158 CHAPTER 4 DATA ANALYSIS AND FINDINGS 4.1 Introduction The five sections in the body of this chapter presents details and elaborations of the findings. The first section presents a descriptive analysis on the demographic information of the survey respondents and the overall response rate. The second section presents an assessment of the reflective measurement model in the Structural Equation Modeling (SEM), elaborating on the data screening process which includes identification of missing data, as well as tests for normality, outliers, linearity, homoscedasticity and multi-collinearity. The third section presents the results of the confirmatory factor analysis which include i) pooled confirmatory factor analysis (CFA) or the goodness of fit of the measurement model in terms of unidimensionality, validity, and reliability and ii) structural model for hypotheses testing. The fourth section discusses the multi group analysis to test the moderating effect of trust on the variables, and the chapter ends with a discussion on the final structural model for the study and the conclusion. 4.2 Descriptive Analysis This section presents a descriptive analysis of the respondents’ demographic information as well as discusses the overall response rate. The study’s respondents are 352 employees from 26 toll concessionaire companies in Malaysia. 159 4.2.1 Demographic Information Table 4.1 shows the distribution of the respondents who participated in this study explained using two demographic features – level of management and working experience (number of years). In this regard, 162 respondents (46.1%) were from the top management, while 190 respondents (53.9%) were from middle management employees. The majority of respondents (63.9%) had over 11 years working experience, 20.5% had 6 to 10 years working experience, and the remaining 15.6% who had less than 5 years. Table 4.1: Distribution of Respondents according to Demographics (n=352) Frequency Percentage (%) Management Level Top Management 162 46.1 Middle Management 190 53.9 Working Experience More than 11 years Between 6 – 10 years Less than 5 years 225 72 55 63.9 20.5 15.6 4.2.2 Response Rate In compliance with data collection requirements, 400 questionnaires were administered to 400 employees as described in 4.2.1 above. From the 400, 352 questionnaires were returned and completed. The response rate was thus 88%, which matched the required response rate for a survey exercise. Table 4.2: Survey Response Rate Survey Administered Survey Returned Complete Match 400 352 352 (100%) (88%) 160 4.3 Assessment of the Reflective Measurement Model in Structural Equation Modeling (SEM) An assessment of the reflective measurement model in SEM was conducted using AMOS Version 23.0. The process started with data screening, including identification of missing data, followed by testing of statistical assumptions such as multivariate normality, detecting outliers, linearity, homoscedasticity, and multi- collinearity between constructs. Hair et al., (2010) quoted that it is vital to test statistical assumptions to avoid violations that may jeopardise the validity of the results. 4.3.1 Data Screening Various methods were used for data screening including detecting missing values and outliers, conducting normality, linearity and homoscedasticity tests, and checking for multi-collinearity. 4.3.2 Missing Data Sekaran and Bougie (2010) states that missing data in any research is a serious issue as it can significantly affect the survey’s results. In this study, the survey was conducted online, and there were no missing values among the 352 responses. 