PLEASE SCROLL DOWN FOR ARTICLE

This article was downloaded by: [ABM Utvikling STM / SSH packages]
On: 21 July 2009
Access details: Access Details: [subscription number 792960683]
Publisher Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK
Ships and Offshore Structures
Publication details, including instructions for authors and subscription information:
http://www.informaworld.com/smpp/title~content=t778188387
Cost modelling of deepwater oil and gas facilities: A case study of spars and
tension leg platforms
C. Jablonowski
a
; A. Strachan
b
a
Department of Petroleum and Geosystems Engineering, the University of Texas at Austin, Austin, USA
b
Wood Mackenzie Canada Ltd, Calgary, Canada
First Published:March2009
To cite this Article Jablonowski, C. and Strachan, A.(2009)'Cost modelling of deepwater oil and gas facilities: A case study of spars
and tension leg platforms',Ships and Offshore Structures,4:1,69 76
To link to this Article: DOI: 10.1080/17445300802371139
URL: http://dx.doi.org/10.1080/17445300802371139
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.
Ships and Offshore Structures
Vol. 4, No. 1, 2009, 6976
Cost modelling of deepwater oil and gas facilities: A case study of spars and tension leg platforms
C. Jablonowski
a
and A. Strachan
b
a
Department of Petroleum and Geosystems Engineering, the University of Texas at Austin, Austin, USA;
b
Wood Mackenzie Canada Ltd, Calgary, Canada
(Received 15 February 2008; nal version received 23 July 2008)
Cost models for deepwater oil and gas facilities can be valuable tools for concept comparison and selection, eld development
planning and optimisation, and benchmarking performance. This study estimates simple cost models for spar and tension leg
platform projects using public and private data on 24 major projects. In addition to providing an analysis of the variables that
affect cost, the study investigates the complexity of model specication. We evaluate sensitivity to modelling assumptions,
sample selection bias and other model specication issues which can lead to improper conclusions.
Keywords: cost models; regression analysis; sample selection; spars; tension leg platforms
1. Introduction
Cost models for deepwater oil and gas facilities have a vari-
ety of uses. Cost models can be employed early in prospect
development to estimate expected lease values, to prepare
bidding strategies and to select prospects and manage the
portfolio. As exploration prospects mature into dened de-
velopment projects, cost models can be used to inform con-
cept comparison and selection, to aid in eld development
planning and optimisation and to benchmark projected and
actual costs.
Given these virtues, it is surprising to nd so little in
the offshore production facility literature on empirical cost
models or on cost-modelling methods.
1
We believe that the
primary causes of this gap are (i) the proprietary nature
of cost and technical data that greatly increases the effort
required to collect the data for detailed models and (ii) the
underlying nature of the facility selection decision process
that results in a non-trivial sample selection problem.
2
We
address both of these issues in this study. Firstly, we use
private cost data carefully organised by industry experts
based on public information and interviews with operating
companies. The specications are parsimonious, allowing
us to use publicly available technical data, but without a
sacrice in model explanatory power. Secondly, we dene
a constrained facility choice model that facilitates diagnosis
and treatment for sample selection.

