single-fatality and
single-injury incidents. Finally, the computation of the value
at risk metric for pipeline accidents highlights the significant
difference in the expected losses when using a power-law
distribution instead of the more common thin-tailed distri-
bution. In particular, the results reveal that distinguishing
between safety measures that focus on reducing the number
of incidents and safety measures that seek to mitigate the
damage from incidents is very important in risk analysis
when “fat-tailed” distributions of damage are considered.
Acknowledgments
The authors wish to thank Edward S. Rubin, W. Michael
Griffin, Andrea Consiglio and Tim D. Gordon for helpful
discussions related to this research.
r e f e r e n c e s
[1] Association of Oil Pipe Lines, Report on Shifts in Petroleum
Transportation, 2011.
[2] Carnegie Mellon University, Learning from U.S. Energy Supply
Catastrophes, Technical Report, Department of Engineering
and Public Policy, Department of Social and Decision Sciences,
Pittsburgh, Pennsylvania, 〈www.epp.cmu.edu/undergraduate/
pdfs/2010/Energy%25%2020catastrophes-FINAL%20report%
203-16-11.pdf〉, 2011.
[3] A. Clauset, C. Shalizi, M. Newman, Power-law distributions in
empirical data, SIAM Rev. 51 (4) (2009) 661–703.
[4] F. Felder, A critical assessment of energy accident studies,
Energy Policy 37 (12) (2009) 5744–5751.
[5] S. Hirschberg, P. Burgherr, G. Spiekerman, R. Dones, Severe
accidents in the energy sector: Comparative perspective, J.
Hazard. Mater. 111 (1–3) (2004) 57–65.
[6] P. Jorion, Value at Risk: The New Benchmark for Managing
Financial Risk, McGraw-Hill, New York, 2007.
[7] H. Montiel, J. Vilchez, J. Arnaldos, J. Casal, Historical analysis
of accidents in the transportation of natural gas, J. Hazard.
Mater. 51 (1–3) (1996) 77–92.
[8] National Transportation Safety Board, Pipeline Special
Investigation Report: Evaluation of Pipeline Failures during
Flooding and of Spill Response Actions, San Jacinto River
near Houston, Texas, October 1994, Technical Report PB96-
917004, Washington, DC, 1996.
[9] Pipeline and Hazardous Materials Safety Administration,
Distribution, Transmission and Gathering, LNG and Liquid
Accident and Incident Data, Washington, DC, 2011.
[10] Pipeline and Hazardous Materials Safety Administration,
Natural Gas Transmission, Gas Distribution and Hazardous
Liquid Pipeline Annual Mileage, Washington, DC, 2011.
[11] W. Press, S. Teukolsky, W. Vetterling, B. Flannery, Numerical
Recipes in C: The Art of Scientific Computing, Cambridge
University Press, Cambridge, United Kingdom, 1992.
[12] C. Restrepo, J. Simonoff, R. Zimmerman, Causes, cost
consequences and risk implications of accidents in U.S.
hazardous liquid pipeline infrastructure, Int. J. Crit.
Infrastruct. Protect 2 (1–2) (2009) 38–50.
 international journal of critical infrastructure protection 7 (2014) 257 –269 269

[13] J. Simonoff, C. Restrepo, R. Zimmerman, Risk management
of cost consequences in natural gas transmission and
distribution infrastructures, J. Loss Prevent. Process Ind. 23
(2) (2010) 269–279.
[14] B. Sovacool, The cost of failure: a preliminary assessment of
major energy accidents, 1907–2007, Energy Policy 36 (5) (2008)
1802–1820.
[15] U.S. Energy Information Administration, Annual Energy Review
2009, Technical Report DOE/EIA-0384(2011), Washington, DC,
〈www.eia.gov/totalenergy/data/annual/pdf/%20aer.pdf〉, 2010.
[16] J. Xu, K. Kochanek, S. Murphy, B. Tejada-Vera, Deaths: final
data for 2007, Nat. Vital Stat. Rep 58 (19) (2010).