Cost Overruns and Claims Management in Highway Construction: Lessons from International Project Management and Emerging Methodological Advances

Abstract

Avoiding highway infrastructure construction cost overruns and reducing associated claims and disputes continues to be a challenge in many countries. Research is needed in identifying notable project planning and management deficiencies that are likely to cause cost overruns. The literature suggests numerous potential causes of cost overrun but the clustering of cause variables and relative importance of clusters has not been researched. The research reported here addresses this knowledge gap using predictive models developed with data contributed by several agencies in participating countries and suggests mitigation measures. Following a review of methods and data sources, a methodological framework is advanced that encompasses statistical methods well suited for providing a scientific basis for identifying important clusters of cost overrun variables. Fifty-three completed questionnaires contributed by knowledge experts and experienced managers from Canada, the United States, the Middle East, and Australia met the sample requirements of statistical methods. Starting from 53 variables, the principal component-supported factor analysis method identified clusters of cost overrun variables and their relative importance was inferred with developed logistic regression models. Deeper insights into the causes of cost overruns obtained from this research suggest mitigation measures (e.g., improved qualification and experience of personnel, enhanced planning and design practices, risk analysis of inputs to cost estimation process) that are within reach of managers. The results can enhance infrastructure planning and management practice including a reduction in claims and disputes.

1. The Context of Project Development and Estimating Costs

The framework for transportation planning and the implementation of infrastructure elements [1], also called the logical framework approach to transportation planning [2], is the starting point for defining the context of project development and estimation of costs. Figure 1 shows three phases over which a pre-project concept progresses from project characterization to the implementation steps.

A pre-project study takes place at the system/corridor level. At this concept level, traditionally termed as “needs study,” information is gathered for a follow-up planning study. At the outset of the planning study, goals, objectives, evaluation criteria, and design standards are defined. These enable the generation of alternative solutions (including infrastructure elements), analysis in terms of cost and other factors, and evaluation using well-defined benefit–cost or effectiveness methods. Initial project characterization takes place at the planning phase to produce necessary costs and other estimates required to support the evaluation step. The evaluation results form the basis of choosing the preferred alternative that advances to the design phase [1,2].

In the design phase, details of cost estimates can be developed using agency guidelines. In the implementation phase, property acquisition takes place, permits are awarded, and construction takes place [1].

2. Problem Definition and Research Objectives

It can be inferred from the context information described above that highway construction project cost estimates are developed first at the planning phase and revised when design, site, and other details become known. Transportation departments are expected to continuously update cost estimates using the latest available information. Consideration of the various risks in cost estimation during all phases of project development and construction is encouraged. These practices result in a series of successive estimates that benefit from increased knowledge and consequently should reflect a greater level of confidence than the preceding estimate [1].

Future cost of highway project items cannot be estimated with certainty. But professional practice demands the minimization of risk in the cost estimation process. Transportation agencies have become aware of some commonly encountered cost overrun variables that are subject to risk [1,2,3]. These are tabulated in a later section of the paper.

Many variables relate to pre-construction activities (e.g., incomplete site and environmental studies, policy and planning issues, design issues, cost estimation and financial issues, qualifications of personnel) [4,5]. This paper is intended to contribute a scientific basis for the study of reducing risk in cost estimates of highway projects.

At the time of project completion, any deviation of the actual cost from the pre-construction budget (that usually includes the contingency component) is of much concern in project planning and management professional practice [3]. Most transportation authorities around the world define this difference as the “cost overrun.” In the current practice, the contingency factor is intended to go beyond previous attempts to account for the risk encountered at the planning level as well as pre-construction cost estimates [6].

Projects of all sizes can cost more than the budget that commonly includes a contingency fund [7]. In addition to the adverse effect of the cost overrun phenomenon at the construction phase, the reliability of the life cycle cost estimate that supports investment decisions can be compromised [8]. For this reason, avoiding cost overrun using planning and management actions, including risk analysis of life cycle costs, has become an important research subject [9,10].

The construction cost overrun is a complex problem to address within the broader subjects of project planning and construction management. Many causes can result in cost overrun [11]. The record of claims and disputes in many countries tracked by a commercial entity provides an initial view of some issues [12]. But many other cost overruns that become necessary for project completion are covered without dispute.

At the current state of knowledge, there is lack of research on the views of high-level management (e.g., a state/provincial department of transportation) regarding notable causes of cost overrun, including those that have resulted in claims and disputes [11,12]. These can potentially originate at the project planning and design levels and materialize while construction is in progress. These are not well-researched, especially as clusters of cost overrun variables. If such issues are identified, these can be addressed prior to construction.

While studying cost overruns in local or state/provincial highway projects within a single country is useful, expanding the investigation to a broader geographic area can provide even better insights.