Corresponding author. Email: cjablonowski@mail.utexas.edu

1
Publications that examine design optimisation and cost estimating for specic technologies and/or projects are abundant (see for example
Brooks and Carroll (1994), Zimmer (1994), and Stokes et al. (1996)). The gap we are referring to is in the area of aggregated analysis
across projects; the most recent published study known to the authors is Karlik (1991).
2
Sample selection problems in statistics and regression analysis occur when the sampling is not random. In this case study, the samples of
spars and TLPs are the result of a selection process assumed to be based on prot maximisation. Therefore, the observations cannot be
construed as a random sample and additional computations are required to correct for this feature of the data.
This case study estimates cost functions for spars and
tension leg platforms (TLPs) using data from 24 completed
projects. A two-stage regression model is specied that
accounts for the underlying facility selection process. The
remainder of this study is organised as follows. Section
2 provides a brief overview of oil and gas projects and
production facility selection, Section 3 species a complete
regression model of facility choice and cost that accounts
for sample selection bias, Section 4 denes the independent
variables and describes the data set and Section 5 provides
a conclusion.
2. Offshore oil and gas projects and production
facility selection
The general progression of an offshore oil and gas project
begins when an oil company acquires exploration and
production rights from the host government. The oil
company drills exploration and appraisal wells based on
geological data and analysis. If economic quantities of hy-
drocarbons are discovered, the prospect is developed with
additional production wells and associated surface- and
subsea-production facilities. A recent overview of the
project development and stage-gate project management
systems as applied in the oil and gas industry is available
in Walkup and Ligon (2006).
ISSN: 1744-5302 print / 1754-212X online
Copyright C
2009 Taylor & Francis
DOI: 10.1080/17445300802371139
http://www.informaworld.com
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
70 C. Jablonowski and A. Strachan
In deepwater settings, the selection and conguration
of production facilities is critical because of high costs and
the quasi-irreversible nature of the initial capital investment.
The analysis of competing options, or concept selection, is
often complicated because several different production fa-
cility types and congurations are feasible for a specic
project, and there is often considerable uncertainty regard-
ing subsurface and other project attributes. The concept
selection process typically consists of numerous scenario
and sensitivity analyses of the competing options. Under-
standing the concept selection process is relevant to cost
modelling. At rst, this link may not be obvious, but the
relationship and implications will be made clear in the spec-
ication of the cost models in Section 4.
In this study, we examine 24 deepwater projects where
either a spar or a TLP was ultimately selected as the primary
production facility. As shown in Figure 1, these projects
span a wide range of water depths and reserves (as measured
in million barrels of oil equivalent), and there is a substantial
overlap in the feasibility of the two production technologies.
While in very deep water, spars appear to hold an advantage,
and in the largest reserve cases, TLPs appear to dominate;
this observation does not exclude the possibility that the
alternate concept is technically feasible if not economically
competitive. Similar overlap is observed when the data is
plotted using other project attributes. For these reasons,
we assume that for all of these projects, both spar- and
TLP-production facilities were considered during concept
selection.
3. Model specication
A model of the decision to employ a spar or a TLP that
accounts for the underlying cost functions of each option
is specied. A latent (unobservable) variable y

i
is dened
which indicates that the propensity to employ a TLP on
project i:
y

i
= Z
i
+u
i
. (1)
However, y

i
is not observable; instead, one observes y
i
,
which takes on values of 0 (spar) or 1 (TLP) according to
the rule:
y
i
=
_
1, if y

i
> 0
0, otherwise
, (2)
where Z
i
is vector of prospect attributes, is vector of
coefcients for prospect attributes and u
i
is random error
term, N(0,
2
).
From here one can derive the standard binary probit
model:
Pr (y
i
= 1) = Pr
_
y

i
> 0
_
Pr (y
i
= 1) = Pr (u
i
> Z
i
) (3)
Pr (y
i
= 1) = Pr
_
u
i

> Z
i

_
.
Dividing by in Equation (3) is convenient because the
quantity u
i
/ is distributed as standard normal (zero mean
0
500
1000
1500
2000
0 50 100 150 200 250 300 350 400 450 500
Reserves (mmboe)
W
a
t
e
r

D
e
p
t
h

(
m
)
TLP Spar
Figure 1. Distribution of spar and TLP projects by water depth and reserves (24 projects).
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
Ships and Offshore Structures 71
and unit variance).
3
This leads to the following result:
Pr (y
i
= 1) =
_
Z
i

_
and therefore, Pr (y
i
= 0)
= 1
_
Z
i

_
, (4,5)
where () is the cumulative density function of the stan-
dard normal distribution. If there are n observations in total,
the likelihood function L for the sample is
L =
n

i=1
_
1
_
Z
i

__
1y
i
_

_
Z
i

__
y
i
(6a)
For purposes of optimisation, the log-likelihood func-
tion ln(L) = l is often employed:
4
l =
n

i=1
_
(1 y
i
) ln
_
1
_
Z
i

__
+y
i
ln
_

_
Z
i

___
(6b)
A complete model of the facility selection decision is
possible, based on the underlying costs of each development
concept. The decision maker develops expectations of these
costs during concept selection, based on intrinsic prospect
attributes and design criteria for each option. This can be
represented as follows:
Cost of spar-based project = S
i
= X
1i

