Predictive Modeling for Transportation Infrastructure Preconstruction Costs

Abstract

This article presents a model to enable managers to make better decisions regarding infrastructure construction management. The model accurately estimates preliminary engineering (PE) costs (synonymous with Preconstruction Services Costs [PCS]) of bridge infrastructure projects so that state, local, and private transportation departments can know what these activities actually cost and can better plan future budgets. It also enables managers of these public works to quickly identify potentially troublesome projects, thus enabling corrective actions to be initiated early. This article describes a comprehensive study of the factors affecting PE costs, a database containing 2001-2009 bridge project data, and models to estimate PE costs. The findings indicate that bridge projects exhibited a historical mean PE cost ratio of 27.8% of estimated construction costs, significantly more than expected or commonly believed.

Keywords

project management preliminary engineering preconstruction services cost estimation bridge preconstruction regression modeling

During a project’s preconstruction phase, PE encompasses both planning activities and engineering design (both preliminary and detailed). PE efforts begin years in advance of the project’s construction being released for construction bids. Typically, departments of transportation estimate PE costs as a fixed percentage of construction costs disregarding other project-specific parameters. An efficient and accurate method to estimate PE costs would benefit state, local, and private transportation departments by enabling them to accurately know what these activities actually cost. This would enable public works managers to better plan future budgets.

This article describes a comprehensive study of the factors affecting PE costs, a database containing 2001-2009 bridge project data from the North Carolina Department of Transportation (NCDOT), and models to estimate PE costs more accurately using both multiple linear regression (MLR) and multilevel Dirichlet process linear modeling (MDPLM), a nonparametric regression technique. The findings indicate that bridge projects exhibited a historical mean PE cost ratio of 27.8% of estimated construction costs, significantly more than expected or commonly believed. Furthermore, we found that when both analytical techniques were applied to a validation sample, the MDPLM was more effective in predicting high PE cost ratios (that were greater than twice the historical mean) achieving an R² value of .78 compared to an R² of .33 for MLR for these projects.

Background

Over the last 30 years, public works projects have increased in number, complexity, and cost, especially during the recent American Recovery and Reinvestment Act of 2009. This has been especially true in the transportation sector. Accordingly, the accuracy of project costs overall (and consequently preconstruction costs) has become a larger concern, both for the managing agencies and the public taxpayer. To address this concern initial focus was placed on construction cost accountability. Managing agencies have more recently adopted tighter financial controls applied to additional project cost components such as right-of-way (ROW), utilities, mitigation, and PE costs (synonymous with Preconstruction Services Costs [PCS]).

A search of infrastructure literature determined that most public transportation agencies (at all government levels) use a constant or sliding percentage of estimated construction costs to develop a PE budget. The most frequent percentage cited is 10% of estimated construction costs (Washington State Department of Transportation [WSDOT], 2002). However, there can be a wide range in PE costs depending on project type and complexity. Although extreme (high and low) PE costs may be considered exceptional or unusual, early identification of these situations can focus public works management’s attention onto such projects before PE costs begin to escalate. This article focuses on such atypical projects to enable their rapid identification.

Intuitively, factors such as project type, project complexity, when PE was conducted, and whether PE efforts were performed in-house or by consultants, should have significant impacts on PE costs. However, there has not yet been a comprehensive study defining the full set of factors and analyzing their effects on PE costs. In the literature there is a large body of work on cost estimation in general, and on construction process cost, but little is available that specifically focuses on PE cost estimation as we do herein.

Preliminary Engineering (PE) Defined

We define PE as the effort required to both plan and design (preliminary and detailed) a project for construction. PE begins when policy makers first authorize funding for a specific project for planning or design activities. The delivery of the construction documents for project bidding marks the end of PE. In general, projects have three cost components: PE, ROW, and construction costs. Consistent with other researchers’ definitions, PE in this study does not include ROW acquisition or construction costs (Turochy, Hoel, & Doty, 2001; WSDOT, 2002). It is important also to point out that we are using the term PE synonymously with preconstruction services costs.

In addition, other project costs such feasibility studies or mitigation efforts are excluded from PE because they are tracked separately. We also used a time marker to restrict PE costs to only those costs occurring AFTER a project is identified or funding is authorized, and BEFORE the project’s construction contract has been formalized.

