See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/299445970

Reducing Data-Collection Efforts for Conceptual Cost Estimating at a Highway

Agency

Article  in  Journal of Construction Engineering and Management · March 2016

DOI: 10.1061/(ASCE)CO.1943-7862.0001174

CITATIONS READS
16 454

3 authors, including:

Douglas D. Gransberg Hyungseok David Jeong

Gransberg & Associates Inc. Iowa State University
416 PUBLICATIONS   2,533 CITATIONS    119 PUBLICATIONS   1,237 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Optimizing Construction Equipment Systems View project

Alternate Design / Alternate Bid Process for Pavement Type Selection View project

All content following this page was uploaded by Douglas D. Gransberg on 16 October 2017.

The user has requested enhancement of the downloaded file.

 Final accepted manuscript. Published as:
1 Gardner, B.J., Gransberg, D.D., and H.D. Jeong, “Reducing Data Collection Efforts for Conceptual Cost Estimating at a Highway
Agency,” Journal of Construction Engineering and Management, 10.1061/(ASCE)CO.1943-7862.0001174, 2016, pp.

22 1 Reducing Data-Collection Efforts for Conceptual Cost

Estimating at a Highway Agency

Y
3 3

4 Brendon Gardner 1; Douglas D. Gransberg 2; and H. David Jeong 3

L
5 Abstract: Data-driven models using historical project attributes to estimate future construction costs, such as multiple-regression analysis
6 and artificial neural networks are both proven techniques that highway agencies could adopt for conceptual cost estimating. This research
7 found literature using those techniques has been solely focused on estimating model performance with little to no attention to the level of
8 effort required to conduct the conceptual estimate. It is commonly believed using more input data enhances estimate accuracy. However, this
9 paper finds for the highway agency studied that using more input variables than necessary in the conceptual estimate does not improve

N
10 estimate accuracy. Conceptual estimates using the minimum amount of input data to produce an estimate with a reasonable level of confidence
11 is more cost effective. This paper quantifies the effort expended to undertake conceptual estimates using data from a highway agency and
12 concludes that input variables that have a large influence on the final predicted cost and require a low amount of effort are desired in data-
13 driven conceptual cost-estimating models for the agency studied. DOI: 10.1061/(ASCE)CO.1943-7862.0001174. © 2016 American Society
14 of Civil Engineers.

O
15 Author keywords: Conceptual cost estimating; Highway infrastructure; Artificial neural networks; Multiple-regression analysis; Cost and
16 schedule.

17 Introduction derived with the least amount of project-specific technical informa- 40

tion (Bode 2000; FHWA 2015). If the figure is too conservative, the 41
18 4 In public works, the budget for a project is often established at a project may not be selected to advance to the next preliminary en- 42
19 point in project development where the estimator has the least gineering stage. As a result, it becomes important to take the initial 43
F
20 amount of design detail from which to compute an estimate (Bode cost estimate seriously and use the available information that has 44
21 2000). Taking federally-funded highway projects as an example, the highest influence on the bottom-line and not allocate precious 45
22 the budget is formally set when the project is assigned a federal time and resources to a project that ultimately will not advance 46
23 project identification number (PIN) and included in the Statewide to construction. Additionally, the time period to conduct the es- 47
24 Transportation Improvement Program (FHWA 2015; Anderson timate is typically limited in the feasibility stage (Gunduz et al. 48
O

25 et al. 2007). The estimate is typically used during early planning 2011), but the estimate requires sufficient accuracy for benefit- 49
26 stages to conduct initial feasibility studies, and both engineers and cost analysis and prioritizing budgets (Anderson et al. 2007). 50
27 planners realize that the accuracy of the initial cost estimate is a Therefore, the objective of this paper is to explore a solution that 51
28 function of the level of design detail available at the time of the can be used to complete critical initial estimates with high-impact 52
29 estimate. To account for the anticipated change in project scope as data that requires the minimum level of effort for the estimator to 53
30 the development process proceeds, a standard contingency based obtain. 54
O

31 on a percentage of the total estimate is added (Minassian and Jer-

32 geas 2009). This kind of estimate is termed a top-down estimate
33 because it relies on parametric cost factors such as lane-miles, lo- Conceptual Cost Estimating in Highway Agencies 55
34 cation, project type, etc., rather than a bottom-up estimate whose
35 basis are the quantities of materials needed on the project (Kim et al. Conceptual cost estimating (CCE) is the first construction cost es- 56
36 2012). timate completed for a project, as shown in Fig. 1. At the concep- 57
tual stage there is little information known about a project, and the 58
R

37 The conundrum faced by engineers in public works is that in

38 order to receive the authorization to expend funds, to advance detailed design has not yet begun. As the importance of the infor- 59
39 the project development, the official budget is based on a figure mation required for the conceptual estimate increases, so too does 60
the need to expend additional design and planning, and this in turn 61
1
Graduate Research Assistant, Dept. of Civil, Construction, and extends the project planning period. Further design and planning 62
Environmental Engineering, Iowa State Univ., Ames, IA 50011-3232 details can be included in later, more confident, estimates. 63
P

(corresponding author). E-mail: brendon.gardner@gmail.com Highway agencies cannot afford to overinvest their design time 64
2
Professor, Dept. of Civil, Construction, and Environmental Engineering, and effort in projects at the conceptual stage. If less effort can be 65
Iowa State Univ., Ames, IA 50011-3232. E-mail: dgran@iastate.edu expended at the conceptual stage, then an estimator’s time can be 66
3
Associate Professor, Dept. of Civil, Construction, and Environmental better applied in the later design estimating stages shown in Fig. 1. 67
Engineering, Iowa State Univ., Ames, IA 50011-3232. E-mail: djeong@
Any investment in the project at the conceptual stage could be ren- 68
iastate.edu
Note. This manuscript was submitted on November 13, 2015; approved dered worthless if a project is not selected for further development 69
on March 3, 2016No Epub Date. Discussion period open until 0, 0; following a benefit-to-cost analysis or a needs assessment. In the 70
separate discussions must be submitted for individual papers. This paper context of structural steel buildings, only 15% of those that reach 71
is part of the Journal of Construction Engineering and Management, the conceptual stage ever get constructed (Moselhi and Siqueira 72
© ASCE, ISSN 0733-9364. 1998). 73

© ASCE 1 J. Constr. Eng. Manage.