1
+e
1i
(7)
Cost of TLP-based project = TLP
i
= X
2i

2
+e
2i
, (8)
where X
i
is a vector of independent variables for project
i, is a vector of coefcients and the e
i
s are uncorrelated
random error terms, N
_
0,
2
_
. All that is typically ob-
served, however, because data availability (the present case
is no exception) is the eventual choice of facility, y
i
:
y
i
=
_
1, if TLP
i
< S
i
0, otherwise.
(9)
Recall the variables of the function for the unobserved
latent variable, y

i
= Z
i
+u
i
. A binary choice model,
such as the probit, based only on the observed organisa-
tional choice is straightforward but only informs on the
relative difference in the coefcients of the underlying cost
3
The quantity u
i
_
is standard normal because u
i
has been trans-
formed by subtracting its mean and dividing by its standard devi-
ation.
4
The log-likelihood function is globally concave and is maximised
numerically. The asymptotic covariance matrix is computed by
evaluating the negative inverse of the Hessian at the maximum
likelihood estimates. Also observe that and are not individ-
ually identied; therefore, it is convenient to normalise to be
one.
equations. That is
Pr (y
i
= 1) = Pr (TLP
i
< S
i
)
= Pr (X
2i

2
+e
2i
< X
1i

1
+e
1i
)
= Pr (e
2i
e
1i
< X
1i

1
X
2i

2
) . (10)
If X
1
and X
2
have common elements and if there is un-
certainty about the effect of a specic independent variable,
even a very general interpretation of this models results be-
comes difcult. Therefore, a different approach is required
if one is to properly estimate coefcients in Equations (7)
and (8).
When one can observe costs, the cost equations can
be estimated directly. Such estimates, however, suffer from
selectivity bias because one cannot observe costs for the
option not selected. To show this, examine the expectation
of the error term of Equation (7):
E (e
1i
|u
i
< Z
i
) = E
_

e
1
,u
u
i
|u
i
< Z
i

_
=
e
1
,u
_
(Z
i
)
(Z
i
)
_
=
e
1
,u

1i
,
(11)
where and are the probability density function (pdf) and
cumulative density function (cdf) of the standard normal
distribution, respectively. Similarly,
E (e
2i
|u
i
> Z
i
) =
e
2
,u
_
(Z
i
)
(Z
i
)
_
=
e
2
,u