Significance of the Research

PE costs usually comprise a significant portion of total project costs. Accurate PE cost estimation can help public works managers make the best possible programming and budgeting decisions. With better PE cost estimates, funding allocations can be proactive, matching the specific needs of each project. This article reports on a method to improve the accuracy of determining PE costs. The data collected and analyzed during this research were specific to the NCDOT. However, the research findings should prove helpful to other public agencies at all governmental levels as well.

Many State DOTs estimate PE costs using a fixed percentage method, which turns out to be inefficient over the project cycle as well as inaccurate. Some projects require less PE funding while others require more and may neccessitate the need for a new funding authorization. Such underallocation or overallocation necessitates management actions to redistribute PE funds. Avoiding such redistributions improves total project cost control. In addition, by having a project-specific PE cost estimate generated at the beginnning of each project’s preconstruction phase, the PE budget status becomes trackable as a performance metric.

Literature Review

In a 2001 study, the Virginia Transportation Research Council (VTRC) reported on the current state of practice among nine DOTs with regard to cost estimating of highway projects during the planning phase. The VTRC researchers noted that, “ROW and PE are the states’ most difficult cost categories to estimate and often present the greatest challenges and deviations during the cost estimation process” (p. 18). Most respondents reported that PE costs were derived from a percentage of estimated construction costs, ranging between 5% and 20%. Texas estimates PE cost as a function of ROW width on some projects. Delaware utilizes a detailed form to guide how PE costs should be estimated based on project size. VTRC suggested that the PE cost of large projects can be estimated as a percentage of construction costs, whereas small projects should estimate required man-hours to determine PE costs. Moderate sized projects may utilize a combination of both estimating methods (Turochy et al., 2001).

As part of a comparative analysis of construction costs, Washington State DOT (WSDOT, 2002) collected information from 25 DOTs whose members served on the AASHTO Subcommittee on Design. The average PE cost among respondents was 10.3% of construction costs and with a range of costs reported between 4% and 20%. NCDOT participated in the survey and reported PE costs of 10% of construction costs. Figure 1 summarizes geographically the PE costs acquired from the two surveys. Responses from 28 DOTs were acquired and have been mapped in Figure 1 (Turochy et al., 2001; WSDOT, 2002).

Figure 1.

DOT reported PE costs as a percentage of construction costs.

Schofer et al. (2010) reported on the recommendations developed during a conference organized by the Transportation Review Board and sponsored by the U.S. DOT Research and Innovative Technology Administration. The conference aimed to define essential research directions needed to manage and preserve the nation’s surface transportation infrastructure. Conference participants shared recommendations aligned into six research themes, one of which was valuation methods to support infrastructure management processes. Schofer et al. (2010) identified the need to “develop objective, quantitative, and monetary methods and models” to determine the cost required to keep the transportation system in a state of good repair.

Planning Component of PE

Just one reference on estimating the PE costs associated with planning could be found in the literature. WSDOT noted in their 2002 survey that the preconstruction efforts required to meet environmental compliance requirements are highly variable between projects. Instead of asking survey respondents to quantify environmental compliance costs, WSDOT attempted to capture how these costs typically change during the preconstruction phase. Twenty-one of the 25 respondents (84%) indicated variability ranges from 0% to 10%. Three other respondents (12%) indicated higher variability in the 11% to 20% range (WSDOT, 2002).

Design Component of PE

More information on the design component of PE can be found in the literature than on the planning component. The work of Nassar, Hegab, and Jack (2005) provides a significant contribution to the literature specifically addressing PE design costs for transportation projects. Nassar et al. sought to create a model to estimate costs of design consultants’ efforts. The model was based on data from 59 Illinois Department of Transportation projects. The best fit model was a log transformation using only one independent variable, the initial estimated construction cost, to predict consultant design costs (Nassar et al., 2005). The prediction error of the model was not reported.

