Evaluating Fracture-Fluid Flowback in Marcellus Using Data-Mining Technologies
Evaluating Fracture-Fluid Flowback in Marcellus Using Data-Mining Technologies
Table 1—Summary of mean and data range for stimulation completion parameters and production.
Data-mining approaches can avoid issues of physical assump- with long, calculated shut-in times may be attributed to a delay in
tions, such as those made from the preceding simulation studies, well connection to production line. Volumetric flowback-recovery
and extract the implicit and interesting information from real pro- efficiency was calculated as cumulative recovered fracture-fluid
duction data in a large database. Data-mining approaches have volume after 3 weeks divided by total injected fracture-fluid vol-
been applied extensively for gas/oil-production analysis in shale ume. Three weeks was used because of the availability of data, but
reservoirs (Centurion 2011; Centurion et al. 2012, 2013; Cunning- 3-weeks cutoff could be referred to the work of Barbot et al. (2013)
ham et al. 2012; Esmaili et al. 2012a, 2012b; LaFollette and Hol- and Seales (2015), as well. According to Seales’ model, approxi-
comb 2011; LaFollette et al. 2012; Shelley et al. 2008), but such mately 80% of recoverable fracture fluid was flowed back within
studies were found rarely in the literature for fracture-fluid-flow- the first 3 weeks. The conclusive remarks achieved in this paper are
back analysis. Lan et al. (2014) performed experimental analysis based on 3 weeks of flowback data. As longer periods of flowback
for fracture-fluid flowback in shale reservoirs on the basis of sam- data are collected in the future, this analysis could be expanded to
ples from limited wells from shale members of the Horn River develop more-comprehensive insights. Finally, it should be noted
Basin and Montney tight gas formation. The controlling factors that all the wells in this paper are multistaged horizontal wells
for fracture-fluid-flowback volume change have not been identi- stimulated by slickwater fracturing.
fied clearly by previous published research, and correlations Before proceeding with data analysis, the outliers were
between fracture-fluid flowback and geology, completion, and detected by use of the Mahalanobis distance approach (Mahalano-
stimulation parameters have not been investigated adequately. bis 1936), which is defined as the distance between a data point
This paper is focused on assessing trends in flowback-recovery and the centroid of a multidimensional space, with the centroid
efficiency through data-mining techniques. The analysis includes determined as a function of the mean values of all variables.
two parts: univariate engineering analysis used to identify correla- Points with the Mahalanobis (1936) distance larger than a certain
tions of completion and stimulation parameters that impact flow- cutoff were treated as outliers. The outliers are assumed to bias
back-recovery efficiency and a multiple-variable linear-regression the analysis result and were removed from the data set. After this
model with variable-importance calculation used to identify the screening, the data set decreased from 187 to 173 observations.
controlling engineering factor that impacts flowback-recovery ef- However, the 173 wells retained after this screening were further
ficiency. In the second part, the effects from geological conditions scrutinized to identify the set of wells most appropriate for use in
were examined by principal-component analysis (PCA) and clus- this analysis. Some wells calculated shut-in periods of negative
tering analysis. The analysis procedures and methods used in this duration, those observations were removed. As a result, the total
paper are discussed in the following section. well count is reduced to 86; all the following analyses are based
on those 86 wells.
According to the thermal-maturity (indicated by vitrinite re-
Data Gathering, Screening, and Grouping flectance Ro) distribution and the formation-thickness (h) distribu-
Details of the data-set development and screening criteria are pro- tion (Bruner and Smosna 2010), we divided the study domain into
vided in Zhou et al. (2014). Completion parameters detailed in the four groups (Fig. 1). The data set was first partitioned as Group A
database include proppant type, proppant mass, fracture-fluid type, and Group B by Ro ¼ 1.6. The green region, Group A, represents
fracture-fluid volume, vertical depth, lateral length, number of the region in which Ro falls between 1.0 to 1.6, and the production
stages, number of perforation shots, perforation intervals, treatment should contain natural gas and oil together. The yellow region,
rate, average treatment pressure, maximum treatment pressure, in- Group B, represents the region in which Ro ranges between 1.6
stantaneous shut-in pressure (ISIP), acid volume, and shut-in time. and 2.5, and production is expected to contain natural gas only.