2i
(12)
Therefore, estimating cost models without accounting
for the selectivity bias would be inappropriate, unless it is
shown that the bias is insignicant. A summary of these
results and the underlying statistics is available in Maddala
(1983).
Results from the rst stage facility choice estimation
( ) can be used to compute the expected bias for each ob-
servation,
i
. This new variable is added as a regressor to
Equations (7) and (8), where its coefcient is an estimate
of
e
1
,u
and
e
2
,u
. This family of two stage models is gener-
ally attributed to Heckman (1979). Examples of empirical
studies employing this general approach are Masten et al.
(1991), who examine organisational choice in shipbuild-
ing, and Jablonowski and Kleit (2006), who examine the
decision to execute offshore drilling projects internally or
to use turnkey. In this study, the dependent variable in the
cost models is the total project cost, which includes the
host facility, the subsea system and the initial development
wells.
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
72 C. Jablonowski and A. Strachan
4. Data collection and diagnostics
The dependent variable in the cost models is the total
project cost. The cost data was obtained from a private
database of E&P projects maintained by a global E&P con-
sulting practice in project analysis and economics.
5
All
spars and TLPs included in the nal data set were lo-
cated in the US Gulf of Mexico and were installed be-
tween 1986 and 2006. Costs were adjusted to constant 2006
US dollars.
6
Costs were assumed to occur in equal annual
amounts between the start and the end year of project de-
velopment. More rened cost adjustment is possible, but
it requires itemised costs, detailed information about the
timing of capital spending, and drilling schedules and rigs
used. An investigation by the authors into the feasibility
of organising such a database concluded this effort to be
prohibitive.
The independent variables for the model (choice and
cost functions) that are believed to inuence the cost and the
choice of the production facility are now dened. This list
is not comprehensive of all potential independent variables
(we are constrained by the availability of data), but we
believe the most important variables are captured in this
data set.
Water depth is dened as the water depth at the site of
the host facility measured in metres. The primary impact
of water depth is on the cost of TLP tendons and the cost
of mooring-related components for both facility types.
Oil rate is dened as the maximum oil production de-
sign rate (nameplate) of the facility measured in thousand
barrels per day (mbopd). Gas rate is dened as the maxi-
mum gas production design rate of the facility measured in
million cubic feet per day (mmcfpd). Higher design rates
result in more piping, tanks and related production equip-
ment, thus increasing the topside weight and the cost of the
production facility.
Topside weight is primarily a function of maximum to-
tal liquids design rate (oil plus produced water). But water
production rates tend to increase as oil production rates de-
crease, and the maximum design rates for oil and produced
water are typically not occurring simultaneously. Therefore,
using both maximum design rates to model cost will be
inappropriate in most cases. Our approach is to use the
maximum oil production rate as a proxy for maximum total
liquids design rate.
7
5
Cost data obtained from Wood Mackenzie, Houston, Texas.
6
Costs adjusted using the US Department of Commerce Producer
Price Index for Oil and Gas Field Machinery and Equipment,
Series ID: PCU333132333132.
7
We do not use maximum total liquids design rate directly be-
cause of problems validating some of the water production rate
data. Based on validated data, we nd high correlations between
topsides weight and oil rate, typically greater than 0.85, and we
conclude that oil rate is an acceptable proxy for total liquids design
rate.
Slot count is dened as the number of well slots avail-
able on the host platform for dry trees.
8
As the slot count
increases, additional space and weight capacity are required,
increasing the size and the cost of the production facility.
Rig type is a binary variable that takes on a value of
1 when an integrated drilling rig is included in the facility
design and 0 otherwise. Deepwater production facilities can
be built with an integrated drilling or workover rig,
9
or with
the capability to host a temporary drilling or workover rig,
either wholly on the platform or via tender-assist. When an
integrated drilling rig is included in facility design, costs
increase relative to all other platform-based rig options.
Spar type is a binary variable that takes on a value of 1
when the design is a classic spar and 0 otherwise. The choice
of the spar type inuences performance characteristics and
cost.
Reserves is dened as the estimated reserves that are
to be produced through the host facility from the primary
reservoirs measured in million barrels oil equivalent (mm-
boe). Large elds can produce at higher rates, all things
being equal, and this affects the choice of facility type,
production capacities and cost.
Topside weight is dened as the weight of the deck and
production facilities measured in metric tons (mt). Topside