Gransberg, Lopez Del Puerto, and Humphrey (2007) investigated the correlation between design fees and construction “cost growth from the initial estimate” (termed CGIE). Using 31 projects of the Oklahoma Turnpike Authority, Gransberg et al. confirmed that as design fees decrease, construction cost growth (from initial estimate to final closeout) increases. Correlating design fees to design quality, the results support the premise that allocating sufficient funding in design reduces the likelihood of construction cost increases from the initial estimate and not allocating sufficient initial design funding results in construction cost increases. Gransberg et al. determined that the CGIE for all projects in the study was 9.65%. In addition, Gransberg et al. reported that design costs for roads were 2% and for bridges design costs were 7.6% of construction costs. The researchers concluded that “bridge design projects should command a relatively higher design fee than roadway projects due to the increased complexity of design” (Gransberg et al., 2007, p. 407).

In-House Versus Consultant PE Providers

Wilmot, Deis, Schneider, and Coates (1999) presented a concise summary of 17 studies that compared the design costs of in-house staff with that of consultants. These studies typically concluded that consultant design costs were greater than in-house costs. Interestingly however, the magnitude of this difference varied significantly depending on the comparison methodology utilized. Wilmot et al. analyzed 37 projects (20 designed in-house and 17 designed by consultants) completed between 1995 and 1997 by the Louisiana Department of Transportation and Development. After including cost factors such as overhead rates, space rental, and insurance, the overall comparison found that in-house costs are approximately 80% of consultants’ costs. This difference was found to be statistically significant at the 95% level. However, the difference was largely accounted for by the additional in-house effort required to prepare and supervise the consultants’ contract (Wilmot, 1999).

VDOT had previously studied the cost difference between roadway design performed in-house compared with using consultants (Kyte, Perfater, Haynes, & Lee, 2004). Consultant design costs were found to be 50% higher than in-house designs. The cited research of both Wilmot et al. (1999) and Kyte et al. (2004) addressed design functions. Both researchers agree that consultant design costs more than in-house design.

Gen and Kingsley (2007) performed a detailed investigation into the use of consultants by Georgia DOT. Consultants’ use to perform engineering design had grown from 10% to 60% over a 15-year span. GDOT evaluated the effectiveness of consultant use by the “quality of work, speed of production, and notably, managerial burden” of consultant contract administration. Cost savings was not reason for consultant selection, and generally consultants’ preconstruction services cost more than similar services performed by agency staff (Gen & Kingsley, 2007).

This article reports on the allocation of costs for both planning and design for selected NCDOT projects. A correlation between PE planning costs and PE provider, similar to the findings reported above for design costs, is suggested.

Estimating Practices in the Transportation Industry

DOT personnel perform project estimating for most projects. Approximately half of all the state DOTs organize their estimators into a unit dedicated to estimating. In 42 states where external consultants also prepare cost estimates, DOT personnel review those estimates in detail. DOTs use three general approaches to estimating (Schexnayder, Weber, & Fiori, 2003):

Parametric estimating using historical cost figures;

Detailed estimating using quantity takeoff and pricing of labor, equipment, and materials;

A combination of parametric and detailed techniques.

Parametric estimating can be used when scoping information is very limited (project definition level is less than 5%; Anderson, Molenaar, & Schexnayder, 2007). Parametric estimates rely on historical cost databases and defined relationships between cost items. Many DOTs use Trns*port, an AASHTO sponsored software package for parametric estimating (Schexnayder et al., 2003).

When projects are fully scoped and detailed design efforts are underway (project definition level exceeds 50%), detailed estimating techniques are commonly used. Detail estimating of scope line items use quantity takeoffs with material, labor, and equipment pricing, but Schexnayder et al. (2003) found that DOTs performing detailed estimates do so only for the major work items and that these account for 65% to 80% of project costs. The remaining work items are estimated using a combination of parametric tools (Schexnayder et al., 2003).

Infrastructure Management Systems

Investigation into state’s infrastructure management systems by Jimenez and Pagano (2011) showed that states are generally improving in their capital planning efforts and project monitoring efforts. For example, NC’s performance on these two factors was rated as a weakness in 2005 but improved to a “midlevel” grade in 2008. States are seeking to improve their infrastructure management systems, and use of a PE estimating tool can be one strategy to assist in that improvement, especially in the project monitoring arena.

Miller and Lantz (2010) interviewed 27 project professionals inquiring about the role and impact of project scoping on achievement of project outcomes. Their findings indicated problems associated with scoping efforts exist and potential process changes to improve project scoping are possible. Since scoping establishes a project’s cost, timing, and purpose (in response to an identified need), successful scoping is associated with efficient project management. Responses from interviews conducted by Miller and Lantz (2010) indicated that “imperfections” in planning and programming (a significant component of PE) “adversely affect scoping.”