Treating the uncertainties in cost overrun continues to be a research challenge. In professional practice, there is general awareness of the stochastic nature of many cost overrun causes. But there is lack of information on the identification and effect of correlated stochastic causes of cost overrun. The availability of this information can lead to guidance on how to reduce their impact.

Given the above background, the research reported here is intended to contribute knowledge on project planning and management actions needed to avoid cost overruns including those that may result in claims and disputes. Specifically, the objectives of research reported here are to:

(1) Obtain the views of experienced managers on the causes of cost overrun using an international questionnaire survey;

(2) Find clusters of variables (commonly referred to as factors) that can be analyzed instead of the many individual variables included in the questionnaire;

(3) Infer the relative importance of factors from the results of statistical methods;

(4) Identify causes of cost overrun that can be mitigated at the project planning and design stages prior to construction and while the construction is in progress;

(5) Define the role of risk analysis to reduce the impact of stochastic correlated cost overruns.

3. Study Organization

Figure 2 presents the study organization. That is, it serves as the roadmap for the contents of the paper. A brief introduction to the components of the study organization is provided here and details are covered throughout the paper. Following the problem definition, as presented in previous sections, a literature review of methods and data sources leads to research steps to address the knowledge gaps noted above. For enhancing the geographic scope of data, transportation agencies in Canada, the USA, the Middle East, and Australia were invited to share information on their cost overrun experience, including reasons for such overruns.

Given the many variables included in the survey (53), the principal component analysis-supported factor analysis method enabled the extraction of factors (i.e., clusters of variables) needed for the development of predictive logistic regression models. These probability-based models identify important causes of cost overrun.

3.1. Literature Review

3.1.1. Approaches to Curb Cost Overrun

The mission is to build or improve a highway facility according to agreed design and quality, within the specified time and budget [13]. Although the common objective is to curb highway project cost overrun, diverse approaches are reported in the literature [14]. The following review includes academic research papers and technical reports published by government/public agencies for use as guides to reduce the likelihood of cost overrun.

As an example of a technical report, the Ministry of Transportation and Infrastructure (British Columbia) guidelines include methods for cost estimation and the development of the contingency estimate [3]. The contingency fund is intended to offset uncertainties and risks. Two methods are commonly applied. One is a detailed item-by-item cost estimation method. The other is the historical bid-based approach which requires adjustments to be made to reflect current prices.

Approaches for developing contingency factor estimates to address common highway project-related risks that impact construction costs are described in reference [6]. The reviewed methods include simple deterministic formulas as well as complex probabilistic methods. Reference [9] reports the best practices for developing cost estimates as well as guidance for overseeing and checking the process of developing such estimates. Developing and applying probabilistic cost estimates is encouraged.

An approach to reducing cost overrun, termed the Reference Class Forecasting (RCF), addresses assumed reasons for cost and schedule overrun in projects, including the tendency to produce optimistic estimates and strategic misrepresentation. If applicable, these actions can logically result in the underestimation of project costs, leading to overrun. As a mitigation, this approach requires explicit, actual data-based adjustments to estimates obtained from past local projects or similar projects in other areas. In defining adjustments, unique characteristics of the project under study are considered. This method has been used by some public agencies [15] and endorsed by academic researchers [16]. Additional information on this subject is provided by [17,18]. Reference [19] adds performance of risk-based estimation to this subject.

The promotors of the RCF method suggest its application to preliminary estimates with the intent to address optimism bias and avoid strategic misrepresentation. Some applications have lowered cost overrun and improved time estimates. A limitation of this method is the guidelines to be followed by analysts, which may not always be applicable.

3.1.2. Studies on Occurrence of Cost Overrun

Information on highway infrastructure cost overrun claims and disputes in many countries, reported in Reference [12], implies the need for further research in cost overrun with the ultimate objective to improve project planning and management. Disputed transportation infrastructure cost cases for U.S., Canada, the Middle East, Australia, and many other countries show several causes of claims and disputes, including planning and design-related issues [12].

Many review papers relying only on the literature identified diverse causes of cost overrun. Reference [20] describes a systematic risk management approach to study causes of cost overrun. Reference [21] studied factors that affect contractors’ risk of cost overburden. Reference [22] describes cost overruns in large-scale projects. Reference [23] explores basic causes of cost overruns. References [24,25,26] present literature reviews on this subject. Reference [27] focuses on cost overrun research from a methodological perspective.

Several research papers describe the application of cost overrun data, obtained from published sources or from public agencies, in modelling the extent of cost overrun using probability distributions. For example, reference [28] presents probability-based methods for the investigation of risk in cost overrun estimates. But these studies could not obtain causes of cost overrun information from primary sources.

Some studies reported in the literature used questionnaires or group discussions to obtain cost overrun causes and data for developing linear multiple regression or other predictive models. An example is a Purdue University research project on the Indiana Department of Transportation (INDOT) construction projects regarding experienced cost overrun, delays, and change orders. Also, causes for such problems were identified by this study. To compare the INDOT experience with other jurisdictions, data from other states were collected and analyzed with an array of statistical methods [29]. Other examples of statistical approaches are presented by [30,31,32].