 Y
F1:1 Fig. 1. Construction cost estimating timeline

74 The CCE techniques currently used by highway agencies vary Data-Driven CCE Models—Prior Studies 128
75 by state. Byrnes (2002) and Turochy et al. (2001) have completed

L
76 surveys on cost estimating at the planning stage. Those two studies The CCE techniques reviewed in this research include ANN and 129
77 found that CCE approaches used by highway agencies are gener- MRA models; both are commonly suggested in the literature 130
78 ally classed into one the following three categories: and will be referred to as data-driven CCE models. The benefits 131
79 • cost-per-mile of typical sections of highway or bridge; of data-driven techniques are the ability to use historical project 132
80 • estimating approximate quantities of major work items; or information for forecasting and the speed at which this can be 133
• no documented or uniform method, instead using experience achieved. Gunduz et al. (2011) recognized this, stating “reliable

N
81 134
82 and engineering judgment. cost estimates are required within a very limited time period at 135
83 Despite these techniques, CCE at highway agencies still re- the feasibility stage.” 136
84 quires improvement. Flyvbjerg et al. (2002) investigated 258 public Performance of data-driven CCE models is subject to variations 137
85 transportation projects and found that 86% of those projects had in model architecture and parameters; this includes the input var- 138
86 experienced cost growths since the initial estimate; on average they iables used, number of hidden layers and nodes in the ANN model, 139

O
87 were 28% higher than the initial estimate. Further, Schexnayder and data-set size. The effects of model architecture and parameters 140
88 et al. (2003) stated that recent publicity has called into question have been studied in data-driven CCE models published in the lit- 141
89 the “ability of departments of transportation to forecast accurately erature (Setyawati et al. 2002; Mahamid 2011; Petroutsatou et al. 142
90 and to control the final cost of their projects”. 2012). Bell and Ghazanfer (1987) selected their final MRA model 143
91 Data-driven techniques using artificial neural networks (ANNs) by building many models through trial and error, choosing the 144
92 and multiple regression analysis (MRA) have been frequently sug- model that produced the least error. This technique has been used 145
93 gested in the literature for CCE and show results equal to, if not in at least four other studies (Creese and Li 1995; Hegazy and Ayed 146
94 superior than, those currently used by highway agencies (Bell and 1998; Gunduz et al. 2011; Petroutsatou et al. 2012). 147
Input variables selected have a large effect on the prediction
F
95 Ghazanfer 1987; Hegazy and Ayed 1998; Moselhi et al. 1992). The 148
96 performance of these models are well within the acceptable estimate capability of the CCE model (Rueda-Benavides and Gransberg 149
97 range for the planning stage suggested by the American Association 2014). Bell and Ghazanfer (1987) concluded this using MRA to 150
98 of State Highway and Transportation Officials (AASHTO) in the predict the construction cost of highway projects. The same deduc- 151
99 Cost Estimating Guidelines (AASHTO 2013). In 2002, Byrnes tion has been reached by at least two other authors of data-driven 152
100 (2002) found that no state highway agency is yet employing math- CCE models (Gunaydin and Dogan 2004; Setyawati et al. 2002). 153
O

101 ematical models; NCHRP Report 574 (Anderson et al. 2007) Bell and Ghanzanfer investigated the accuracy of the input varia- 154
102 reached the same conclusion. bles in their research. Gunyadin and Dogan made reference to the 155
103 Both MRA and ANNs link a historical database of project attrib- selection of the input variable types, and Setyawati et al. referred to 156
104 utes to the actual final construction cost of each project. The rela- optimizing the number of input variables to achieve better predic- 157
105 tionship identified within the data can then be used to forecast the tion accuracy. Model creators usually only have a one-time com- 158
106 final construction cost of future projects. Using MRA links the in- mitment to collecting the cost predictors (Smith and Mason 1997). 159
O

107 formation with a linear equation to the construction cost (Turochy If model creators select cost predictors that require a large amount 160
108 et al. 2001). Each attribute is assigned a weight when the linear of data collection and processing effort, then it will burden the use- 161
109 equation is developed such that the error in forecasting the con- fulness of the model as a CCE tool. 162
110 struction cost is minimized. On the other hand, ANNs use artificial Despite the amount of previous research in data-driven CCE 163
111 intelligence to find patterns to describe the construction cost from a models, no work reported in the construction literature studied 164
112 historical database of project attributes (Pewdum et al. 2009). His- quantifies the effort required to conduct the cost estimate. Collec- 165
R

113 torical data is used to train the ANN model and recognize relation- tion and storage of data from historical projects requires time and 166
114 ships within the database. This trained model is then used to resources of which highway agencies have a limited quantity. Fur- 167
115 forecast future construction costs by looking for similar patterns. ther cost influencing information gathered later in the project life- 168
116 No matter the CCE technique employed by highway agencies or cycle can be included in more detailed bottom-up design stage 169
117 suggested in the literature, a particular level of project definition (or estimates. 170
118 design effort) is required in order to conduct the cost estimate.
P

119 Sanders et al. (1992) observed this balancing act between efforts
Literature Analysis 171
120 expended and estimate accuracy, stating “there is an inverse rela-
121 tionship between the accuracy of an estimate and its preparation Previous authors of data-driven CCE model research have remained 172
122 cost. At some point, increased accuracy cannot justify the addi- silent on the effort to collect, store, and use databases to conduct the 173
123 tional costs incurred.” The sooner that the initial estimate is devel- cost estimates. As a result, this research analyzed the data-driven 174
124 oped, the smaller the level of project definition required for CCE CCE models published in the literature to observe how many input 175
125 with commensurately lower cost and effort. This then means that variables are being used and resultant error. The literature analysis 176
126 estimators and designers can focus their efforts on projects that are was a starting point of the research to investigate if additional inputs 177
127 beyond the planning stage and are likely to reach construction. improve estimating accuracy. 178

© ASCE 2 J. Constr. Eng. Manage.