Dirichlet Modeling

In response to increasing complexity of data, modeling methods have become sophisticated and complicated. However, the current art of prediction modeling for complicated data sets is either human-aided or a brute-force search to identify the covariate–response relationship (linear or nonlinear), the form of the variance component (homoscedasticity or heteroscedasticity), the degree of heterogeneity (the number of statistically different components), the choice of covariate (fixed or random effects), or the existence of interactions between covariate variables or component variances, for example. Such uncertainty factors have significant effects on both prediction performance and computational efforts. Unfortunately, there is no model or systematic way of modeling that can deal with all such uncertainties simultaneously.

The goal of this study was to identify a way of modeling complex data, such as bridge projects’ PE cost ratio, to create an accurate prediction model. In contrast to existing models, where a model-maker resolves uncertainty through numerous statistical tests, this proposed model is created in a layer-by-layer fashion with minimal assumptions on uncertainty.

A Dirichlet process linear model (DPLM) is a nonparametric regression technique suitable for handling a heterogeneous data population. A data set is assumed to be sampled from a heterogeneous population if the sample can be partitioned into sets with each having statistically significant different parameters. The prediction capability of DPLMs is robust because its averaging estimate over competing models accommodates the variation of prediction and avoids overfitting. A DPLM is also flexible in that the degree of heterogeneity (the number of statistically different mixture components) does not need to be decided before modeling. Therefore DPLM provides a convenient way of minimizing statistical tests and arbitrary thresholds decisions.

DPLM assumes that random error of mixture components is independent and identically distributed; therefore it has difficulty in modeling the case of data where variance components exhibit heteroscedasticity and interactions. To address this difficulty, a Dirichlet process generalized linear model (DPGLM; Hannah, Blei, & Powell, 2010; Mukhopadhyay & Gelfand, 1997) allows for a nonlinear relationship between the linear predictor and response mean, similar to other generalized linear models (McCullagh & Nelder, 1989; Wedderburn, 1974). The DPGLM model also makes possible the change of random error based on a covariate by incorporating variance components into the linear predictor. However, the choice of the covariate–response distribution and the form of the variance component requires human efforts throughout possible candidates, and the random errors of mixture components are still assumed to independent.

Thus we chose to use a multilevel Dirichlet process linear model (MDPLM) to overcome the deficiencies of DPGLM models while keeping the simple assumptions of DPLM. The MDPLM can be considered to be a nonlinear system where the data set is input, a performance measure is output, and the random effect variables generated in layers are noises (Benzi, Parisi, Sutera, & Vulpiani, 1983; McNamara & Wiesenfeld, 1989).

Method

Our analytical techniques described in this section focus on two modeling approaches applied to a NCDOT bridge project data set: MLR and MDPLM. These approaches were undertaken simultaneously by separate members of the project team so that a comparison could be made regarding the predictive value of each approach. This article introduces the MDPLM model, presents MDPLM modeling results, and compares the results from MDPLM to MLR.

Database Compilation

The research team initially obtained project descriptive data, cost estimates, and actual cost expenditures for bridge projects let for construction from NCDOT during January 1, 1999 through June 30, 2008. We queried NCDOT’s project management systems to acquire actual PE costs for the bridge projects during this letting period. Our database relied heavily on NCDOT’s data collected for the National Bridge Inventory System (NBIS) that meets the standardized Federal Highway Administration (FHWA) inspection and reporting guidelines. However, complete NBIS bridge project data were only available for projects let since January 2001. To keep the target data set size at approximately 500 projects, data on bridge projects let between July 2008 and June 2009 were added to the data set.

A typical NCDOT bridge project involves replacement of an existing structure, using cast in place concrete, over water, 30 m wide by 12 m long. Median construction cost is US$910,000 and the mean construction cost is US$1.2 million.

The ratio of PE costs to estimated construction costs was selected as the preferred response (dependent) variable for cost regression analyses. Using a cost ratio allowed modeling across all project sizes and eliminated conversion of cost values to a common base year to account for inflation. The ratio of actual PE cost to the estimated construction cost was tabulated for all bridge projects. This ratio is referred to as the project’s PE cost ratio.