Because of insufficient data in some wells, only the fracture-fluid The white area is not considered in this study. Group A is then
volume, proppant mass, vertical depth, lateral length, treatment partitioned into Group A1 and Group A2 by Ro value equal to 1.3,
rate, number of stages, and shut-in time were considered as varia- above which a large amount of gas is known to be generated.
bles for this study. Table 1 shows the mean and ranges for each of Group B is subpartitioned into Group B1 and B2 by h ¼ 80 ft. The
those parameters over the entire spatial domain and in each geologi- detailed partition rules and principles are discussed in Zhou et al.
cal group. The shut-in time is defined herein as the number of days (2014). Table 2 lists the number of wells and the thermal-maturity
between end of treatment and beginning of flowback. Some wells range and thickness range in each geological group.
W E
S Marshall
Monongalia
Wetzel A2
Tyler A1 Marion Preston
Taylor
Doddridge
Harrison
Barbour
B1
B2
Upshur
Allegheny Front
0 25 50 100
Miles
Fig. 1—Groups A1, A2, B1, B2 distribution on the map of West Virginia. Data set was first partitioned as Groups A and B by
Ro 5 1.6; Group A was subpartitioned by Ro 5 1.3; Group B was subpartitioned by h 5 80 ft.
Start
A (52) B (34)
PCA MLR
K-Means LMG
Finish
averaging sequential sums of squares over all orderings of regres- each cluster are calculated. Distances between each observation
sors (predictors), is commonly applied to observational data prob- and the centroid are calculated and summed. The final clustering
lems such as that considered herein. In this context, “sequential” result is achieved when the sum of the distances between
means that the regressors are entered into the model in the order observed value and centroid in each cluster is at minimum after
in which they are listed. Division of sequential sums of squares by several iterations. This optimization process can be summarized
the total response sum of squares yields sequential R2 contribu- by Eq. 3, where jCkj means the number of clusters and (xij – xi0 j)
tions. This method takes into account the dependence on order- is the Euclidean distance between every two observations in a
ings by averaging over orderings, using simple unweighted k cluster:
averages. Eq. 2 shows the corresponding function, where lmg(xk) X
is the variable importance of xk calculated by the LMG method, K 1 X Xp 2
minimizeC1 …Ck i;i0 2C ðx ij x i0jÞ :
seqR2 is the sequential sum of squares, n is the data size, p is the k¼1 jC j
k
k j¼1
number of predictors, and S is the data set excluding xk: ð3Þ
1
lmgðxk Þ ¼ PCA. PCA refers to the processes that compute the principal
p!
X components and perform analysis on the basis of these compo-
2
Sðx1 ;…xp Þnðxk Þ nðSÞ!½ p nðSÞ 1!seqR ½ðxk ÞjS: ð2Þ nents in understanding the data pattern. The principal components
are a relatively small subset of the original set of variables, and
collectively explain the majority of variance of the data set. The
Clustering Analysis. Clustering analysis is used to identify sub- generation of principal components on the basis of the theory of
groups (clusters) in a data set by partitioning objects with similar matrix decomposition and eigenvector is to find a low-dimen-
characteristics. Because it is often applied to problems in which sional set of features of the data set. The PCA can be used as
the number of groups is not known a priori, many clustering dimension reduction for visualization, such as 2D visualization.
approaches prescribe criteria to objectively define what it means This function is applied in this study in combination with the clus-
for two or more observations to be similar or different. Commonly tering method.
used clustering algorithms include K-means, hierarchical cluster-
ing, spectral clustering, and density-based clustering. In this
study, the K-means clustering method is applied to partition the Results and Discussion
set of 86 screened observations into subgroups by stimulation and Single-Parameter Study. A set of simple univariate analyses
completion parameters because of its simplicity and efficiency. In was performed to examine the correlation of each parameter with
K-means clustering, the data are initially randomly assigned to K the cumulative water recovery within a 3-week flowback to find
clusters, and centroids representing the mean of all variables in potentially statistically significant correlations between the 3-
10 12
8.75
9.71
10
8
2
2
0 0
Group A Group B Group B1 Group B2
Fig. 3—Three-week-flowback recovery between Group A (52 Fig. 4—Three-week-flowback recovery between Group B1 (17
cases) and Group B (34 cases); the flowback recovery increased cases) and Group B2 (17 cases); the flowback recovery
as Ro increased. decreased as h increased.
week fracture-fluid flowback and well geologic, stimulation, and Group B supports the decision to define this geological boundary.