 Table 1. Construction Cost-Estimating Models Studied
Input ANN estimating MRA estimating
T1:1 Author variables error (%) error (%) Brief project scope
T1:2 Petroutsatou et al. (2012) 5 4.65 — Tunnels in Greece
T1:3 Mahamid (2011) 9 — 13.0 Highway (various sizes)
T1:4 Gunduz et al. (2011) 17 5.76 2.32 Light rail track works in Turkey

Y
T1:5 Lowe et al. (2006) 12 — 19.30 Buildings in U.K.
T1:6 Petroutsatou et al. (2006) 5 — 9.6 Tunnels in Greece
T1:7 Kim et al. (2004) 9 3.0 7.0 Residential buildings in Seoul, Korea
T1:8 Gunaydin and Dogan (2004) 8 7.0 — RC 4-8 story residential buildings in Turkey
T1:9 Emsley et al. (2002) 5 16.6 — Buildings
T1:10 Setyawati et al. (2002) 2 13.4 9.2 Education building construction

L
T1:11 Al-Tabtabai et al. (1999) 9 9.1 — Highway construction
T1:12 Hegazy and Ayed (1998) 10 19.33 — Highway construction in Newfoundland, Canada
T1:13 Elhag and Boussabaine (1998) 4 17.80 — School construction
T1:14 Moselhi and Siqueira (1998) 4 10.77 14.76 Steel framed low-rise buildings
T1:15 Creese and Li (1995) 3 8.24 — Timber bridges
T1:16 Sanders et al. (1992) 10 — 6.0 Urban highway bridge widening in Alabama

N
T1:17 Bell and Ghazanfer (1987) 5 — 17.0 Highway construction maintenance projects
Note: — indicates that data is not applicable to that publication.

179 A total of 16 publications that included results from data-driven with a smaller reduction in error as each input variable is added; this 202

O
180 CCE models were studied. Publications were selected that involved is highlighted by the best fit curves being negative power curves. 203
181 either ANN or MRA prediction algorithms to output the construc- The relationship is much stronger with the MRA models in the lit- 204
182 tion cost of the project using input variables at the early design erature, with the power curve R2 value being 0.7211. When the 205
183 stage. From each of the publications both the performance and the obvious outlier in Fig. 2(a) is removed, then the R2 value in that 206
184 number of input variables used to produce their best performing plot increases to 0.335. 207
185 model were collected. The results of this research is shown in A weakness of this conclusion is that the literature is for projects 208
186 Table 1. Some publications tested both the ANN and MRA tech- of many different scopes. Additionally, none of these past studies 209
187 niques, which resulted in a total of 20 models for comparison, have converted their input variables into perceived effort, which 210
F
188 shown graphically in Fig. 2. means that effort and performance cannot be directly compared. 211
189 Performance of both the ANN and MRA models were measured An assumption of this literature analysis is that each input variable 212
190 using the mean average percentage error (MAPE). This method is requires equal estimating effort (Gransberg and Riemer 2009). The 213
191 commonly used by authors of data-driven CCE models (Petroutsa- results of this study specifically quantify input variable effort and 214
192 tou et al. 2012; Gunduz et al. 2011; Mahamid 2011; Hegazy and prove that not all input variables require the same level of effort to 215
193 Ayed 1998). Calculation of the MAPE is furnished using Eq. (1) compute. 216
O

194 (Mahamid 2011) The requirement to minimize estimating effort for CCE is also 217

 n   recognized in other industries outside of construction. Verlinden 218
100% X  P i − Ai 
  et al. (2008) created an ANN to calculate the cost of sheet metal 219
MAPEð%Þ ¼  A  ð1Þ
n i¼1 i manufacturing for customers; their research recognized the neces- 220
sity to provide customers a swift quotation, albeit at the cost of 221
195 where n = number of data-points used to test the model; Pi = pre- possibly reduced accuracy. In another study, Walczak (2001) cre- 222
O

196 dicted construction cost using the data-driven CCE model for the ated an ANN to predict a foreign exchange rate. Walczak’s study 223
197 ith project; and Ai = actual construction cost from the historical found there was no need to use the entire available database and that 224
198 records collected for the ith project. only a few years of data was necessary to provide reasonable con- 225
199 The results from the literature analysis in Figs. 2(a and b) show fidence. Walczak concluded that this would have a significant effect 226
200 that previous publications achieved lower error through more input on model development cost savings, where “the cost is not only fi- 227
201 variables. Both plots in Figs. 2(a and b) show diminishing returns nancial, but also the development time and effort” (Walczak 2001). 228
R

25 35
30
Error [MAPE] (%)
Error [MAPE] (%)

20
25
15 R = 0.1462 20
P

R2 = 0.3335 (excluding outlier)

10 15
10 y = 60.609x-0.984
5
y = 18.386x-0.469 5 R = 0.7211
0 0
0 5 10 15 20 0 5 10 15 20
(a) Number of Input Variables (b) Number of Input Variables
Neural Network Regression

F2:1 Fig. 2. Literature analysis of inputs versus error: (a) number of input variables versus model prediction error of artificial neural networks; (b) number
F2:2 5 of input variables versus model prediction error of multiple regression analysis

© ASCE 3 J. Constr. Eng. Manage.

 was recorded using the proposed input selection order (Step 3a), 264
and then it was repeated using the input selection order in reverse 265
(Step 3b). Finally, Step 4 compares the cumulative perceived effort 266
for each construction cost estimate to the estimating error achieved. 267
In this step the proposed input selection method (3a) is compared to 268
completing the task in reverse order (3b) in order to validate frame- 269

Y
work effectiveness. 270

F3:1 Fig. 3. Proposed dual-objective hierarchy tree for conceptual cost Survey 271
F3:2 estimates
A survey was conducted at MDT to understand the perceived level 272
of effort required to estimate the construction cost of a project at 273

L
the conceptual stage. Firstly, two days of interviews at MDT estab- 274
229 Research Objective lished the key attributes of a project that influence the construction 275
230 This paper proposes a new CCE objectives hierarchy, illustrated in cost to aid the survey development. Following these interviews and 276
231 Fig. 3, to evaluate data-driven CCE models. Previous data-driven a review of literature, 29 variables were identified that have an in- 277
CCE models have been focused on the prediction accuracy (Objec- fluence on the construction cost of MDT’s highway projects; these 278
232