Figure 2 presents a comparative history of the annual number of NCDOT bridge projects and the annual mean PE cost ratio for 2001 through midyear 2009. The historical mean PE cost ratio for this time period was 27.8%. The 95% confidence interval (CI) for the mean PE cost ratio was 26.0% to 29.6%. Note that the mean PE cost ratio for years 2001-2005 was 24.0%, and for 2007-2009, the mean increased to 39.6%. The year 2006, with a mean of only 9.2%, indicated an anomaly although no specific agency procedural changes could be identified to account for these changes.

Figure 2.

Historical comparison of NCDOT bridge projects.

The database compilation process was first introduced in Hollar, Arocho, Hummer, Liu, and Rasdorf (2010) and more detailed discussion can be found in this reference.

Multiple Linear Regression (MLR) Modeling

Selecting the “best” model using multiple linear regression (MLR) can be difficult if there are a large number of independent variables. To assist in MLR model selection considering our 28 variables, we utilized the GLMSLECT procedure within the Statistical Analysis Software (SAS) package. In addition to forward-, backward-, and stepwise-selection, GLMSELECT provides two additional variable selection methods: least angle regression (LAR) and least absolute shrinkage and selector operator (LASSO). The GLMSELECT procedure provides an efficient starting point for model selection. Model refinement can then follow using intuitive insights gained from data familiarity (Cohen, 2006).

Before MLR processing, the 2001-2009 bridge data set was divided into a modeling set (containing 391 projects) and a validation set (containing 70 projects). Only the modeling set was used for constructing MLR models. We considered both numerical and categorical variables when utilizing the GLMSELECT procedure for variable selection. We compared the adjusted R² values associated with the GLMSELECT iterations to assess model fit. After considering the input requirements from the user’s perspective, the recommended MLR model utilized four numerical (N) and four categorical (C) variables including first-level interactions, achieving an adjusted R² of .33 with a MAPE of 43% when applied to the validation set. The eight selected variables are listed below:

Right of Way Cost to Estimated Construction Cost (N);

Roadway Percentage of Construction Cost (N);

Estimated Construction Cost (N);

Bypass Detour Length (N);

Project Construction Scope (C);

NCDOT Division (C);

Geographical Area of State (C);

Planning Document Responsible Party (C).

Hierarchical Stratified Data Subsets

To address the heterogeneous nature of the data, we investigated applying MLR to data subsets created from the hierarchical classification shown in Figure 3. Three of the four categorical variables identified by MLR model development were used to create the hierarchical subsets. Following project assignment to a specific subset, MLR modeling was repeated using the four numerical variables previously identified and adding a fifth numerical variable, project length.

Figure 3.

Hierarchical data subsets subjected to MLR analyses.

Liu et al. (2011) provides the complete set of regression parameters for all models fit to each subset of bridge projects. Overall, applying traditional MLR techniques to smaller subsets of projects did not provide a significant improvement in prediction capability compared to using MLR applied to the entire database. Often we encountered insufficient numbers of projects in some subsets to support robust regression analyses.

The MLR regression analyses supported our assumptions that project data are heterogeneous, that the relationships between the PE cost ratio and other variables are complicated, and that multiple relationships exist between variables. The modeling of these complex situations may be improved by moving beyond traditional MLR regression techniques and employing a more sophisticated technique such as MDPLM as described in the next section.

Multilevel Dirichlet Process Linear Model (MDPLM)

Estimating a bridge project’s PE cost ratio is difficult because the relationships between PE cost ratio and other variables are complicated and difficult to define. The multilevel Dirichlet process linear model (MDPLM) can deal with more complex situations than existing regression techniques (such as generalized linear mixed models; Breslow & Clayton, 1993), multilevel mixed linear or nonlinear models (Goldstein, 1986, 1991), or Dirichlet process generalized linear models (Hannah et al., 2010; Mukhopadhyay & Gelfand, 1997) and produce more robust estimates with less variance. An MDPLM consists of DPLM layers using different variables and is a sequential variance estimate model. The variance of each layer is estimated in the successive layer.