Several studies used the questionnaire approach to obtain data from a limited number of sources within a country. Following basic statistical analyses, linear regression models were developed to quantify cost overrun as a function of causal variables. This approach produces predictive models but has limitations in quantifying the importance of causal factors in probabilistic terms. Examples of this approach are reported in [33,34,35].

3.2. Need for Methodological Advances and Data Contributed by Experienced Managers

The literature review presented in Section 3.1.1 and Section 3.1.2 suggests the need for methodological advances in researching the complex highway construction cost overrun problem. Also, in order not to limit the scope of research to a part of a country or to one country, a multi-country information acquisition approach is needed. Further, due to the absence of studies based on data contributed by experienced managers who have overseen many projects, this source of knowledge is beneficial for inclusion in a research study on cost overrun.

Considering that the initial inclusion of many potentially applicable variables is necessary for a realistic scope of research, the statistical method of principal component-based factor analysis is selected for reducing many variables into a smaller number of clusters (factors) for use in further analysis without loss of information [36,37,38]. A logical next research step is to develop probability-based predictive logistic regression models using the identified factors.

As compared to multiple linear regression modelling approach, the logistic regression models address the stochastic nature of the cost overrun phenomenon. That is, these models are better suited for inferring the importance of a factor in causing overrun in probabilistic terms. For the information of interested readers, studies that are not on highway cost overrun, but have applied the logistic regression tool to factor analysis outputs, are reported in [39,40,41]. Further information on logistic regression modelling, including the rationale for its use, is provided in a later section of the paper.

3.3. International Questionnaire Survey

Tracked cost overrun claims and disputes suggest that these occur in many countries [12]. Therefore, a study of root causes should aim to obtain detailed information from international sources. In the planning phase of this research, the following options were considered for obtaining data:

• Crowd sourcing: Construction industry members (e.g., contractors) in a country or in several countries could be asked to respond to survey questions. This could potentially result in a large database, but the necessary detailed knowledge of respondents cannot be assured.
• Agents of claims and disputes: Although the transcripts provide real life information on causes, the agents are not likely to respond to questions on many potential causes of cost overrun.
• Experienced managers (e.g., executive officers in a provincial/state department of transportation) in selected countries: This option was selected for questionnaire implementation for the reasons that these managers have knowledge and experience, and they are likely to participate for knowledge generation reason [11].

A questionnaire survey was identified as the most effective way to obtain data from government transportation departments, large construction company managers, and relevant institutes. This observation is based on a review of the application of this research method in social life [42], in business [43], and in organization studies [44].

3.4. Cost Overrun Variables

A detailed literature review resulted in variables/causes of cost overrun for inclusion in the questionnaire. The variables cover applicable issues, categorized as policy, planning, regulations, design, finances, project type, construction-specific problems, qualification and experience of personnel, and many other potential causes of cost overrun (

Table 1. Cost overrun variables.

Variable Classification, Number, and Description

POLICY V1 Changes in government funding policies V2 Deal termination due to changes in law, government policy or protocols V3 Change in regulations DESIGN, CONSTRUCTION AND SCHEDULING V4 Complexity of the project (e.g., project size, project type, scope of work) V5 Design changes during construction work V6 Re-work due to the construction errors V7 Unexpected technical problem V8 Design errors that represent insufficient deliverables V9 Changes by owner on the completion date of the project V10 Scope changes by Owner during construction V11 Delays related to owner or owner representative (e.g., stop work) V12 Unrealistic project scheduling V13 Acceleration to maintain schedule V14 Delays in sending important documents to construction site (e.g., drawings, design changes) V15 Type of construction contract (e.g., unit price contract) V16 Unnecessary practices, specifications, procedures, and documentation requirements forced onto the construction site workers V17 Replacing unsatisfactory subcontractors from site by hiring new subcontractors V18 Delay by subcontractor SITE CONDITIONS/ENVIRONMENT V19 Poor site management V20 Unexpected weather conditions V21 Accidents due to poor site safety MATERIALS AND EQUIPMENT V22 Shortage of materials and equipment on site V23 Damages in materials and equipment in transit to the construction site V24 Late delivery of materials and equipment at the construction site V25 Equipment Selection Changes V26 Construction variations due to equipment selection V27 Defective materials LABOUR, STAFF AND VENDOR V28 Shortage of skilled labour V29 Lack of staff at the time of construction V30 Labour strikes and vendor strikes PERMITS AND APPROVALS V31 Delays and approval of shop drawings and installation procedures V32 Building permit to the construction contractor V33 Government/municipal approvals ESTIMATION/BUDGET AND FINANCIAL V34 Lack of expertise in setting the budget V35 The approved budget was too low V36 Absence of a detailed estimate plan V37 Changes in prices of items that have already been approved V38 Economic and financial factors V39 Inappropriate and inadequate procurement (e.g., payment terms, pricing) V40 Shortage of contingency and management reserve funds V41 Unaddressed overtime work or multiple shifts that was not included in the base estimate V42 Bankruptcy of subcontractors and vendors during construction work V43 Currency fluctuations OTHER FACTORS V44 Bad luck V45 Lack of technical qualifications of the client V46 Overly high expectations V47 Poor communication and coordination between all parties V48 Disputes between parties (designer, contractor, owner) V49 Political Factors V50 Technological risk V51 Land acquisition issues within right-of-way V52 Quality assurance and quality control V53 Inexperienced project managers, estimators and planners