N
are shown in Table 2. The research team then assigned the attributes 279
233 tive 1). This research adds another parameter for consideration: it
into one of three categories: 280
234 investigates the influence of effort expended (Objective 2) in gath-
• Roadway: an attribute associated with information about the 281
235 ering the input information for the models.
proposed project location; 282
236 The objective of this paper is to evaluate the effort expended
• Design: an attribute determined during the design process; or 283
237 for data-driven CCE models. Specifically the paper focuses on two
• Construction Administration: an attribute related to the con- 284
238 questions:

O
struction activity. 285
239 1. Can a framework be created to select inputs that help meet the
These categories were selected to reflect the location where the 286
240 dual-objective goal of maximum performance with minimal
data was being received from at MDT. For example, the majority of 287
241 effort?
roadway characteristics were generally sourced from the Data and 288
242 2. Is there an optimum number of input variables that highway
Statistics Bureau at MDT, which stores Geographical Information 289
243 agencies should be collecting to minimize the effort for data-
Systems (GIS) on roadway attributes. 290
244 driven CCE models?
Respondents of the survey were asked, among other questions, 291
245 The outcomes of this research help both researchers and practi-
to answer the following on each of the 29 attributes identified: 292
246 tioners to focus on both objectives during the CCE stage, allowing
• Rate the typical effort required to compute or identify this vari-
F
293
247 them to estimate the projects’ construction cost at an early stage of
able, and 294
248 project development with the least amount of effort but with opti- • How influential do you believe this variable is on the construc- 295
249 mal performance. tion cost of a project? 296
250 Data was furnished from one state highway agency, Montana The questions were designed with an ordinal (categorical) scale 297
251 Department of Transportation (MDT), to conduct this study. The in which respondents are required to select the most suitable answer 298
O

252 research scope has been limited to one major work-type, pavement as shown in Fig. 5 (Fink 2009; Fowler 2009). 299
253 preservation of highway projects. To validate the findings two data- The survey was distributed at MDT through an email link to all 300
254 driven methods, ANN and MRA, were tested. 84 preconstruction personnel that were deemed suitably qualified 301
to respond. A total of 35 responses were received with four of these 302
excluded as nonresponses. This resulted in a 37% response rate. 303
255 Research Methodology Responses were received from all five bureaus and from a large 304
O

range of job titles. Although there is “no agreed-upon standard for 305
256 To validate the input selection framework and determine if an op- a minimum acceptable response rate” (Fowler 2009), the research- 306
257 timum level of input variables exist, a combination of perceptional ers were satisfied that the 37% response rate reflected the entire 307
258 survey data was used with real project data to predict the construc- population. 308
259 tion cost. The research steps are shown in Fig. 4 below. In Step 1, a
260 survey was conducted to grasp perception on the level of effort re-
Input Variable Selection 309
R

261 quired for different inputs to the conceptual estimate. The dual-
262 objective input selection method, proposed as part of this research, To meet the dual-objective CCE goal, it was proposed that input 310
263 was then used in Step 2. Next, the estimating error for each model variables be selected starting with those that require a low level 311
P

F4:1 Fig. 4. Research steps

© ASCE 4 J. Constr. Eng. Manage.

 Table 2. Cost-Influencing Attributes Identified at MDT of effort to compute or identify but that also have a high influence 312
T2:1 Reference on the construction cost of the project. This is shown in Fig. 6 with 313
number Cost-influencing attribute Attribute category the input variables suggested to be selected in the bottom right hand 314
quadrant. 315
T2:2 1 Design AADT Design
T2:3 2 Design speed Design To validate this selection process, the research team combined 316
T2:4 3 Start and end stations, length Design the perceptive survey results with performance of a data-driven 317

Y
and width CCE model created specifically using projects that the survey re- 318
T2:5 4 Intersection signalization and Design spondents design and manage at MDT. Two data-driven CCE mod- 319
signage eling techniques, ANN and MRA, were used with databases 320
T2:6 5 Horizontal and vertical Design provided by MDT to predict the construction costs of projects. In- 321
alignment put variables were systematically added to the data-driven CCE 322
T2:7 6 Extent of changes to the existing Design model starting with those in the bottom right quadrant of Fig. 6 323

L
intersections
to meet the dual objectives of the main CCE goal. Further inputs 324
T2:8 7 Typical section Design
T2:9 8 Curb, gutter and sidewalk Design were added based on their distance from the bottom right quadrant 325
T2:10 9 Bridge type and complexity Design in Fig. 6; this is explained in more detail later on in this paper. In 326
T2:11 10 Volumes of excavation and Design each of the models, the performance and total perceived effort from 327
embankment all input variables used were recorded. 328

N
T2:12 11 Geotechnical—subsurface and Design
slope recommendations
T2:13 12 Bridge deck area Design
Results 329
T2:14 13 Hydraulic recommendations Design
and culverts
T2:15 14 Storm sewer extents Design Survey Response 330

O
T2:16 15 Bridge span lengths Design
T2:17 16 Foundation complexity of the Design The average results of the survey from 31 respondents are shown in 331
bridge Fig. 7; the data-point numbers reference the 29 attributes from 332
T2:18 17 Right-of-way acquisition and Design Table 2. Respondents rated the effort on a 1–3 ordinal scale whereas 333
costs the influence of this variable on the construction cost was rated on a 334
T2:19 18 Extent of utility relocations and Design 1–4 ordinal scale; these scales are shown in Fig. 5. As such quad- 335
costs rants were arbitrarily assigned on both scales to visually divide up 336
T2:20 19 Urban or rural project Roadway the results and aid the input variable selection process. The units on 337
T2:21 20 Construction on Native Roadway
both axis correspond to the ordinal response scale from Fig. 5; they 338
F
American reservations
T2:22 21 Site topography Roadway are referred to as “points” from here on. 339
T2:23 22 Existing surfacing conditions Roadway Visually, there are a number of interesting results shown in 340
and depths Fig. 7. Firstly, only 5 of the 29 attributes shown in Table 2 fall 341
T2:24 23 Number of intersections in Roadway in the bottom right quadrant of the plot: attributes MDT perceives 342
project as requiring a low amount of effort to collect that also have a high 343
O