completion features. Fig. 4 shows the average 3-week-flowback recovery within Group
B. It seems that as the thickness increases, the flowback recovery
decreases. To further understand the relationship between forma-
Geological Parameter. The difference in geological setting could tion thicknesses and flowback recovery, the stimulation and com-
have significant effects on flowback performance. Thermal matu- pletion parameters need to be considered.
rity, used as a practical index for the hydrocarbon composition and
hydrocarbon phases in the reservoir, is important for the 3-week-
flowback evaluation. In addition to thermal maturity, formation Engineering Parameter. Stimulation operational choices were
thickness also significantly impacts 3-week-flowback performance. also analyzed to explore which engineering parameters have sig-
Fig. 3 shows a comparison of the results on 3-week-flowback nificant influence on the flowback recovery (magnitude of correla-
recovery in Groups A and B. In Group A, the average 3-week- tion as absolute value), and the type of correlation (positive or
flowback recovery is 4.46%, while in Group B, the average is negative). Fig. 5 shows results of that univariate correlation anal-
8.75%. This relatively large difference between Group A and ysis; the coefficients of correlations (r) between single-engineering
30 30 30
Recovery Efficiency (%)
10 10 10
0 0 0
0 2 4 0 5 10 0 5,000 10,000
Total Fracture Fluid (bbl) ×105 Total Proppant (lbm) ×106 Vertical Depth (ft)
30 30 30
Recovery Efficiency (%)
20
20 20
10
10 10
0
0 0
6,000 8,000 10,000 0 20 40 50 100 150
Lateral Length (ft) Number of Stages Treatment Rate (bbl/min)
30
Recovery Efficiency (%)
20
10
0 100 200
Shut-In Time (days)
Fig. 5—Analysis of single-engineering parameter on 3-week-flowback recovery; shut-in time has a negative correlation with flow-
back recovery (86 cases).
Table 3—Correlation coefficient and P-value between completion parameters and flowback recovery through 3 weeks post-stimulation.
parameters (total fracture-fluid volume, proppant mass, vertical Ranking of Engineering Parameters. To identify key engineer-
depth, lateral length, number of stages, treatment rate, shut-in time) ing factors affecting the well flowback performance, multiple-
and 3-week-flowback recovery are presented in Table 3. regression analysis between 3-week flowback and engineering pa-
Those results show that the shut-in time has the largest coeffi- rameters within each geological group is implemented. In this
cient of correlation compared with other parameters, followed by analysis, it is assumed that within each geological group, the geo-
lateral length, with r value as –0.3828 and 0.3713, respectively. logical characteristics are similar and the wells are operated using
The vertical depth, proppant mass applied, and lateral length ex- the same strategy. Because the engineering-parameter units vary
hibit positive correlation with the 3-week-flowback recovery (i.e., significantly between wells, a scaling process is applied before
the flowback recovery increases as those parameters increase). multiple-regression analysis is performed. To make the compari-
Other parameters, such as fracture-fluid volume, number of frac- son more meaningful, the total fracture-fluid volume, total prop-
ture stages, treatment rate, and shut-in time, have a negative rela- pant mass, and lateral length are divided by the number of stages.
tionship with the 3-week-flowback recovery (i.e., the flowback Table 4 shows the correlation between each factor with the water-
recovery decreases as these parameters increase). Fracture-fluid production performance by the sign and value of coefficient. The
volume, proppant mass, vertical depth, and number of stages have multiple-regression equation for each group is shown as Eq. 4. To
P-values larger than 0.05, and therefore are determined to not overcome the range problem of flowback recovery, the logistic
have a statistically significant relationship with the flowback re- transformation was applied (Kutner et al. 2004). Eq. 4 could be
covery. All of those correlations are weak (even the ones with P- applied to estimate 3-week flowback with a level of confidence,
value greater than 0.05). Still, this analysis provides some insights as shown in Table 4.