). The choice of variables was guided by real-world events that caused cost overruns. These events occur during the planning, design, and construction phases of a project as described in Section 1. Risks and uncertainties do play a role, but attempts can be made by technical and managerial personnel to lessen their impact with updated information.

The respondent was requested to rate the causes of cost overrun on a five-point Likert scale (i.e., very high, high, medium, low, and very low). This aspect of the data acquisition design was necessary for statistical analyses and predictive model development.

All aspects of the questionnaire study design were examined by the University Ethics Review Board, including the plan for quantitative studies. Following approval, about 150 questionnaires were forwarded to departments of transportation in Canadian provinces and USA states, professional infrastructure agencies/managers in the Middle East, and institutes in Australia. The responding organizations were requested to provide responses completed by high-level experienced managers. Fifty-three completed questionnaires contributed by Canadian, United States, Middle East, and Australian experienced managers and knowledge experts met the sample requirements of statistical methods.

The survey responses (data sets) were used for factor extraction with the use of the Statistical Package for Social Sciences (SPSS) Version 25. The outputs of analysis enabled the development of logistic regression models. The mathematical functions used in the SPSS are not reproduced in this paper. But to explain the theoretical basis of the predictive logistic modelling approach, a limited number of equations are presented.

3.5. Identification of Cost Overrun Factors

Although the Likert scale-based results of the questionnaire study described in previous sections of the paper provided useful general information, it was decided to apply scientific methods to obtain details inherent in the database. Also, the study of stochastic characteristics of the cost overrun phenomenon necessitated the application of advanced statistical methods.

Statistical analysis of many variables requires the use of the principal component method-supported factor analysis to identify a few factors that retain the properties of the constituent larger number of variables. References [37,38] describe data analysis techniques and Reference [39] provides an example application. Well-defined methodological steps shown in Figure 3 were followed for factor analysis. The rationale was that if successfully formed, these factors will enable the study of the cost overrun phenomenon captured by many variables included in the questionnaire without loss of inherent information.

3.6. Suitability of Survey Data

To check the statistical suitability of data obtained with the use of Likert scale, the Cronbach’s Alpha (α) test of reliability was applied for this purpose [45]. Although Alpha can range from zero to one, for robust analysis a high value is necessary. For variables rated by respondents, the Cronbach’s Alpha was 0.98, which implies that the survey instrument (i.e., Likert scale) and resulting data were reliable.

3.7. Filtering Data and Adequacy Tests

Following the development of the correlation matrix, data were filtered to remove variables that showed correlation coefficient lower than 0.3 (considered to be a weak correlation) [46]. The objective was to include variables in factor analysis that correlate well with other variables. This strategy resulted in 31 variables. According to the histogram of correlation coefficients shown in Figure 4, 60% are above 0.6 and 83.7 percent exceed 0.5.

Before conducting factor analysis, the Kaiser–Meyer–Olkin (KMO) measure of sample adequacy is used to check if the KMO > 0.5 requirement based on Kaiser’s criterion is met [47].

Table 2. KMO and Bartlett’s Tests.

Test	Values
Kaiser–Meyer–Olkin Measure of Sampling Adequacy	0.816
■ Bartlett’s Test of Sphericity Approximate Chi-Square Degree of Freedom (df) Significance (sig.)	2189.718 465 0.00

shows that the KMO was 0.816, which is acceptable. Also, the Bartlett Test of Sphericity was performed to measure the strength of the relationship between variables [48]. The value of the Bartlett Test of Sphericity was 2189.718 and the associated significance level was 0.000 (p < 0.05). Therefore, the data of the 31 variables were suitable for factor analysis.

3.8. Factor Extraction

Although the use of the principal component analysis (PCA) to analyze 31 variables can produce many principal components (factors) (at most 31), the objective of the factor analysis is to extract a smaller number of factors that can explain a very high percentage of variance [49]. The PCA transforms the correlated variables into a set of linearly uncorrelated principal components (factors) [50]. The choice of factors to retain for further analysis depends on the variance explained. That is, the decision to retain the number of components is based on the cumulative percentage of variance explained.