4.3.3 Multivariate normality A normality assessment is conducted evaluate the skewness for every variable, and was undertaken using the skewness and kurtosis. A skewness value of 1.0 or lower means the data was normally distributed. The AMOS SEM using the Maximum 161 Likelihood Estimation (MLE) is a relatively vigorous test to identify skewness more than 1.0 in absolute value, especially in cases with large sample size. Meanwhile, the Critical Region (CR) for the skewness does not exceed 8.0. In MLE, a sample size greater than 200 is considered large even though the data distribution is slightly abnormal. Therefore, with a sample size greater than 200, the researcher was able to proceed with the analysis with the absolute skewness up to 1.5 (Zainudin, 2015). For this study, the skewness range is -1.164 to -0.499, which is below than the stipulated +-1.5 (Hair et al., 2010), as seen in Table 4.3. Therefore, normality for this data is achieved. Table 4.3: Assessment of Normality VARIABLE MIN MAX SKEW C.R. KURTOSIS C.R. INV1 1.000 5.000 -.499 -3.820 .675 2.583 INV2 1.000 5.000 -.586 -4.490 .590 2.258 INV3 1.000 5.000 -.676 -5.180 1.214 4.648 INV4 1.000 5.000 -.653 -5.002 .824 3.154 INV5 1.000 5.000 -.879 -6.729 1.738 6.655 INV6 1.000 5.000 -.651 -4.989 1.276 4.886 BI1 1.000 5.000 -.619 -4.743 .943 3.613 BI2 1.000 5.000 -.656 -5.025 1.490 5.706 BI3 1.000 5.000 -.556 -4.257 .463 1.773 BI4 1.000 5.000 -.629 -4.819 .606 2.320 BI5 1.000 5.000 -.719 -5.508 .985 3.773 PV1 1.000 5.000 -.717 -5.492 .516 1.977 PV2 1.000 5.000 -.895 -6.853 1.890 7.238 PV3 1.000 5.000 -.662 -5.069 .998 3.820 PV4 1.000 5.000 -.815 -6.245 1.483 5.680 162 VARIABLE MIN MAX SKEW C.R. KURTOSIS C.R. PV5 1.000 5.000 -.774 -5.928 1.456 5.577 PV6 1.000 5.000 -.660 -5.054 1.211 4.640 PV7 1.000 5.000 -.798 -6.110 1.120 4.287 PV8 1.000 5.000 -.742 -5.682 .360 1.378 GI6 1.000 5.000 -.865 -6.625 1.214 4.649 GI7 1.000 5.000 -.653 -4.998 .229 .879 GI8 1.000 5.000 -.628 -4.808 .488 1.870 GI9 1.000 5.000 -.832 -6.372 .744 2.849 GI0 1.000 5.000 -.804 -6.157 1.087 4.162 GI11 1.000 5.000 -.888 -6.804 .993 3.804 ATT1 1.000 5.000 -1.164 -8.916 2.657 10.175 ATT2 1.000 5.000 -.533 -4.085 1.804 6.910 ATT3 1.000 5.000 -.796 -6.097 1.506 5.766 ATT4 1.000 5.000 -1.150 -8.810 2.943 11.271 ATT5 1.000 5.000 -.997 -7.638 2.398 9.183 ATT6 1.000 5.000 -.991 -7.592 2.281 8.735 ATT7 1.000 5.000 -1.084 -8.306 2.338 8.954 SN1 1.000 5.000 -.764 -5.849 1.109 4.247 SN2 1.000 5.000 -.521 -3.993 1.014 3.881 SN3 1.000 5.000 -.530 -4.061 .950 3.640 SN4 1.000 5.000 -.708 -5.420 1.230 4.710 SN5 1.000 5.000 -.664 -5.084 1.015 3.886 PBC1 1.000 5.000 -.754 -5.774 .724 2.773 PBC2 1.000 5.000 -.623 -4.772 .533 2.042 PBC3 1.000 5.000 -.881 -6.744 1.222 4.680 PBC4 1.000 5.000 -.626 -4.794 .629 2.410 PBC5 1.000 5.000 -.625 -4.791 .408 1.563 PBC6 1.000 5.000 -.638 -4.884 .555 2.127 PBC7 1.000 5.000 -.618 -4.735 .389 1.488 Multivariate 790.320 116.526 163 4.3.4 Outliers Byrne (2010) refers to outliers as any observation that is numerically different in comparison with the overall dataset. Schumacher & Lomax, (1996) state that outliers have the ability to affect the parameter estimates. Table 4.4 shows 43 cases (where squared Mahalanobis distance values exceed the critical chi-square value, which in this case is 73.402 (refer Appendix 1), and are thus considered outliers. Table 4.4: Outliers Observation number Mahalanobis d-squared p1 p2 335 195.263 .000 .000 261 184.676 .000 .000 248 178.504 .000 .000 97 175.251 .000 .000 247 158.798 .000 .000 141 132.102 .000 .000 234 130.662 .000 .000 81 119.898 .000 .000 228 118.890 .000 .000 3 118.422 .000 .000 242 116.703 .000 .000 292 112.484 .000 .000 260 109.863 .000 .000 294 104.730 .000 .000 274 100.369 .000 .000 164 Observation number Mahalanobis d-squared p1 p2 100 98.458 .000 .000 101 98.458 .000 .000 332 96.188 .000 .000 282 93.797 .000 .000 352 93.797 .000 .000 166 93.196 .000 .000 21 92.236 .000 .000 305 91.856 .000 .000 6 91.746 .000 .000 227 91.095 .000 .000 263 88.344 .000 .000 90 87.056 .000 .000 146 86.591 .000 .000 236 85.720 .000 .000 201 85.383 .000 .000 14 84.864 .000 .000 91 84.730 .000 .000 229 83.970 .000 .000 73 82.534 .000 .000 74 82.261 .000 .000 185 80.855 .001 .000 240 80.666 .001 .000 259 80.557 .001 .000 165 Observation number Mahalanobis d-squared p1 p2 223 79.380 .001 .000 235 78.577 .001 .000 23 76.710 .002 .000 43 76.697 .002 .000 103 76.681 .002 .000 94 72.944 .004 .000 217 72.593 .004 .000 218 71.748 .005 .000 According to Pallant (2011), prior to coming to a decision to maintain or delete the value, the researcher is required to look into Cook’s Distance value to identify whether those values (cases) have any unjustified influence on the results. Tabachnick and Fidell, (2007), however, added the cases with Cook’s Distance value of larger than 1 are those found to have potential problems. In the case of this study, the Cook’s Distance