The bridge project data set contains heterogeneous data with uncertainty regarding the variables and response relationships, interactions among variables, degree of heterogeneity, and variance. In contrast to other modeling techniques (where a model-builder resolves uncertainty through numerous statistical tests), MDPLM builds a model layer-by-layer with minimal assumptions regarding uncertainty. It is not necessary to describe how data are distributed or structured, nor specify the random effects. MDPLM can fit complex data with less variance. By finding hidden or unobservable random effects, the reduction of variance in prediction is achieved. The modeling procedure decides the optimal layer of prediction to avoid overfitting by using cross-validation and stochastic resonance theory techniques.

We used MDPLM to construct a model to predict the PE cost ratio of bridge projects. The data set was slightly different from the MLR data set and contained 505 bridge projects let for construction between January 1999 and June 2008. Unlike the MLR analysis, the MDPLM predictive variables did not rely on NBIS data thereby allowing use of the 1999 and 2000 bridge data initially collected. The MDPLM technique used 13 independent variables. The response was the PE cost ratio (Lee, 2011). The 13 variables are listed below. In comparison to the MLR model, variables 1 to 5 were used by both modeling methods (MLR and MDPLM).

Right of Way Cost to Estimated Construction Cost;

Roadway Percentage of Construction Cost;

Estimated Construction Cost;

Geographical Area of State;

Planning Document Responsible Party;

Structure Percentage of Construction Cost;

Road System Classification;

NEPA Document Classification;

Project Length;

Roadway Functional Classification;

Estimated Right-of-Way Cost;

Number of Spans;

Structure Type (bridge or culvert).

From the DPLM (trained by 505 bridge project observations), the expected number of mixture components was approximately 10. The size of a validation set was set to 25. We set the number of training-validation data set pairs to 10 for cross-validation. Figure 4 shows the predictive performance of the MDPLM averaged over 10 training-validation data set pairs by layer transition. Mean square error (MSE), mean absolute error (MAE), and mean absolute relative error (MARE) were used as base error measurements in panels (a), (b), and (c), respectively. The prefix “t” or “v” designates which data set (training or validation) was used to compute the error measurements. Throughout the plots in Figure 4, a clear trend is observed for the validation data sets (dashed line): error measures, E(vMSE), E(vMAE), and E(vMARE), decrease as the layer of prediction advances from 0 to 4, but then error measures start to increase after Layer 4. With the stochastic resonance theory, we have showed that the prediction at Layer 4 balances observable (training) and unobservable (validating) data. Layer 4 provides the optimal predictive performance for the MDPLM.

Figure 4.

MDPLM expected error measurements.

The final bridge model was trained with all 505 projects generating the layer-by-layer scatter plot of predicted values shown in Figure 5. Layer 0 is equivalent to a DPLM model. As the MDPLM layer increases from 0 to 4 (associated with increasing model complexity), the variance of the predicted values decreases. The advantage of MDPLM over DPLM is evident by the obvious variance reduction as the additional Layers 1 through 4 of the MDPLM were created.

Figure 5.

MDPLM layer-wise predictive results.

Results, Conclusions, and Recommendations

As previously noted, the recommended MLR model utilized eight predictor variables (with first-level interactions) and achieved an adjusted R² of .33 with a MAPE of 43% when applied to the validation set. The predictor variables used in MLR modeling relied heavily on data from the National Bridge Inventory System (NBIS). These variable values were available for NCDOT bridge projects let for construction between January 2001 and June 2009. The MLR model was developed using 461 bridge projects.

The MDPLM used 13 predictor variables and minimized error measures at the fourth layer. The model definition established at Layer 4 was applied to the same validation set used in MLR analysis resulting in an R² of .79 with a MAPE of 22%. The MDPLM analysis used a data set of 505 NCDOT bridge projects let between January 1999 and June 2008. This data set differs from the MLR data set because the MDPLM analyses did not rely on data fields related to the NBIS.

Figure 6 conveniently illustrates the comparative results of the MLR and MDPLM techniques in predicting PE cost ratio of 70 bridge projects used as a validation set. MDPLM generated a higher R² value (.79) than MLR (.33) and, most notably, shows that MDPLM is superior to MLR in predicting a project’s PE cost ratio in situations where the cost ratio exceeds the mean (27.8 %) by a factor of 2 or more. That is, the MDPLM technique handles data outliers.