T2:25 24 Number of bridges in the project Roadway influence on the construction cost of the project. It was not a sur- 344
scope prise that three are roadway characteristics, easily identified once a 345
T2:26 25 Traffic control—closures or Construction administration project has been selected and its location confirmed. These char- 346
detours
acteristics include whether the project is going to be in an urban 347
T2:27 26 Environmental permitting Construction administration
requirements—wetlands environment, the topography of the road, and the number of bridges 348
T2:28 27 Letting date Construction administration within the limits of the project. There was only one design factor 349
O

T2:29 28 Context sensitive design issues, Construction administration identified in the bottom right quadrant. 350
controversy Secondly, all the attributes in the top right quadrant of the Fig. 7 351
T2:30 29 Contract time Construction administration are design factors. This is intuitively logical as design requires 352
R

Question 1) Rate the typical effort required to compute or identify this variable:
Rating: L = Low effort, M = Medium time H = High effort
information and effort involved. Possibly site
available, desktop visits, site
study investigations and
approximations.
P

Scale: 1 2 3

Question 2) How influential do you believe this variable is on construction cost:

Does not Minor Average Major influence
Answer: influence cost influence influence
Scale: 1 2 3 4

F5:1 Fig. 5. Ordinal scale used for the two survey questions

© ASCE 5 J. Constr. Eng. Manage.

 Y
L
F6:1 Fig. 6. Selecting input variables to meet the dual objectives of CCE
Fig. 8. Preference for selecting input variables F8:1

N
Identifying Effort and Influence of Input Variables
where xi = average perceived cost influence from the survey; 373
Effort expended to identify/calculate (points)

3
A ¼ 4, the maximum construction cost influence based on the 374
18 17
5
ordinal survey rating and the ideal value as shown Fig. 8; yi = aver- 375
13 11
age perceived effort from the survey; B = 1, the minimum effort 376

O
2.5 6 16
14 10 rating based on the ordinal survey rating and the ideal value as 377
8
15 shown Fig. 8; and i = input attribute being measured, ranging from 378
25 9
2
26 7 1 to 29. 379
4 28
22 A total of 189 pavement preservation projects were provided to 380
3
24 the research team from MDT. The projects were made available 381
29
1.5 12 from existing databases and conceptual project reports completed 382
27 1 21
during the planning phase of each project. The research team then 383
23
2
20
compiled a database that included as many of the 29 input variables 384
F
19
1 for the ANN and MRA model as possible. Because the survey was 385
2 2.5 3 3.5 4
created for generic project types, some of the project attributes were 386
Influence on the Construction Cost (points)
not relevant to pavement preservation projects. For the purposes of 387
Design Roadway Construction Administration this case study, 13 input variables relevant to pavement preservation 388
were selected. These were selected based on guidance from MDT 389
O

F7:1 Fig. 7. Results of MDT cost-estimating survey personnel, and selected attributes are shown in Table 3. An example 390
is the exclusion of right-of-way acquisition costs; it was determined 391
at MDT to be very unlikely that pavement preservation projects 392
would involve such occurrence. 393
353 significant effort to be expended, and the outcome should have a
Input variables were added by selecting them in order starting 394
354 large effect on the construction cost. Finally, very few variables
with the shortest distance from the ideal input variable to the largest 395
355 occupy the top left quadrant. Those that do occupy this quadrant
O

distance. The average survey results for the influence and effort are 396
356 are bordering other quadrants, implying that any attribute requiring
shown in Table 3 along with the calculated distance to the “ideal 397
357 a significant amount of effort to be expended by MDT is going to
input variable” shown in Fig. 8. Each time a new input variable was 398
358 have a significant influence on the construction cost of the project.
added to the model, the MAPE of the model with the test data was 399
359 This observation is also reinforced by the fact that two-thirds of all
recorded. To verify the usefulness of the input selection method, the 400
360 variables are in the bottom left or top right quadrant (i.e., variables
process was repeated in the reverse order (starting with the largest 401
361 are either low-effort/low-influence or high-effort/high-influence
distance from the ideal input variable).
R

402
362 variables).
To be able to compare the results from all the models, the same 403
152 projects were used to train each model, and the same 38 proj- 404
363 Case-Study ects were used to test the model and calculate the MAPE. The ran- 405
364 The findings from the survey were used to validate the dual- domly selected 38 test projects accounted for 20% of the database; 406
365 objective input variable selection method proposed as part of this this proportion of testing to training data was based on previous 407
P

366 research. The research team proceeded to build a data-driven CCE literature (Petroutsatou et al. 2012; Moselhi and Siqueira 1998). 408
367 model that has the least amount of effort with suitable performance.
ANN Results 409
368 As such, many of the 29 attributes were included in the model, one
A commercially available ANN modeling software package was 410
369 at a time, starting with the variable closest to the most preferred to
used to train and then test the database. Initially, only one input 411
370 the least preferred attribute (as shown in Fig. 8). The formula to
variable with the shortest distance to the ideal input variable shown 412
371 calculate each distance was based on the Euclidean distance,
in Fig. 8 was used to train and then test the first model. Input var- 413
372 and shown in Eq. (2) (Danielsson 1980)
iables were then added to the model one at a time, getting further 414
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
from the ideal input variable. Each time, the MAPE and cumulative 415
Distance to ideal input ðpointsÞ ¼ ðxi − AÞ2 þ ðyi − BÞ2 ð2Þ
effort points of the prediction model was recorded. The process was 416

© ASCE 6 J. Constr. Eng. Manage.

 Table 3. Input Variables Selection Order and Distance from Ideal Input
Average perceived Average perceived Distance to ideal input
T3:1 Proposed input variable selection order influence (points) effort (points) (points) refer to Eq. (2)
T3:2 19. Urban or rural project 3.48 1.10 0.56
T3:3 21. Site topography (steep, flat, or undulating terrain) 3.26 1.29 0.80
T3:4 3. Start and end stations, length and width 2.97 1.71 1.25

Y
T3:5 1. Design AADT 2.74 1.29 1.29
T3:6 7. Typical section (depths of surfacing and aggregate) 3.19 2.03 1.31
T3:7 2. Design speed(s) 2.67 1.16 1.34
T3:8 4. Intersection signalization and signage 2.87 1.90 1.44
T3:9 25. Traffic control—closures or detours 2.84 2.00 1.53
T3:10 8. Curb and gutter and sidewalk 2.97 2.13 1.53