value, seen in Table 4.5, shows a maximum value is 0.822 (less than 1). Based on this, all 43 cases were retained (Pallant, 2011). Table 4.5: Cook’s Distance Value Min. Max. Mean Standard Deviation N Cook's Distance .000 .822 .006 .045 352 4.3.5 Linearity and Homoscedasticity The assessment of linearity was conducted through a residual analysis that resulted from the regression analysis. The homoscedasticity test was conducted using 166 a scatter plot diagram of standardised residuals, and results showed the existence of homoscedasticity in the set of independent variables, and the variance of the dependent variable. Pallant (2011) assumptions of linearity and homoscedasticity can be identified by examining a scatter plot of standardised residuals. Figure 4.1 shows the scatter plot for linearity and homoscedasticity scores are concentrated in the centre (along with the 0 points). Therefore, the linearity and homoscedasticity are achieved for this data. Figure 4.1: Scatter plot of the Standardised Residuals 4.3.6 Multi-collinearity Multi-collinearity can be defined as the component variables being close equal in linearity (Zainuddin, 2015). Multi-collinearity is basically when two or more variables are not independent and is seen as simply a matter of identifying the degree of closeness, to allow proper treatment to be taken in the study. When variables are 167 used as predictors, a sufficiently high degree of interdependence can render the model results inadequate and misleading. As such, the Structural Equation Modeling (SEM) is a powerful method for managing multi-collinearity in sets of predictor variables. The importance of multi- collinearity is to define the correlation value between variables where a value not exceeding 0.85 shows that the variable is free from the redundant items (Zainudin, 2015). Table 4.6: Correlation Matrix PBC SN ATT GI PV BI INV Perceived Behavioural Control (PBC) - Subjective Norm (SN) .697 - Attitude (ATT) .566 .623 - Governmental Influence (GI) .543 .559 .486 - Project Viability (PV) .611 .615 .601 .656 - Behavioural Intention (BI) .635 .590 .542 .607 .586 - Involvement (INV) .592 .609 .595 .567 .682 .588 - Multi-collinearity occurs when the correlation matrix between any two variables show an extremely high value, i.e. a value of above 0.85 for any variable is deemed as a redundant variable (Zainudin, 2015). Table 4.6 shows the value for perceived behavioural control, subjective norm, attitude, governmental influence, project viability, behavioural intention and involvement variables are between 0.486 to 0.697, where the correlation among all the variable is below 0.85 and thus indicating that there are no redundant items between variables. 168 4.4 Hypothesis Testing This section will be present on the results and analysis of the hypotheses tested for this study, which predominantly uses the SEM methods. In this study, four stages of analysis were used: i. Measurement Model; ii. Structural Model; and iii. Multi Group Analysis, which is used to determine the moderation effects of the trust between behavioural intention and involvement behaviour. 4.4.1 Measurement Model This section explains the process of analysing data using the Confirmatory Factor Analysis (CFA) method. This includes the processes of assessing measurement validity, uni-dimensionality and reliability. Prior to modelling the structural model and executing SEM, the study is first required to validate all latent constructs involved in the model (Zainudin, 2015). This validation procedure is known as CFA, which can be executed using two methods – the single construct CFA and the Pooled-CFA for all constructs. This study utilises the Pooled-CFA method as it is seen to be more suitable, efficient and thorough, as well as able to avoid any model identification problems, as indicated by Zainudin (2015). Under Pooled-CFA, all the constructs are pooled together and linked using the double-headed arrows to assess the correlation among them. In this case, all the constructs were pooled and assessed simultaneously. 