There are two other considerations that can potentially guide the choice of factors. Kaiser’s criterion [47] suggests that factors with eigenvalues greater than 1 can be retained. Also, a scree plot of the eigenvalues of factors, presented later in this section, can help visualize the role of each factor. At the cutoff of eigenvalue 1.0, the number of factors can be clearly seen for further analysis.

Based on the above noted theory, eigenvalues and variances are the basis for factor extraction.

Table 3. Factor extraction using initial eigenvalues (IEV) and rotation results.

Factor	Initial Eigenvalue (IEV)	IEV % of Variance	IEV Cumulative %	Following Rotation % of Variance	Following Rotation Cumulative %
1	19.70	63.55	63.55	18.64	18.64
2	1.88	6.06	69.61	18.02	36.66
3	1.36	4.38	73.99	15.10	51.76
4	1.13	3.65	77.64	14.77	66.53
5	1.06	3.40	81.04 *	14.51	81.04 *
		81.04		81.04 *

NOTE: * 81.04% of the total variance is accounted for by the first 5 factors.

shows eigenvalue and variance results that guided the factor extraction decision. The factors are arranged in descending order, based on “explained variance”. The first five factors have eigenvalues greater than 1. These five factors account for 81.04% of total variance.

The next step in analysis was the application of the varimax rotation method to adjust the variance of each of the extracted factors. The adjustment process redistributed 81.04% of total variance over the extracted five factors. The purpose of the rotation process is to obtain a better balance of variance for the extracted factors for improved factor loads [36,39,51]. By rotating the factors, each factor can have a set of highly loaded variables. The results of the varimax rotation shown in the last two columns of Table 3 suggest that, as compared to column 3 values based on initial eigenvalues, the percent variance explained by factors in column 5 are better balanced.

To explain the rationale for rotation and its effect further, the PCA-based factor analysis did not produce easily interpretable factors. This can be seen in information for before rotation presented in Table 3. The rotation step addressed this issue by mathematically transforming the factor loadings. By rotating the factors, each variable ideally loads highly on one factor while exhibiting minimal loadings on others. This reduces the complexity of understanding which variables belong to which underlying latent factor.

Several observations can be drawn from contents of Table 3. In the initial eigenvalue column of Table 3, the first factor with an eigenvalue of 19.70 accounted for 63.55% of the total variance for the data. The second factor accounted for only 6.06% of the total variance. Similar comments apply to Factors 3 to 5. The results also show that 81.04% of the total variance was attributable to the first five factors. But after rotation, the percentage of variance explained by the first factor was 18.64 and the second factor accounted for 18.02 percent. Similar improved loading of variables was found for Factors 3 to 5 (Table 3).

A scree plot (Figure 5) was developed to illustrate the number of extracted factors. It is a graphical representation of the eigenvalue corresponding to factor numbers. As noted above, the eigenvalue of 1.0 defines the cutoff for factor retention. This illustration confirms that the first five factors should be sufficient for this research study. Therefore, these were included in the factor model, and their factor loadings were calculated.

As noted later, for improved interpretation, Factors 2, 3 and 4 were combined. In accordance with the theory, the extracted factors have loadings on the cost overrun variable greater than 0.5.

3.9. Interpretation of Factors

From the rotated factor matrix, the Factor 1 cluster consists of eight variables (

Table 4. Variables included in Factor 1.

Variable	Original Variable Descriptions	Factor Loading
V3	Change in regulations	0.776
V17	Replacing unsatisfactory subcontractors from site by hiring new subcontractors	0.740
V38	Economic and financial factors	0.735
V1	Changes in government funding policies	0.651
V16	Unnecessary practices, specifications and procedures	0.634
V42	Bankruptcy of subcontractors and vendors during construction work	0.579
V14	Delays in sending important documents to construction site (e.g., drawings, design changes)	0.564
V40	Shortage of contingency and management reserve funds	0.546
	Variance explained 18.64%

). Figure 6 illustrates the classification of these variables defined in Table 1. As previously noted, Factor 1 accounts for 18.64% of the total variance. The relative impact of this factor on the cost overrun is examined later in the logistic regression model.

Table 5. Factors 2 to 4 cluster.