L
T3:11 29. Contract time 2.45 1.58 1.65
T3:12 27. Letting date 2.35 1.29 1.67
T3:13 11. Geotechnical—subsurface and slope recommendations 3.39 2.65 1.76
T3:14 6. Extent of utility relocations and costs 3.26 2.71 1.86

N
417 then repeated until all 13 input variables were included in the ANN stage. Because these two variables were also perceived by MDT as 442
418 model. The process was then conducted in reverse order by adding having a high influence on the construction cost, the input selection 443
419 input variables in the opposite fashion. Fig. 9 illustrates the results process proposed in this research picked these two variables among 444
420 of each approach. the first 6–8 variables. 445
421 Fig. 9 shows that when input variables are added in the order On the other hand, design variables require a high level of effort 446

O
422 suggested by this research, the model can more quickly reach rea- at the early stage. Although they have high influence on the con- 447
423 sonable accuracy with less effort. It also minimizes the number of struction cost, many were excluded from the first 6–8 variables. 448
424 input variables required to achieve the lowest possible MAPE. The Most design factors do have a perceived high impact on the con- 449
425 corresponding model reached around 25% error with a cumulative struction cost, but at the early stage, there is a low level of confi- 450
426 effort of 7.5 points. With the reverse order of input variable selec- dence with those numbers. Two such examples are the geotechnical 451
427 tion, a comparable level of error was not reached until around complexities and utility replacements required. At the early stage, 452
428 17.5–20 points of effort. This is more than twice the level of es- highway agencies only have a very vague estimate of those 453
429 timating effort for the same performance. Both methods show that variables; thus the confidence in the top-down number is very low 454
F
430 there is a point where adding additional input variables, or expend- at the conceptual stage. However, it is recognized that their de- 455
431 ing more effort, results in diminishing returns and little or no im- signed outcome does have a significant impact on the cost. The 456
432 provement in performance in predicting construction costs for the data inputs for design variables in the conceptual estimating model 457
433 additional effort. When the point of diminishing returns is reached, are sourced from project information at the early stage; thus they 458
434 the overall goal of the estimating model is reached: maximum per- are not inputs known to a high level of confidence and contain 459
O

435 formance with minimal effort. plenty of variability from this initial estimate to the final estimate. 460
436 The authors speculate that selecting input variables that require a This is unlike variables such as the length or urban/rural input var- 461
437 low level of effort essentially means that variable is known to a high iables, which are known to a high level of confidence at the early 462
438 degree of confidence at the early estimate stage. Two examples are stage and also have a high impact on the construction cost. 463
439 the length of the project and if the project will be in an urban setting
or a rural setting. These two variables both require a low level of MRA Results 464
440
The same database was used with commercial software for MRA.
O

465
441 effort, thus are known to a high degree of confidence at the early
When the process was repeated with MRA, the rational selection 466
method proposed in this research also proved successful to meet 467
both objectives, as seen in Fig. 10. This helped to validate the se- 468
Analyzing the Performance of ANN Model
lection process. It is evident that the ANN model’s performance 469
250
was superior to the MRA, with a 25% error using ANN compared 470
to 50% error with MRA. These errors are both within the range 471
R
Estimating error [MAPE] (%)

200 suggested by the AASHTO Guideline to Cost Estimating 472

(AASHTO 2013) at the planning stage. The superior performance 473
150 of ANN is in agreement with several data-driven CCE models 474
found in the literature (Petroutsatou et al. 2012; Kim et al. 2004; 475
100
Moselhi and Siqueira 1998). However, this conclusion is not uni- 476
P

versal in the construction literature, with some authors reporting the 477
opposite findings (Gunduz et al. 2011; Setyawati et al. 2002). The 478
50 ongoing debate with both techniques was the reason that this re- 479
search tested the input variable selection framework with both 480
0 ANN and MRA. 481
0 5 10 15 20 25
With the MRA model, using the reverse order of input variables 482
Cumulative estimating effort (points)
never reaches the optimal prediction accuracy of around 50%. Also 483
Dual-objective selection method Reverse
the regression analysis actually performs better with fewer input 484
variables, and after a point the prediction error starts increasing. 485
F9:1 Fig. 9. ANN performance and effort expended
Without a rational input variable selection method, such as trial 486

© ASCE 7 J. Constr. Eng. Manage.

 Analyzing the Performance of MRA model Conclusion 527
250
Conceptual estimating models using ANN and MRA models were 528
Estimating error [MAPE] (%)

200 developed for MDT and debunked the conventional perception that 529
more input variables improves estimate accuracy. A rational input 530
selection method was proposed for that agency to select low-effort 531
150

Y
and high-impact input variables. Adding further input variables be- 532
yond 6–8, using either model technique, resulted in diminishing 533
100 returns for model performance. Findings from this research have 534
positive implications for that highway agency as less data collec- 535
50 tion and storage efforts are required if they are to employ data- 536
driven conceptual cost estimating techniques. 537

L
0 The paper’s primary contribution for both researchers and prac- 538
0 5 10 15 20 25 titioners is to highlight for the first time that although increasing the 539
Cumulative Estimating Effort (points) number of input variables in an early estimate may appear to en- 540
Dual-objective selection method Reverse hance estimate accuracy on an intuitive basis, this is not necessarily 541
true. The MDT case study showed that for both the ANN and MRA 542

N
F10:1 Fig. 10. MRA performance and effort expended approaches, adding detail to the model reached a point of dimin- 543
ishing returns at roughly 6 to 8 high-impact/low-effort variables for 544
that agency. 545

487 and error, commonly employed in the literature (Hegazy and Ayed
488 1998; Kim et al. 2004), one may conclude that a given set of data Acknowledgments 546

O
489 is not capable of predicting the construction costs to reasonable
490 accuracy. The authors would like to acknowledge Montana Department of 547
Transportation for furnishing the data and interviews and assisting 548
with the survey requirements for this research. We are also grateful 549
491 Discussion for their continued support, especially the assistance of Lesly 550
Tribelhorn and Kris Christensen. 551
492 The research in this paper has shown that a data-driven pavement
493 preservation CCE model developed for MDT does not need to in-
F
494 clude all project attributes to predict the construction cost to rea- References 552
495 sonable accuracy at an early stage of project development. This
496 research was conducted with data from only one highway agency AASHTO. (2013). “Practical guide to cost estimating.” Washington, DC. 553
497 and for one major work type; therefore, its conclusions cannot be Al-Tabtabai, H., Alex, A. P., and Tantash, M. (1999). “Preliminary cost 554
498 generalized. The researchers recognize that every agency may have estimation of highway construction using neural networks.” Cost Eng., 555
different business practices and working environments, which 41(3), 19–24. 6 556
499
O