Figure 6.

Comparison of MDPLM and MLR results.

The actual PE cost ratios range from 2.7% to 152%. The MLR model generated a range of predictions from 13.3% to 58.8%, whereas the MDPLM predictions range from 5.8% to 142%. Thus the MDPLM is superior in predicting across the full range of historical PE cost ratios observed, especially in the range of high PE cost ratio values. We conclude that although high PE cost ratios may be considered exceptional or unusual, the MDPLM model identifies them, thus enabling early detection to focus public works managers to pay attention to such projects before PE costs begin to escalate.

NCDOT is currently employing both MLR and MDPLM models in an agency study of recent bridge projects. Although the MDPLM results are deemed more accurate, interpretation of variable effects and sensitivity is difficult because of the multilevel structure. Therefore, the agency is using the MDPLM model to predict a project’s PE cost ratio and using the MLR model to determine variable sensitivity effects.

The authors recommend that current bridge projects (let since 2009) be incorporated into the data set and the analyses repeated. Furthermore, since the 2007-2009 mean PE cost ratio had shown a significant increase over earlier years, a newer data set (only going back to 2007) is recommended for future analyses along with identifying the causes of such a significant PE cost ratio shift in 2007. We also recommend that highway data be modeled in a similar manner and that significant shifts in highway project PE costs also be identified.

The authors reviewed general data collection recommendations shared by Pantelias, Flintsch, Bryant, and Chen (2009). Pantelias et al. (2009) noted that “past practices and staff culture” still largely influence what data is collected and data is typically housed in databases supporting a single asset class (such as bridges or pavements). However, Pantelias et al. (2009) found that most agencies are planning for integrated data systems (across asset classes) and that the agencies are seeking to “optimize the overall agency data collection activities” to meet asset management needs. Within this study, data quality and characteristics were of potential concern. Having access to accurate data would allow further modeling of bridge PE costs and also allow modeling of other project types.

In addition, due to the reliance on bridge project data being extracted from data sets prepared for the National Bridge Inventory System, the authors suggest use of a data management and extraction utility tool such as the tool developed by Karlaftis, Kepaptsoglou, and Lambropoulos (2005).

Footnotes

Authors’ Note

The contents do not necessarily reflect the official views or policies of either the North Carolina Department of Transportation or the Federal Highway Administration at the time of publication.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors acknowledge the research funding provided by the North Carolina Department of Transportation Research and Development Unit as part of NCDOT project 2010-10.

Author Biographies

Donna A. Hollar is an assistant professor of construction management at East Carolina University. Her research interests include preconstruction project cost analyses, optimization methods and modeling, and GIS applications applied to infrastructure sustainability.

William Rasdorf is a professor of civil engineering at North Carolina State University. His interests span construction, transportation, environmental, and computer-aided engineering. His research focuses on information technology and the modeling of engineering processes, facilities, and operations.

References

Anderson

Molenaar

Schexnayder

(2007). Guidance for cost estimation and management for highway projects during planning, programming, and preconstruction (NCHRP Report 574). Washington, DC: Transportation Research Board.

Benzi

Parisi

Sutera

Vulpiani

(1983). A theory of stochastic resonance in climatic change. SIAM Journal on Applied Mathematics, 43, 565-578.

Breslow

N. E.

Clayton

D. G.

(1993). Approximate inference in generalized linear mixed models. Journal of the American Statistical Association, 88(421), 9-25.

Cohen

R. A.

(2006). Introducing the GLMSELECT PROCEDURE for model selection (Paper 207-31). Proceedings of the Thirty-first Annual SAS Users Group International Conference, SAS Institute, Cary, NC. Retrieved from http://www2.sas.com/proceedings/sugi31/207-31.pdf

Gen

Kingsley

(2007). Effects of contracting out engineering services over time in a state Department of Transportation. Public Works Management & Policy, 12, 331-343.

Goldstein

(1986). Multilevel mixed linear model analysis using iterative generalized least squares. Biometrika, 73(1), 43-56.

Goldstein

(1991). Nonlinear multilevel models, with an application to discrete response data. Biometrika, 78(1), 45-51.