into the general trend between each stimulation or completion pa- 0 1
rameter and the 3-week-flowback recovery of Marcellus shale in
northeastern West Virginia. B recovery C
log@ recoveryA ¼ b0 þ b1 Fracture Fluid=Stage
1
100
Multiple-Parameter Study. The lack of strong correlation
between the reservoir and stimulation characteristics and well þ b2 Proppant Mass=Stage þ b3 Lateral Length=Stage
flowback recovery from univariate correlation analysis may stem þ b4 Vertical Septh þ b5 Average Rate þ b6 Shut-in Time:
from a lack of data normalization for geologic and engineering ð4Þ
parameters and the colinearity among those variables. In an
attempt to overcome these issues and identify meaningful patterns For each geological group, the variable importance was calcu-
in the data, multivariable regression, LMG, PCA, and clustering lated as shown in Appendix A. The factor of greatest importance
methods are considered in combination. for each of the four defined geological subgroups and the corre-
sponding variable-importance value are shown in Fig. 6. The
most important engineering factor for Groups A1 and B2 is the
shut-in time; for Group A2 and Group B2, the key factors are
0.6
Shut-in time Shut-in time proppant mass/stage and fracture-fluid volume/stage, respectively.
Proppant/stage To further understand what causes the different dominated en-
0.5
Fracture gineering-factor results, the proppant size and proppant concentra-
Variable Importance
40% 40%
A1 36.11% A2 37.14%
35% 35%
30% 27.78% 30%
Frequency 25% 25%
Frequency
22.22%
20% 20% 17.14% 17.14%
15% 14.29%
15%
10% 10%
5.56% 5.56%
5% 2.78% 2.78% 5% 2.86% 2.86% 2.86% 2.86% 2.86%
0% 0%
0.80 0.93 1.07 1.20 1.33 1.47 1.60 0.40 0.53 0.67 0.80 0.93 1.07 1.20 1.33 1.47
Proppant Concentration (lbm/gal) Proppant Concentration (lbm/gal)
Fig. 7—Proppant-concentration distribution for Group A1 (28 Fig. 8—Proppant-concentration distribution for Group A2 (24
cases). Approximately 86% of wells in this group with prop- cases). Approximately 17% of wells in this group with small
pant-concentration ranges from 1.2 to 1.47 lbm/gal. proppant-concentration of 0.4 lbm/gal, which is not observed in
the other three groups.
35%
B1 31.03%
30% 35%
B2 30.43%
25% 30%
26.09% 26.09%
20.69% 20.69%
Frequency
20% 25%
Frequency
15% 20%
10% 15%
6.90% 6.90% 6.90%
5% 3.45% 3.45% 10%
4.35% 4.35% 4.35% 4.35%
0% 5%
0.67 0.80 0.93 1.07 1.20 1.33 1.47 1.60 0.00%
0%
Proppant Concentration (lbm/gal) 0.53 0.67 0.80 0.93 1.07 1.20 1.33 1.47
Proppant Concentration (lbm/gal)
Fig. 9—Proppant-concentration distribution for Group B1 (17
cases). Approximately 72% of wells in this group with prop- Fig. 10—Proppant-concentration distribution for Group B2 (17
pant-concentration ranges from 1.2 to 1.47 lbm/gal. cases). Approximately 82% of wells in this group with prop-
pant-concentration ranges from 1.2 to 1.47 lbm/gal.
Therefore, the only difference among Groups A1, B1, and B2
is proppant size. According to the proppant size, Groups A1 and In Group A2, the concentration distribution of applied prop-
B2 show characteristics of typical slickwater fracturing to carry pant is smaller than that in the other three geological groups. In
proppant of small size. Slickwater fractures typically generate other words, with the same fracture-fluid volume, the total prop-
longer hydraulic fractures with larger matrix-fracture contact area pant mass applied would be less in Group A2 as compared with
and lower fracture conductivity. For those wells, flowback is most the other three geological groups. Because proppant is required to
sensitive to shut-in time. Longer shut-in times may result in more hold open stimulated fractures, insufficient proppant loading may
fracture-fluid imbibition in matrix or natural fracture and return of result in increased fracture closure during flowback period and
less to the surface. reduction in water-recovery efficiency. Relatively low proppant-
Larger-sized proppant, such as 20/40-mesh proppant (420 to mass loading is, therefore, proposed to be the dominating factor
840 mm), applied in Group B1 may result in greater fracture width influencing flowback performance in Group A2.
and higher fracture conductivity near the wellbore (Ahn et al.
2014). With good fracture conductivity, fracture fluid could flow
easily to the wellbore—such a mechanism may be supported by Geological Parameter. The effect of geological properties on
data shown in Table 1. The volumetric flowback recovery may be 3-week-flowback recovery was also examined using PCA and
more directly affected by the fracture-fluid volume per stage K-means clustering methods to normalize the effects from stimu-
rather than shut-in time. lation and completion parameters and generate different clusters
of similar stimulation- and completion-parameter values.