Anderson, S., Molenaar, K., and Schexnayder, C. (2007). “Guidance for 557
500 could lead to a different result. Despite this limitation, the findings
cost estimation and management for highway projects during planning, 558
501 that continuous addition of variables may not necessarily increase programming and preconstruction.” NCHRP Rep. 574, National 559
502 the accuracy of early estimating is significant. Future research Cooperative Highway Research Program (NCHRP), Transportation 560
503 could analyze perceptional data to establish high-impact/low-effort Research Board of the National Academics, Washington, DC. 561
504 variables for other agencies. Bell, L. C., and Ghazanfer, A. B. (1987). “Preliminary cost estimating for 562
505 If this concept holds true for other highway agencies intending highway construction projects.” AACE Trans., C6.1–C6.4. 7 563
8
O

506 to employ data-driven methods for CCE, then the implications of Bode, J. (2000). “Neural networks for cost estimation: Simulations and 564
507 this research highlight that: pilot application.” Int. J. Prod. Res., 38(6), 1231–1254. 565
508 • A rational input selection method, such as the one suggested in Byrnes, J. E. (2002). “Best practices for highway project cost estimating.” 566
509 this paper, could be used to yield suitable input variables with M.S. thesis, Arizona State Univ., Mesa, AZ. 567
Creese, R. C., and Li, L. (1995). “Cost estimation of timber bridges using 568
510 low effort and contribute to acceptable performance;
neural networks.” Cost Eng., 34(5), 17–22. 9 569
511 • Once highway agencies are confident in the input variables re-
R

Danielsson, P.-E. (1980). “Euclidean distance mapping.” Comput. Graphics 570

512 quired to estimate the conceptual cost of projects, the collection Image Process., 14(3), 227–248. 10571
513 of further information is obsolete; it only consumes data storage Elhag, T. M. S., and Boussabaine, A. H. (1998). “An artificial neural system 572
514 space and requires time and effort from personnel whose efforts for cost estimation of construction projects.” Assoc. Res. Constr. 573
515 could be better applied elsewhere; and Manage., 1, 219–26. 11574
12
516 • The results imply that suitable confidence in estimating the con- Emsley, M. W., Lowe, D. J, Duff, A. R., Harding, A., and Hickson, A. 575
517 ceptual costs of projects can be achieved with lower project de- (2002) “Data modelling and the application of a neural network ap- 576
P

518 finition if the correct input variables are selected. proach to the prediction of total construction costs.” Constr. Manage. 577
519 The final implication of this study is the most important one Econ., 20(6), 465–472. 13578
520 for MDT: at the conceptual stage of a project lifecycle, using the FHWA (Federal Highway Administration). (2015). “Fact sheets on 579
521 models developed in this research, less effort is required to highway provisions—Statewide planning.” 〈http://www.fhwa.dot.gov/ 580
safetealu/factsheets/statewide.htm〉 (Aug. 17, 2015). 581
522 achieve satisfactory accuracy at the conceptual estimating stage.
Fink, A. (2009). How to conduct surveys: A step-by-step guide, 4th Ed., 582
523 This is better for a highway agency than unnecessarily expending SAGE Publications, Thousand Oaks, CA, 1–125. 583
524 design resources to improve project detail at the conceptual Flyvbjerg, B., Skamris Holm, M., and Buhl, S. (2002). “Underestimating 584
525 development stage for no improvement in the cost estimate costs in public works projects: Error or lie?” J. Am. Plann. Assoc., 585
526 accuracy. 68(3), 279–295. 586

© ASCE 8 J. Constr. Eng. Manage.

587 Fowler, F. J. (2009). Survey research methods, 4th Ed., SAGE Publications, Moselhi, O., Hegazy, T., and Fazio, P., (1992). “Potential applications of 622
588 Thousand Oaks, CA, 1–199. neural networks in construction.” Can. J. Civ. Eng., 19(3), 521–529. 21623
589 Gransberg, D. D., Lopez del Puerto, C., and Humphrey, D. (2007). “Relat- Moselhi, O., and Siqueira, I., (1998). “Neural networks for cost estimating 624
590 ing cost growth from the initial estimate to design fee for transportation of structural steel buildings.” AACE Int. Trans., 6.1–6.4. 22625
23
591 projects.” J. Constr. Eng. Manage., 10.1061/(ASCE)0733-9364(2007) Petroutsatou, C., Lambropoulos, S., and Pantouvakis, J.-P. (2006). “Road 626
592 133:6(404), 404–408. tunnel early cost estimates using multiple regression analysis.” Oper. 627
593 Gransberg, D. D., and Riemer, C. (2009). “Impact of inaccurate engineer’s Res. An Int. J., 6(3), 311–322. 24628

Y
594 estimated quantities on unit price contracts.” J. Constr. Eng. Manage., Petroutsatou, K., Georgopoulos, E., Lambropoulos, E., and Pantouvakis, J.-P. 629
595 10.1061/(ASCE)CO.1943-7862.0000084, 1138–1145. (2012). “Early cost estimating of road tunnel construction using neural 630
596 Gunaydin, H. M., and Dogan, S. Z., (2004). “A neural network approach networks.” J. Constr. Eng. Manage., 10.1061/(ASCE)CO.1943-7862 631
597 for early cost estimation of structural systems of buildings.” Int. J. .0000479, 679–687. 632
598 14 Project Manage., 22(7), 595–602. Pewdum, W., Rujirayanyong, T., and Sooksatra, V., (2009). “Forecasting 633
599 Gunduz, M., Ugur, L. O., and Ozturk, E. (2011). “Parametric cost estima- final budget and duration of highway construction projects.” Eng. Constr. 634