weight is primarily a function of the quantity of steel em-
ployed. As topside weight increases, there is a direct effect
on the cost of the topsides and an indirect effect on the size
and cost of the hull.
Data on the independent variables was obtained from
Nutter and Albaugh (2005, 2006) and from the operat-
ing company and other industry websites. Different data
sources occasionally contain conicting data. Reasonable
effort was made to conrm all data. If a data entry could
not be conrmed across multiple sources, the observation
was excluded from the nal specications. Summaries for
the spar and TLP data sets are reported in Tables 1 and 2.
Correlation among the independent variables increases
the potential for collinearity-related estimation problems.
In the spar data, for example topside weight is correlated
with oil rate (0.85), rig type (0.96) and reserves (0.90). The
TLP data exhibit a similar structure. Tables 3 and 4 re-
port the correlation coefcients for the spar and TLP data.
Choosing which variables to retain and which to exclude
is not an exact science. One must compare the theoretical
8
When a wellhead is located on the host platform, it is called a dry
tree. Access to the well for maintenance and remedial operations
is accomplished using a drilling or workover rig located on the
platform. For reservoirs where frequent well interventions are
anticipated, dry trees are generally preferred as is an integrated
rig. When a wellhead is located on the seaoor, it is called a
wet tree. Access to the well is accomplished using a oating rig
(e.g. semisubmersible rig). Because of the high cost and uncertain
availability of oating rigs, wet trees are generally preferred when
infrequent well interventions are anticipated.
9
Supra.
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
Ships and Offshore Structures 73
Table 1. Summary of spar data (11 projects).
Standard
Variable Units Type Mean Deviation
Total cost $2006 Continuous 694.2 280.0
Water depth m Continuous 1105.2 348.9
Oil rate mbopd Continuous 59.8 26.0
Gas rate mmcfpd Continuous 154.1 80.3
Slot count slots Continuous 11.5 5.1
Rig type Binary 0.27
Spar type Binary 0.27
Reserves mmboe Continuous 160.0 81.8
Topside weight mt Continuous 7529.1 4836.8
arguments for competing specications and weigh the ben-
ets of including additional independent variables against
the costs of potentially degenerate results.
10
This process is
more complicated when the sample size is small as it is in
this study. The nal specications reported here represent
a careful balancing of these issues. Because of collinear-
ity problems, topside weight is excluded from all of the
nal model specications. Reserves was considered in the
choice model but was ultimately excluded for collinearity
and other reasons.
5. Analysis and results
In this section, we estimate the model as specied above,
starting with the choice model and then use the results to
correct for selectivity bias in the cost functions. The depen-
dent variable in the facility choice model is binary, and it
takes on a value of 1 for projects where a TLP was installed
and 0 for projects where a spar was installed. The appro-
priate variables to include in the choice model are those
that would be dened (xed) in a design basis. An example
of this type of variable is water depth. Variables that may
have been different between competing concepts must be
excluded.
11
This feature of the model complicates the spec-
ication. We discuss a fewvariables in detail to demonstrate
this point prior to dening the nal specications.
The number of well slots on the host facility, or slot
count, can be constrained by the spatial distribution of re-
serves, that is if the reserves are dispersed across a wide
area, fewer wells can be drilled from a common host.
10
When highly correlated independent variables are included in a
regression, the variances of the coefcient estimates tend to in-
crease, resulting in low t-statistics. That is, two variables known
to positively inuence the level of the dependent variable can be
found to be statistically insignicant. Another awkward outcome
may assign opposite signs to the variables, canceling out their
effect. Thus, judging whether a regression result is reliable de-
pends on the analysts knowledge of the technical relationships
and experience.
11
This constraint results because we do not have data on design
details for unselected production facility options.
Table 2. Summary of TLP data (13 projects).
Standard
Variable Units Type Mean deviation
Total cost $2006 Continuous 806.5 459.4
Water depth m Continuous 865.1 275.0
Oil rate mbopd Continuous 75.6 53.8
Gas rate mmcfpd Continuous 180.8 130.4
Slot count slots Continuous 12.7 10.8
Rig type Binary 0.31
Reserves mmboe Continuous 161.0 164.4
Topside weight mt Continuous 10,132.8 7977.7
Also, reservoir properties affect expectations regarding
production problems and the frequency of well interven-
tions, and this can constrain the slot count. While these
intrinsic prospect variables inuence the slot count, the
constraints are not always binding, and in many cases
the operator can compare multiple congurations of each
type of the production facility. In these cases, the slot count
may vary between competing concepts, and therefore it