5 The effect of thermal maturity was first examined. PCA and
clustering methods were applied on Group A wells, which share
4
3.87 the same formation-thickness range, but have different thermal
3-Week Flowback (%)
Soeder, D. J. 1988. Porosity and Permeability of Eastern Devonian Gas USA, 23–25 February. SPE-131767-MS. http://dx.doi.org/10.2118/
Shale. SPE Form Eval 3 (1): 116–124. SPE-15213-PA. http:// 131767-MS.
dx.doi.org/10.2118/15213-PA. Zammerilli, A. M. 2010. Projecting the Economic Impact of Marcellus Shale
Sweeney, J., Filer, J., Patchen, D. et al. 1986. Stratigraphy and Petroleum Gas Development in West Virginia: A Preliminary Analysis Using Public
Production of Middle and Upper Devonian Shales, Northwestern West Available Data. Report, DOE/NETL-402/033110, US Department of
Virginia. Presented at the SPE Unconventional Gas Technology Sym- Energy/National Energy Technology Laboratory (31 March 2010).
posium, Louisville, Kentucky, USA, 18–21 May. SPE-15222-MS. Zhou, Q., Dilmore, R., Kleit, A. et al. 2014. Evaluating Gas Production
http://dx.doi.org/10.2118/15222-MS. Performances in Marcellus Using Data Mining Technologies. Journal
Ward, J. A. 2010. Kerogen Density in the Marcellus Shale. Presented at of Natural Gas Science and Engineering 20: 109–120. http://
the SPE Unconventional Gas Conference, Pittsburgh, Pennsylvania, dx.doi.org/10.1016/j.jngse.2014.06.014.
Appendix A
Figs. A-1 through A-4 show the order of the variable importance for the four geological groups.
0.6
A1
0.5
Variable Importance
0.4
0.3
0.2
0.1
0
Fracture Proppant Lateral Vertical Average Shut-in
Fluid/Stage Mass/Stage Length/Stage Depth Rate Time
Fig. A-1—Rank of variable importance of each engineering factor in Group A1 (28 cases). Shut-in time is the most-important engi-
neering factor.
0.3
A2
0.25
Variable Importance
0.2
0.15
0.1
0.05
0
Fracture Proppant Lateral Vertical Depth Average Rate Shut-in Time
Fluid/Stage Mass/Stage Length/Stage
Fig. A-2—Rank of variable importance of each engineering factor in Group A2 (24 cases). Proppant mass/stage is the most-impor-
tant engineering factor.
0.25
B1
0.2
Variable Importance
0.15
0.1
0.05
0
Fracture Proppant Lateral Vertical Depth Average Rate Shut-in Time
Fluid/Stage Mass/Stage Length/Stage
Fig. A-3—Rank of variable importance of each engineering factor in Group B1 (17 cases). Fracture fluid/stage is the most-important
engineering factor.
0.25
B2
0.2
Variable Importance
0.15
0.1
0.05
0
Fracture Proppant Lateral Vertical Depth Average Rate Shut-in Time
Fluid/Stage Mass/Stage Length/Stage
Fig. A-4—Rank of variable importance of each engineering factor in Group B2 (17 cases). Shut-in time is the most-important engi-
neering factor.
The PCA method is implemented here to replace the seven varia- pcaComp1 ¼ /11 Fracture Fluid Volume þ /21
bles (fracture-fluid volume, proppant mass, vertical depth, lateral Proppant þ /31 Depth þ /41 Lateral Length þ /51
length, number of stages, treatment rate, and shut-in time) that are Number of Stages þ /61 Trearment Rate þ /71
linearly correlated with flowback-water volume as two linearly
uncorrelated variables (principal components), referenced herein Shut-in Time: ðB-2Þ
as pcaComp1 and pcaComp2. Each component is the linear com- pcaComp2 ¼ /12 Fracture Fluid Volume þ /22
bination of certain coefficients (principal component loading vec-
tors) from the seven variables, as shown in Eqs. B-2 and B-3. The Proppant þ /32 Depth þ /42 Lateral Length þ /52
coefficients for pcaComp1 and pcaComp2 are shown in Table B- Number of Stages þ /62 Trearment Rate þ /72
1 and the corresponding clustering results are shown in Figs. B-1 Shut-in Time: ðB-3Þ
and B-2. The best choice for the number of clusters is made
2
1
1
pcaComp2
pcaComp2
0
0
–1
–1
–2
–2
–3
–3
–3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3
pcaComp1 pcaComp1
2
2
1
1
pcaComp2
pcaComp2
0
0
–1
–1
–2
–2
–3
–3
–3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3
pcaComp1 pcaComp1
Fig. B-1—Clustering results for geologic Group A (52 cases). Each point is one well, and each circle is one cluster.