Factors and Variables	Original Variable Descriptions	Factor Loading	Variance Explained
Factor 2			18.02%
V4	Complexity of the project (e.g., project size, project type, scope of work)	0.837
V15	Type of construction contract (e.g., unit price contract)	0.785
V53	Inexperienced project managers, estimators and planners	0.748
V5	Design changes during construction work	0.683
V8	Design errors that represent insufficient deliverables	0.656
V9	Changes by owner on the completion date of the project	0.547
V13	Acceleration to maintain schedule	0.521
V31	Delays and approval of shop drawings and installation procedures	0.514
Factor 3			15.10%
V21	Accidents due to poor site safety	0.799
V52	Quality assurance and quality control	0.738
V6	Re-work due to the construction errors	0.515
V19	Poor site management	0.507
Factor 4			14.77%
V36	Absence of a detailed estimate plan	0.688
V46	Overly high expectations	0.659
V32	Building Permit to the construction contractor	0.603
V18	Delay by subcontractor	0.577
V39	Inappropriate and inadequate procurement (e.g., payment terms, pricing)	0.572
V2	Deal termination due to changes in law, government policy or protocols	0.570
V34	Lack of expertise in setting the budget	0.555
	Cumulative variance explained		47.89%

presents information on Factors 2–4. The percent variance accounted for by Factors 2–4 are 18.02%, 15.10%, and 14.77%, respectively. The total variance explained by a combination of these factors is 47.89%. These are combined into one factor for further analysis because their constituent variables cluster well as shown in Figure 7. Most variables are categorized into design, construction, scheduling, estimation/budget, financial, and experience issues. Some of these causes of cost overrun, namely those that relate with deficient planning and design, have been reported in transcripts of claims and disputes [12].

As noted in

Table 6. Variables represented by Factor 5.

Variable	Original Variable Descriptions	Factor Loading
V41	Unaddressed overtime work or multiple shifts that was not included in the base estimate	0.700
V26	Construction variations due to equipment selection	0.666
V28	Shortage of skilled labour	0.623
V24	Late delivery of materials and equipment at the construction site	0.599
	Variance explained 14.51%

, Factor 5 consisting of four variables, explains 14.51% of the total variance. Figure 8 illustrates these categories of cost overrun causes.

A summary of the factor analysis results presented in

Table 7. Summary of interpreted factors.

Factor	% of Variance	Comment on Constituent Variables
Factor 1 (8 variables)	18.64	Most issues belong to estimation/budget, finances, design (delays in sending drawings to site, unnecessary practices). Policy and regulatory issues are also noted. See Table 4 and Figure 6. Table 1 shows variable classification and description.
Combined Factor (based on Factor 2 + Factor 3 + Factor 4) (19 variables)	47.89	Most variables are classified as issues with planning, design, construction, scheduling, estimation/budget, finances, inexperience, quality, expectations, permits and approvals, site management, approvals. See Table 5 and Figure 7. Table 1 shows variable classification and description.
Factor 5 (4 variables)	14.51	Most variables relate to issues with materials and equipment. Also, there are variables on estimation/budget, financial, shortage of skilled labour. See Table 6 and Figure 8. Table 1 shows variable classification and description.

suggests that as expected, diverse causes of cost overruns are identified by the statistical method. However, a few themes are becoming noticeable. These mainly relate to planning, design, qualifications, experience, and site management issues. As noted previously, at this stage of analysis, the relative importance (i.e., impact) of these factors on cost overrun could not be inferred. The results of logistic predictive models described next provided an answer to this question.

3.10. Probability-Based Logistic Regression Modelling

3.10.1. Methodological Components

The probability-based predictive logistic regression model development and interpretation of results are the final components of the overall methodology for research reported in this paper. Figure 9 presents inputs, logistic model structure, and outputs. Details are presented in this section and Section 3.10.2.

The model is structured as follows:

$$\mathrm{L}\mathrm{o}\mathrm{g}\left(P\right)=B_0+B_1{X}_1+B_2{X}_2+\cdots +B_i{X}_i$$

(1)

where

P = is the probability of cost overrun occurring (termed the dependent variable);

$B_0$ = model constant;

$B_1, B_2, \dots B_i$ = coefficients;

$X_1, X_2, \dots X_i$ = explanatory factors—the cluster of variables obtained from factor analysis.

In logistic regression modelling, the maximum likelihood method is used to estimate the probability of cost overrun P, model constant B₀, and coefficients (B₁, B₂, …) for the explanatory variables (X₁, X₂,…). The odds ratio is calculated to infer the strength of the effect of explanatory factors (cluster of variables) on P in terms of the likelihood of the occurrence of cost overrun.

The logistic regression model dependent variable P is coded in terms of the odds of an event occurring. The methodology uses the ratio of the probability that it will occur to the probability that it will not occur. If the odds ratio is greater than one (i.e., odds > 1), it indicates that an explanatory factor (cluster of variables) is likely to cause a cost overrun. But a value of less than one (i.e., odds < 1) indicates no effect.

When modelling a binary value, it is necessary to measure the fit of the model to the observed data [52]. The Pseudo R-Square measures the proportion of explained variance in the regression model. There are two types of R-Square in logistic regression analysis. These are Nagelkerke and Cox & Snell [52,53,54,55].