would be inappropriate to include the actual slot count
(from the selected option) in the choice model. Without
intimate knowledge of each prospect, it is not possible to
conclude denitively which condition (more or less con-
strained) holds more often in practice. In the base case, we
assume that the slot count is more often intrinsic to the
prospect rather than a tradeoff variable, and we include it
in the choice model.
Variables such as the total number of wells (dry and wet
tree) and production rates pose a similar problem. In many
cases, the depletion plan for a eld is highly constrained
by the spatial distribution of reserves and initial reservoir
conditions. In such cases, there is little exibility in the
depletion plan, and deviations result in signicant losses in
recovery rates and project value. In other cases, the deple-
tion plan may be more exible, and the operator can assess
various project scales for each concept; in these cases, it
would be inappropriate to include production rates in the
choice model for the reasons already cited. Again, it is not
possible to conclude from the data available which condi-
tion holds more often. In the base case, we assume that oil
rate and gas rate are more often intrinsic to the prospect,
and we include both in the choice model.
Because of the high correlation between oil rate and
reserves, we do not include both of these variables in the
choice model. Reserves would be the preferred variable to
retain for the choice model because it clearly is intrinsic to
the project, and one would expect facility choice to depend
on reserves as described in the denition of this variable
above. But in preliminary specications, we found that this
variable is generally not statistically signicant. Our hy-
pothesis for this result is twofold. One possibility is that the
result is sound, and reserves are in fact not signicant. As
seen previously in Figure 1, there is a signicant overlap in
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
74 C. Jablonowski and A. Strachan
Table 3. Correlation coefcient matrix for spar data (11 projects).
Water depth Oil rate Gas rate Slot count Rig type Spar type Reserves Topside weight
Water depth 1.00
Oil rate 0.65 1.00
Gas rate 0.43 0.36 1.00
Slot count 0.21 0.21 0.61 1.00
Rig type 0.16 0.73 0.23 0.52 1.00
Spar type 0.29 0.09 0.01 0.35 0.54 1.00
Reserves 0.49 0.88 0.46 0.17 0.78 0.20 1.00
Topside weight 0.28 0.85 0.39 0.32 0.96 0.35 0.90 1.00
Table 4. Correlation coefcient matrix for TLP data (13 projects).
Water depth Oil rate Gas rate Slot count Rig type Reserves Topside weight
Water depth 1.00
Oil rate 0.26 1.00
Gas rate 0.54 0.47 1.00
Slot count 0.13 0.57 0.57 1.00
Rig type 0.29 0.49 0.67 0.53 1.00
Reserves 0.34 0.87 0.68 0.75 0.52 1.00
Topside weight 0.48 0.90 0.65 0.52 0.91 0.78 1.00
facility choice throughout the range of reserves. Even if a
relationship exists, it may not be measurable in a statistical
sense. A second possibility relates to data quality. Across
all sources, the data on oil and gas reserves is the most in-
consistent of all of the variables in the study. Reserves data
can be highly proprietary and may not be reported in an
unbiased manner. It is also frequently updated throughout
the early stages of a project. For these reasons, we employed
production rates instead of reserves.
Similar arguments, in addition to the aforementioned
collinearity issues, were evaluated for the remaining inde-
pendent variables. The base case specication of the choice
model contains four independent variables: water depth,
oil rate, gas rate and slot count. We do not have theoreti-
cally derived hypotheses for the signs of coefcients in the
choice model. But based on the distribution of projects in
Figure 1, one might expect that the sign for water depth
to be negative and the signs for oil rate and gas rate to be
positive.
The results of the choice model are reported in Table 5.
Because there is a signicant overlap in the technically
feasible range of applications, it is not surprising that the
results on individual variables are generally weak. The ex-
ception is water depth, which appears to be the driver for
facility choice, all things being equal. We suspect that, given
a larger data set, the signicance of oil rate and gas rate
would increase. The coefcient signs also conform to the
expectations mentioned in the preceding paragraph. While
the likelihood ratio index (pseudo R
2
) is low at 0.24, the
model correctly predicts facility choice in 79% of the ob-
servations. We use this base case specication of the choice
model to compute the expected selectivity bias.
As dened above, the dependent variable in the cost
models is the total project cost, including the host facility,
the subsea system and the initial development wells. In
the cost models, the expectation for the sign of all of the
Table 5. Regression results for choice model (binary
probit).
Variable Coefcient t-statistic
Constant 1.8466 1.61
Water depth 0.0029 2.33
Oil rate 0.0123 1.14
Gas rate 0.0043 1.13
Slot count 0.0383 0.86
n = 24; Log likelihood = 12.54841; Likelihood ratio index
= 0.24.
Table 6. Linear cost model (spars).
Variable Coefcient t-statistic
Intercept 258.8573 1.35
Oil rate 6.0876 3.40
Gas rate 1.1419 1.70
Slot count 30.4256 2.88
Spar type 146.3163 1.78