2
1
1
pcaComp2
pcaComp2
0
0
–1
–1
–2
–2
–3
–3
–4 –2 0 2 –4 –2 0 2
pcaComp1 pcaComp1
2
1
1
pcaComp2
pcaComp2
0
–1
–1
–2
–2
–3
–3
–4 –2 0 2 –4 –2 0 2
pcaComp1 pcaComp1
Fig. B-2—Clustering results for geologic Group B (34 cases). Each point is one well, and each circle is one cluster.
Qiumei Zhou is a PhD-degree candidate in the Energy and mental engineer with the American Academy of Environmen-
Mineral Engineering Department at Pennsylvania State Univer- tal Engineers and Scientists.
sity. Her research interests include numerical simulation and
data-mining analysis for unconventional reservoirs. Zhou holds John Wang is currently an associate professor of petroleum
a BS degree in petroleum geology from Chengdu University of and natural-gas engineering at the John and Willie Leone Fam-
Technology (China) and an MS degree in petroleum and nat- ily Department of Energy and Mineral Engineering, having
ural gas engineering from Pennsylvania State University. joined the faculty of Pennsylvania State University in 2009. He is
also a director of Alfavor Petroleum Corporation. Before joining
Robert Dilmore is a research environmental engineer in the US the university, Wang was a petroleum engineer with a US inde-
Department of Energy’s National Energy Technology Labora- pendent producer in Shreveport, Louisiana, where he worked
tory, Office of Research and Development, working on topics on evaluating and optimizing producing oil/gas fields. He is
related to characterizing technical and environmental per- dedicated to advancing the knowledge and practices in res-
formance of complex engineered geologic systems. His ervoir evaluation and stimulation through teaching undergrad-
recent research foci have included CO2 enhanced recovery uate and graduate courses, working with research scientists/
and storage in conventional and unconventional systems, associates/assistants, and serving the industry and professional
integrated assessment modeling for quantitative risk assess- society in a number of ways. Currently, Wang serves as the edi-
ment of large-scale geologic CO2 storage, characterization tor-in-chief of Elsevier’s Journal of Unconventional Oil and Gas
of potential pathways for unwanted subsurface fluid migra- Resources, and is an associate editor of Springer’s Journal of
tion, and industrial process-waste management. Dilmore holds Petroleum Exploration and Production Technologies. He holds
a BS degree from the University of Delaware and a PhD a PhD degree from Texas A&M University, an MSc degree from
degree in environmental engineering from the University of the University of Houston, and a BS degree from the University of
Pittsburgh. He has authored or coauthored more than 40 pub- Petroleum of China, all in petroleum engineering.
lished papers, technical reports, and book chapters on his
research; is a professional engineer in the commonwealth of Andrew Kleit is a professor of energy and environmental eco-
Pennsylvania; and holds board certification as an environ- nomics and MICASU Faculty Fellow at the Pennsylvania State
University. He is the Founding Chair of the Energy Business and the impacts of mining on political unrest in Peru. He has also
Finance Program, the largest undergraduate major in the uni- taught two courses at ESAN University. Kleit is the author of more
versity’s College of Earth and Mineral Sciences. Before becom- than 70 articles, including publications in the Journal of Law and
ing an academic, Kleit was Economic Advisor to the Director, Economics, Review of Economics and Statistics, Health Affairs,
Bureau of Competition, Federal Trade Commission, and Senior Health Economics, Journal of Regulatory Economics, Resources
Economic Adviser to the Director for Investigation and Research and Energy Economics, Energy Studies Review, and The Energy
(the chief antitrust official in Canada). He has traveled widely in Journal. He is the author or editor of five books, including Elec-
Peru and has deep interest in the economic development of tricity Restructuring: The Texas Story (2009) and Electric Choices:
the country. Together with Oswaldo Morales Tristan of ESAN Uni- Deregulation and the Future of Electric Power (2006). Kleit holds
versity in Lima, Kleit is working on a research agenda examining a PhD degree in economics from Yale University.