Smith and McKenna [54] noted other commonly used Pseudo R-Square indices (i.e., McFadden’s index, Cox–Snell index with or without Nagelkerke). The Nagelkerke R-Square is the most used statistic when interpreting the logistic regression model and the Cox and Snell R-Square only provides an approximate value of variance. According to Campbell [56], both the Nagelkerke and Cox and Snell R-Square values can be interpreted similarly. Therefore, the higher the R-Square value, the better the model fit.

The Cox and Snell’s R-Square is calculated using the following equation:

$$R_{\mathrm{C}\mathrm{S}}^2=1-\mathit{exp}\left({\displaystyle \frac{{-2LL}_{\mathrm{n}\mathrm{e}\mathrm{w}}-{(-2LL}_{\mathrm{I}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}})}{N}}\right)$$

(2)

where

N: sample size;

LL: log-likelihood.

The Nagelkerke R-Square is calculated as follows:

$$R_N^2={\displaystyle \frac{R_{\mathrm{C}\mathrm{S}}^2}{1-\mathit{exp}\left(-\frac{{-2LL}_{\mathrm{I}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}}{N}\right)}}$$

(3)

The log-likelihood (LL) can be used to assess the significance of the logistic regression model. The smaller the value of the log-likelihood, the better the fit of the model. A large value of the log-likelihood indicates poorly fitting models. Sometimes the log-likelihood is referred to as (−2LL).

Chi-square, another goodness of fit test, measures the significance of the model. In the SPSS software, the chi-square model is labelled as the “Omnibus Test of model Coefficients”. The chi-square measure is sometimes called the traditional fit measure. A significance value of less than (0.05) indicates that the model fits well.

3.10.2. Logistic Regression Model Results

Logistic regression models were developed using factors that resulted from the principal component analysis-supported factor analysis method. The process of calibrating the logistic regression models enabled adding factors one by one as explanatory factors. This step resulted in the estimate of their respective unique contribution to cost overrun. Statistics were compiled on odds ratio, log-likelihood, Pseudo R-Square, and chi-square values.

Results based on Factor 1, shown in

Table 8. Logistic regression model statistics for Factor 1.

Model Statistics	Result
Factor 1 odds ratio	0.895
Model Fit Information
−2 log-likelihood (−2LL)	72.391
Model chi-square	0.155
Sig.	0.693
Pseudo R-Square
Cox & Snell R-Square	0.003
Nagelkerke R-Square	0.004

, indicate an odds ratio of less than one. This result implies that the cost overrun is not occurring due to the inclusion of this factor in the regression model. Other results (i.e., the log-likelihood of 72.391, a very small chi-square of 0.155, non-significant at 0.693 (p > 0.05)) indicate that the model does not fit well. Therefore, it can be inferred that Factor 1 does not seem to have any significant effect on cost overrun.

An explanation is noted here on why Factor 1, which accounted for 18.64% variance in the factor analysis study did not produce an odds ratio of one or higher in the predictive logistic regression model. Although within factor analysis, the percent variance accounted by a factor is a useful indicator of the relationship of its constituent variables with cost overrun, the 18.64% explained variance was not sufficient to produce a significant effect on the outcome of predicting a cost overrun. Also, according to theory, the factor analysis and the logistic regression model need not measure the same phenomenon. The magnitude of variance explained by a factor is attributed to data, the odds ratio describes the direction and magnitude of a factor’s effect on the outcome (i.e., in this analysis, the ratio of the probability that cost overrun will occur to the probability that it will not occur).

We proceed with the next model by adding another factor. The results of adding the Combined Factor (i.e., combination of Factors 2–4) to the regression model are shown in

Table 9. Logistic regression model statistics for Factor 1 plus the Combined Factor.

Model Statistics	Result
Factor 1 odds ratio	0.848
Combined Factor (combination of Factors 2–4) odds ratio	13.626
Model Fit Information
−2 log-likelihood (−2LL)	53.664
Model chi-square	18.882
Sig.	0.000
Pseudo R-Square
Cox & Snell R-Square	0.300
Nagelkerke R-Square	0.402

. The results show that the Combined Factor has an odds ratio greater than one, indicating that the addition of the Combined Factor into the model increased the probability of the cost overrun occurring. This model is significant at 0.000 (p < 0.05). The log-likelihood decreased from 72.391 to 53.664, and the chi-square increased from 0.155 to 18.882, indicating that the model fitted well and was a significant improvement over the previous model (based only on Factor 1).

Moreover, the model accounts for between 30% (Cox and Snell R-Square) and 40.2% (Nagelkerke R-Square) of the variance, indicating a moderate to a good association between the explanatory factors and the dependent variable.

The next step in logistic regression modelling was to add Factor 5. The results are presented in

Table 10. Logistic regression model statistics for Factor 1, the Combined Factor, and Factor 5.