Gransberg

D. D.

Lopez Del Puerto

Humphrey

(2007). Relating cost growth from the initial estimate to design fee for transportation projects (American Society of Civil Engineers, Reston, VA). Journal of Construction Engineering and Management, 133, 404-408.

Hannah

L. A.

Blei

D. M.

Powell

W. B.

(2010, May 13-15). Dirichlet process mixtures of generalized linear models (pp. 313-320). Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, Chia Laguna Resort, Sardinia, Italy.

10.

Hollar

Arocho

Hummer

Liu

Rasdorf

(2010). Development of a regression model to predict preliminary engineering costs. Proceedings of the 2010 Technical Conference and Exhibit: Meeting Transportation’s 21st Century Challenges, Institute of Transportation Engineers, Savannah, GA.

11.

Jimenez

Pagano

(2011). What factors affect management quality? State infrastructure management and the government performance project. Public Works Management & Policy, 17, 124-151.

12.

Karlaftis

Kepaptsoglou

Lambropoulos

(2005). A web-supported national bridge inventory management tool. Public Works Management & Policy, 9, 248-258.

13.

Kyte

C. A.

Perfater

M. A.

Haynes

Lee

H. W.

(2004). Developing and validating a tool to estimate highway construction project costs. Transportation Research Record: Journal of the Transportation Research Board, 1885, 35-41.

14.

Lee

(2011). Multilevel Dirichlet Process Linear Model (MDPLM) for heterogeneous data (Unpublished doctoral dissertation). North Carolina State University, Raleigh, NC. Retrieved from http://www.lib.ncsu.edu/resolver/1840.16/7006

15.

Liu

Hummer

J. E.

Rasdorf

Parikh

S. C.

Lee

Gopinath

Hollar

D. A.

(2011). Preliminary engineering cost trends for highway projects (Technical Report No. FHWA/NC/2010-10). Raleigh: North Carolina Department of Transportation.

16.

McCullagh

Nelder

(1989). Generalized linear models (2nd ed.). London, UK: Chapman and Hall.

17.

McNamara

Wiesenfeld

(1989). Theory of stochastic resonance. Physical Review A, 39, 4854-4869.

18.

Miller

Lantz

(2010). Options for improving transportation project scoping. Public Works Management & Policy, 14, 332-350.

19.

Mukhopadhyay

Gelfand

A. E.

(1997). Dirichlet process mixed generalized linear models. Journal of the American Statistical Association, 92, 633-639.

20.

Nassar

K. M.

Hegab

M. Y.

Jack

N. W.

(2005, April). Design cost analysis of transportation projects (pp. 949-956). Proceedings of the Construction Research Congress 2005: Broadening Perspectives, San Diego, CA.

21.

Pantelias

Flintsch

G. W.

Bryant

J. W.

Chen

(2009). Asset management data practices for supporting project selection decisions. Public Works Management & Policy, 13, 239-252.

22.

Schexnayder

C. J.

Weber

S. L.

Fiori

(2003). Project cost estimating—A synthesis of highway practice (NCHRP Project 20-7, Task 152). Washington, DC: Transportation Research Board.

23.

Schofer

J. L.

Evans

Freeman

M. P.

Galehouse

L. L.

Madanat

Maher

Wlaschin

(2010). Research agenda for transportation infrastructure preservation and renewal: Conference report (American Society of Civil Engineers, Reston, VA). Journal of Infrastructure Systems, 16, 228-230.

24.

Turochy

R. E.

Hoel

L. A.

Doty

R. S.

(2001). Highway project cost estimating methods used in the planning stage of project development (Technical Assistance Report 02-TAR3). Charlottesville: Virginia Transportation Research Council.

25.

Wedderburn

R. W. M.

(1974). Quasi-likelihood functions, generalized linear models, and the Gauss-Newton method. Biometrika, 61, 439-447.

26.

Wilmot

C. G.

Deis

D. R.

Schneider

Coates

C. H.

Jr. (1999). In-house versus consultant design costs in state departments of transportation. Transportation Research Record, 1654, 153-160.

27.

Washington State Department of Transportation. (2002, April). Highway construction cost comparison survey. Olympia, WA: Author. Retrieved from http://ntl.bts.gov/lib/12000/12000/12082/Construction_Cost.pdf