e
1
,u
33.8045 0.27
n = 11; R2 (adj) = 0.88.
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
Ships and Offshore Structures 75
Table 7. Linear cost model (TLP).
Variable Coefcient t-statistic
Intercept 54.8080 0.66
Oil rate 3.1686 4.22
Gas rate 1.2924 4.09
Slot count 15.7242 4.30
Rig type 121.7436 1.42

e
2
,u
75.9172 1.05
n = 13; R2 (adj) = 0.95.
coefcients is positive.
12
The results from the estimation of
Equation (7) for spars are reported in Table 6.
The overall t of the model is good and yields an ad-
justed R
2
of 0.88. Oil rate and slot count are signicant
at the 95% level; gas rate and spar type are signicant at
the 85% level. All coefcients have the expected sign. The
selection bias is not signicant.
Water depth is excluded from the nal regression be-
cause it was found to be statistically insignicant across
various specications. The incremental cost of mooring in
deeper water, while positive, is considerably smaller than
the incremental cost of other project attributes and is not
measurable in a statistical sense.
Rig type is excluded because of problems encountered
including it and oil rate in the same specication (the corre-
lation coefcient between the two variables is 0.73). Also,
the rig type and spar type variables are identical except for
two observations, even though the correlation coefcient is
computed to be 0.54 (this results because both variables are
binary, thus reducing the normal usefulness of the corre-
lation coefcient in detecting collinearity). Given that oil
rate and spar type exhibit a low correlation (0.09), rig type
was dropped. The results from the estimation of Equation
(8) for TLPs are reported in Table 7.
The overall t of the model is very good, yielding an
adjusted R
2
of 0.95. Oil rate, gas rate and slot count are
signicant at the 95%level; rig type is signicant at the 80%
level. All coefcients have the expected sign. The selection
bias is signicant at the 70% level.
As it was the case in the spar regression, water depth is
excluded from the nal regression because it was found to
be statistically insignicant across various specications.
This is a somewhat surprising result given the cost of
tendons. Upon further inspection, however, we observe that
two thirds of the TLP observations are clustered within
a 300 metre range. This concentration of data, and the
small number of observations, is the likely cause of this
result.
The coefcients from each of these models can be com-
pared directly. For example, the marginal cost of 1 mbopd
12
There is no expectation for the sign of the selection bias coef-
cient.
of oil capacity on a spar costs about $6 million, whereas
on a TLP the same capacity costs about $3 million. The
marginal cost of gas capacity is similar for each facility.
Each additional well slot on a spar costs about $30 million,
whereas on a TLP it costs about $16 million. Also, a classic
spar design adds almost $150 million to the expected cost,
and an integrated drilling rig on a TLP adds $122 million.
Recall that costs are in 2006 US dollars. If the model is to
be used in a predictive mode, additional adjustment would
be required to account for recent increases in costs. From
2006 to mid-2008, there has been a 16% increase in the
index used for adjustment in this study.
We can compare these results to cost models estimated
without the correction for the sample selection bias. In
the case for spars, the differences in coefcient estimates
are insignicant. This is the expected result given the
statistical insignicance of the bias coefcient. In the case
for TLPs, the intercept term increases by 112% and the
coefcients for oil rate and rig type are reduced by 11%
and 10%, respectively. The remaining coefcients are
essentially unchanged. In most cases, the difference in the
residuals between the two models is small and would not
be material with respect to decision making. But in many
cases, it exceeds $20 million and the maximum difference
equals $85 million. If the sample selection bias is taken
to be signicant, using the uncorrected model during
concept selection could result in an incorrect decision. In a
benchmarking setting, use of the uncorrected model could
lead to an incorrect assessment of the estimated or actual
project cost.
As an alternative to the base case, we also estimated a
choice model using water depth as the only independent
variable. This model represents the opposing arguments to
those of the base case choice model, that is that production
rates and the number of slots are not intrinsic to the project.