Model Statistics	Result
Odds Ratio
Factor 1	0.854
Combined Factor (combination of Factors 2–4)	16.305
Factor 5	0.715
Model Fit Information
−2 log-likelihood (−2LL)	52.674
Model chi-square	19.872
Significance.	0.000
Pseudo R-Square
Cox & Snell R-Square	0.313
Nagelkerke R-Square	0.419

. The overall test of this model was significant at p < 0.05, indicating a good fit of the model. However, the odds of Factor 1 and Factor 5 show no effect on cost overrun and the odds of the Combined Factor continue to show an impact on the cost overrun variable.

The −2LL ratio decreased from 53.664 to 52.674, and the chi-square increased from 18.882 to 19.872. The Cox and Snell R-Squared was 31.3% and the Nagelkerke R-Squared of this model was 41.9%. In general, the model fit metrics and the R-Square metrics are like the previous model (shown in Table 9). Therefore, this model shows a weak improvement.

4. Discussion: Key Findings, Recommended Interventions

The principal component analysis-supported factor analysis method in association with the predictive logistic model produced findings of interest to professional staff as well as researchers. From the results of logistic regression models noted in Table 10, specially the odds ratio, the following relative position of factors in contributing to cost overrun is inferred: (1) the Combined Factor (combination of Factors 2, 3, and 4); (2) Factor 1; and (3) Factor 5. Figure 7 illustrates the classification of variables included in the Combined Factor and Table 7 provides summary comments on constituent variables.

It is beneficial for the reader to know that there is an absence of published studies based on data contributed by experienced international managers who have overseen many projects. Therefore, the findings of this research cannot be compared with academic published papers. However, the results of the Combined Factor are in general consistent with claims and disputes transcripts regarding road and highway projects. This consistency can be viewed as external validity of research results obtained with the use of international questionnaire study. Also, it is an indication of possible professional cultural overlap between participating countries.

A summary of frequently reported claims and disputes suggests the following notable issues: design was incomplete, design was incorrect, changes in scope, physical conditions were unforeseen, workman deficiencies, design information was issued late, and contract management and/or administration failure [12]. But the “design information was issued late” was not included in the highly rated Combined Factor. Instead, it is a part of Factor 1 which received an odds ratio less than 1.

While viewing results, the reader is advised that this research paper cannot identify specific geographic and jurisdictional locations of international contributors of data due to the participation agreement. The rationale for this aspect of the survey design was for the benefit of a global technical audience interested in knowing the international experience in causes of cost overrun.

Recommended interventions based on results are noted next.

• Several identified causes of cost overrun can be avoided and mitigated at the project planning and design stage prior to construction. Also, project planners and designers can assist in mitigating cost overrun causes that occur while the construction is in progress.
• The stochastic characteristics of many cost overrun causes that are noticeable throughout this paper call for risk analysis to reduce the impact of uncertainties.
• The results of factor analysis and the probability-based logistic regression model contribute to the risk minimization objective. Additional information on risk analysis methods can be sourced from [3,7,8].
• This research calls for improving the entire process of planning, design, risk analysis, and implementation of highway infrastructure projects to avoid and mitigate cost overruns.

High-level managers can avoid cost overrun claims and disputes by implementing the above recommended interventions. These can be incorporated in guidelines on project planning, design, cost estimation, and construction management. Professional staff qualifications can be improved, and they can be trained to implement these recommendations on a routine basis. Enhanced project planning and design practices including risk analysis of inputs to the cost estimation process is within reach of high-level managers.

5. Conclusions

The methodological framework, the constituent methods, and data contributed by international participants worked well in achieving the research objectives. The literature supplemented by the claims and disputes information are suitable for the identification of potential causes of cost overrun. The selected statistical methods (i.e., principal component analysis-based factor analysis and probability-based logistic regression model) meet the requirements for data analysis and predictive model development. Their output can lead to improved project planning and management practices that are likely to avoid cost overruns.

Data contributed by experienced managers in selected countries reflect their detailed knowledge of conditions that may cause cost overrun. Although there are challenges in obtaining sensitive data on cost overrun experience in a country, the comprehensive questionnaire survey produced the number of responses required for statistical analyses and model calibration.

The factor analysis approach supported by the principal component analysis method enabled the extraction of well-defined factors from many variables for interpretation without loss of information. These extracted factors resulted in a statistically acceptable predictive probability-based logistic regression model used for assessing the effect of identified factors in causing cost overrun.

Based on their experience and the context of their practice, the international respondents have contributed information that resulted in defining measures for reducing the occurrence of cost overrun and avoiding claims and disputes. A high proportion of causes of cost overrun identified in this research relate to pre-construction activities and those that can be avoided while the construction is taking place. Most of these can be addressed by project planners and construction managers. Also, with knowledge of risk management methods, adverse effects can be minimized.

Future studies could attempt to increase the participation of additional countries to enhance the coverage of international project management experience. Also, future studies can apply the methodological advances reported in this paper to other transportation infrastructure (e.g., public transit, airports, marine ports).