L
600 tion system for light rail transit and metro trackworks.” Expert Syst. Archit. Manage., 16(6), 544–557. 635
601 15 Appl., 38(3), 2873–2877. Rueda-Benavides, J. A., and Gransberg, D. D. (2014). “Indefinite delivery/ 636
602 Hegazy, T., and Ayed, A., (1998). “Neural network model for parametric indefinite quantity contracting: A case study analysis.” Transp. Res. Rec., 637
603 cost estimation of highway projects.” J. Constr. Eng. Manage., 10.1061/ 2408, 17–25. 25638
26
604 (ASCE)0733-9364(1998)124:3(210), 210–218. Sanders, S. R., Maxwell, R. R., and Glagola, C. R. (1992). “Preliminary 639
605 Jeong, H. S., and Woldesenbet, A. (2012). “Procedures and models for estimating models for infrastructure projects.” Cost Eng., 34(8), 7–13. 27640

N
606 estimating preconstruction costs of highway projects.” Oklahoma Schexnayder, C. J., Weber, S. L., and Fiori, C. (2003). “NCHRP synthesis 641
17
607 16 Transportation Center. of highway practice: Project cost estimating.” Transportation Research 642
608 Kim, G.-H., An, S.-H., and Kang, K.-I. (2004). “Comparison of construction Board of the National Academics, Washington, DC. 643
609 cost estimating models based on regression analysis, neural networks, Setyawati, B. R., Sahirman, S., and Creese, R. C. (2002). “Neural networks 644
610 18 and case-based reasoning.” Build. Environ., 39(10), 1235–1242. for cost estimation.” AACE Int. Trans., 13.1–13.9. 28645
611 Kim, H., Seo, Y., and Hyun, C., (2012). “A hybrid conceptual cost estimating Smith, A. E., and Mason, A. K. (1997). “Cost estimation predictive mod- 646

O
612 19 model for large building projects.” Autom. Constr., 25, 72–81. eling: Regression versus neural network.” Eng. Econ., 42(2), 137–161. 29647
613 Lowe, D. J., Emsley, M. W., and Harding, A. (2006). “Predicting construc- Turochy, R. E., Hoel, L. A., and Doty, R. S. (2001). “Highway project cost 648
614 tion cost using multiple regression techniques.” J. Constr. Eng. estimating methods used in the planning stage of project development 649
615 Manage., 10.1061/(ASCE)0733-9364(2006)132:7(750), 750–758. VTRC 02-TAR3.” Virginia Transportation Research Council, 1–290. 30650
616 Mahamid, I. (2011). “Early cost estimating for road construction projects Verlinden, B., Duflou, J. R., Collin, P., and Cattrysse, D. (2008). “Cost 651
617 using multiple regression techniques.” Australas. J. Constr. Econ. estimation for sheet metal parts using multiple regression and artificial 652
618 Build., 11(4), 87–101. neural networks: A case study.” Int. J. Prod. Econ., 111(2), 484–492. 31653
619 Minassian, V. K., and Jergeas, G. F. (2009). “A prototype risk analysis Walczak, S. (2001). “An empirical analysis of data requirements for finan- 654
620 for determining contingency using approximate reasoning method.” cial forecasting with neural networks.” J. Manage. Inf. Syst., 17(4), 655
F
621 20 Cost Eng., 51(1), 26–33 203–222. 32656
O
O
R
P

© ASCE 9 J. Constr. Eng. Manage.

 Queries
1. ASCE Open Access: Authors may choose to publish their papers through ASCE Open Access, making the paper freely available
to all readers via the ASCE Library website. ASCE Open Access papers will be published under the Creative Commons-Attribu-
tion Only (CC-BY) License. The fee for this service is $1750, and must be paid prior to publication. If you indicate Yes, you will
receive a follow-up message with payment instructions. If you indicate No, your paper will be published in the typical subscribed-

Y
access section of the Journal.

2. Please provide the ASCE Membership Grades for the authors who are members.

3. "The online citation name format is shown below. Please confirm each author’s surname and first initial are shown correctly.-

L
Gardner, B; Gransberg, D. D; David Jeong, H"

4. Please check the hierarchy of section heading levels.

5. Please check the edits made in Fig. 2 captions.

N
6. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.

7. This query was generated by an automatic reference checking system. This reference could not be located in the databases used

O
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.

8. Please provide the volume and issue number for Ref. Bell and Ghazanfer (1987).

9. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.
F
10. Issue number ’3’ has been inserted in this reference. Please check and confirm the edit made here.

11. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
O

articles as possible.

12. Please provide the issue number for Ref. Elhag and Boussabaine (1998).

13. Issue number ’6’ has been inserted in this reference. Please check and confirm the edit made here.

14. Issue number ’7’ has been inserted in this reference. Please check and confirm the edit made here.
O

15. Issue number ’3’ has been inserted in this reference. Please check and confirm the edit made here.

16. This reference Jeong and Woldesenbet (2012) is not mentioned anywhere in the text. ASCE style requires that entries in the
References list must be cited at least once within the paper. Please indicate a place in the text, tables, or figures where we may
R

insert a citation or indicate if the entry should be deleted from the References list.

17. Please provide the publisher’s location for Ref. Jeong and Woldesenbet (2012).

18. Issue number ’10’ has been inserted in this reference. Please check and confirm the edit made here.
P

19. Please provide the issue number for Ref. Kim et al. (2012).

20. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.

21. Issue number ’3’ has been inserted in this reference. Please check and confirm the edit made here.

© ASCE 10 J. Constr. Eng. Manage.

22. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.

23. Please provide the volume and issue number for Ref. Moselhi and Siqueira (1998).

Y
24. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.

25. Please provide the issue number for Ref. Rueda-Benavides and Gransberg (2014).

L
26. Volume number ’2408’. has been inserted in this reference. Please check and confirm the edit made here.

27. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced

N
articles as possible.

28. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.

O
29. A check of online databases revealed a possible error in this reference. The lpage has been changed from ’167’ to ’161’. Please
confirm this is correct.

30. Please provide the publisher’s location for Ref. Turochy et al. (2001).

31. Volume number ’111’. has been inserted in this reference. Please check and confirm the edit made here.Issue number ’2’ has been
inserted in this reference. Please check and confirm the edit made here.
F
32. This query was generated by an automatic reference checking system. This reference could not be located in the databases used
by the system. While the reference may be correct, we ask that you check it so we can provide as many links to the referenced
articles as possible.
O
O
R
P

© ASCE 11 J. Constr. Eng. Manage.

View publication stats