The results were then used to estimate the sample selection
bias and to estimate the cost functions. Notwithstanding the
relatively weaker explanatory power of the variable choice
model (71% correct predictions; likelihood ratio index =
0.11), the results for the cost functions were very similar
in all respects to the results based on the base case choice
model. This suggests that in this setting, the arguments and
assumptions regarding the appropriate variables to employ
in the choice model are not critical to the estimation of
the cost functions. It is not clear whether this result can
be generalised, and we intend to investigate this question
formally in a future study.
6. Conclusions
This case study estimates cost functions for spars and TLPs
using data from 24 completed projects. A two-stage re-
gression model is specied that corrects for the sample
selection bias by accounting for the underlying facility se-
lection process. We nd that the sample selection bias is
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9
76 C. Jablonowski and A. Strachan
not a signicant problem in cost models for spars, but it
is potentially signicant in cost models for TLPs. The cor-
rected cost models are reported for both facility types and
can be used to value and select prospects, to inform con-
cept comparison and selection, to aid in eld development
planning and optimisation and to benchmark projected and
actual costs.
The regression theory and technique employed in this
study are somewhat advanced, but engineers with some
background and interest in statistics will be capable of im-
plementing this approach. Also, these models are available
in most commercial regression packages. But specifying
models such as this is not a minor undertaking. The data
requirements are signicant, and technical expertise and
understanding of the decision being evaluated is required.
Diagnostics during the preliminary regression stage are also
critical and require some experience in applied regression
analysis.
Acknowledgements
The authors thank Jerry Kepes, who inspired the original
formulation of the problem, and Bill Lamport for assistance
on the design aspects of spars and TLPs. John Lynaugh,
Kannika Yodpetchpongsri, Cengizhan Yenerimand Chaiya-
porn Wiboonkij-Arphakul provided excellent research as-
sistance. The authors also thank Wood Mackenzie for pro-
viding industry cost data. Partial nancial support for this
project was provided by the Cockrell School of Engineer-
ing at the University of Texas at Austin. All errors and
omissions remain the responsibility of the authors.
References
Brooks I, Carroll J. 1994. Drilling and production from a oating
spar. Paper SPE 28699-MS presented at: The SPE Interna-
tional Petroleum Conference and Exhibition of Mexico; Oct
1013; Veracruz, Mexico.
Heckman J. 1979. Sample selection bias as a specication error.
Econometrica. 47(1): 153162.
Jablonowski C, Kleit A. 2006. Transaction costs and risk prefer-
ences: modeling governance in offshore drilling. Paper SPE
101936 presented at: The SPE Annual Technical Confer-
ence and Exhibition; Sep 2427; San Antonio, Texas. doi:
10.2118/101936-PA
Karlik C. 1991. Parametric estimating of oil and gas production
facilities capital costs worldwide. Paper SPE 22015 presented
at: The SPE Hydrocarbon Economics and Evaluation Sympo-
sium; Apr 1112; Dallas. doi: 10.2118/22015-MS
Maddala GS. 1983. Limited dependent and qualitative variables in
econometrics. Cambridge (UK): Cambridge University Press.
Masten SE, Meehan JW, Snyder EA. 1991. The costs of organiza-
tion. J Law Econ Organ. 7(1): 125.
Nutter T, Albaugh E. 2005. Worldwide survey of TLPs, TLWPs.
Offshore Mag, October.
Nutter T, Albaugh E. 2006. Worldwide survey of spars, DDCVs.
Offshore Mag, November.
Stokes E, Koon J, Thompson D. 1996. Tailoring the TLP-evolution
and adaptation of the concept. Paper SPE35298-MSpresented
at: The SPE International Petroleum Conference and Exhibi-
tion of Mexico; Mar 57;Villahermosa, Mexico.
Walkup GW Jr, Ligon JR. 2006. The good, the bad, and the ugly
of the stage-gate project management process in the oil and
gas industry. Paper SPE 102926 presented at: The SPE An-
nual Technical Conference and Exhibition; Sep 2427; San
Antonio, Texas. doi: 10.2118/102926-MS
Zimmer R. 1994. The micro TLP production system for deepwa-
ter Gulf of Mexico. Paper SPE 28681-MS presented at: The
SPE International Petroleum Conference and Exhibition of
Mexico; Oct 1013; Veracruz, Mexico.
D
o
w
n
l
o
a
d
e
d

B
y
:

[
A
B
M

U
t
v
i
k
l
i
n
g

S
T
M

/

S
S
H

p
a
c
k
a
g
e
s
]

A
t
:

0
9
:
4
3

2
1

J
u
l
y

2
0
0
9