169 The seven variables of this study – perceived behavioural control, subjective norm, attitude, governmental influence, project viability, behavioural intention and involvement – which can be further broken down into 44 items is shown in Figure 4.2. Figure 4.2: Pooled-CFA Three assessments for (i) unidimensionality; (ii) validity; and (iii) reliability for measurement models, were required before modelling the structural model. These assessments are explained below: (i) Unidimensionality The requirement of fitness indexes for this study was achieved through item- removal of low factor loading items identified through pooled-CFA. Eight items were deleted – three from governmental influence (G6, G7, G11); two from subjective norm 170 (SN1, SN4); and three under project viability (PV1, PV7, PV8), after which, the model run and the fitness indexes achieved the required levels for the study to continue (Figure 4.3). The fitness indexes were met based on requirements specified by Hair, et al. (2010). There was no issue of Model Identification in this study event with the removal of the eight constructs since the combined constructs would have increased the degrees of freedom for the model (Zainudin, 2015). On a similar note, all items showed positive factor loading values. Figure 4.3: The Pooled-CFA after Removal of Low Loading Items 171 (ii) Validity For validity assessment, three processes were conducted, namely: a. convergent validity; b. construct validity; and c. discriminant validity. (a) Convergent Validity - The assessments for Convergent Validity was made based on the values of Average Variance Extracted (AVE) and Composite Reliability (CR) as shown in Table 4.7. The study computed these values for every construct, with a minimum threshold value for AVE is 0.5 or higher, and CR is 0.6 identified for validity to be achieved (Hair et al., 2010). Table 4.7: The AVE and CR Values for all Constructs Construct Item Factor Loading (above 0.50) CR (above 0.6) AVE (above 0.5) Perceived Behavioural Control PBC1 .74 0.902 0.569 PBC2 .81 PBC3 .79 PBC4 .71 PBC5 .68 PBC6 .75 PBC7 .79 Subjective Norm SN1 Removed 0.754 0.506 SN2 .76 SN3 .70 SN4 Removed SN5 .67 Attitude ATT1 .66 0.892 0.543 ATT2 .72 172 Construct Item Factor Loading (above 0.50) CR (above 0.6) AVE (above 0.5) ATT3 .77 ATT4 .75 ATT5 .78 ATT6 .76 ATT7 .71 Governmental Influence GI11 Removed 0.778 0.538 GI10 .76 GI9 .70 GI8 .74 GI7 Removed GI6 Removed Project Viability PV1 Removed 0.836 0.506 PV2 .68 PV3 .69 PV4 .71 PV5 .80 PV6 .67 PV7 Removed PV8 Removed Behavioural Intention BI1 .64 0.840 0.514 BI2 .77 BI3 .75 BI4 .64 BI5 .77 Involvement INV1 .66 0.877 0.545 INV2 .72 INV3 .73 INV4 .75 INV5 .83 INV6 .73 173 The results in Table 4.7 show the AVE and CR values for all constructs exceeds the threshold value of 0.5 and 0.6 respectively. Thus, the study concludes that the Convergent Validity and Composite Reliability for all constructs in the model have been achieved (Zainudin, 2015). (b) Construct Validity - From the Pooled-CFA results, the researcher intended to identify the Fitness Indexes for the measurement model, the factor loading for each item, and the correlation between constructs. Table 4.8: Fitness Indexes for Fitness of the Constructs Name of category Name of index Index value Comments 1. Absolute fit RMSEA 0.058 Required level achieved 2. Incremental fit CFI 0.904 Required level achieved 3. Parsimonious fit ChiSq/df 2.185 Required level achieved Table 4.8 shows that the required values for the Fitness Indexes have been achieved for Construct Validity as proposed by Zainudin (2012). Based on this, the measurement model is said to have achieved Construct Validity. (c) Discriminant Validity - The final step of the Pooled-CFA method, the study assessed the discriminant validity of the constructs to clarify that they are not redundant of each other. Discriminant validity for the construct is achieved if the correlation among the independent variable in the model does not exceed 0.85 (Zainudin, 2015). The study also developed a Discriminant Validity Index Summary (Table 4.9) for all constructs in the model. 174 Table 4.9: Discriminant Validity Index Summary PBC SN ATT GI PV BI INV Perceived Behavioural Control (PBC) .754 Subjective Norm (SN) .697 .711 Attitude (ATT) .566 .623 .736 Governmental Influence (GI) .543 .559 .486 .733 Project Viability (PV) .611 .615 .601 .656 .711 Behavioural Intention (BI) .635 .590 .542 .607 .586 .716 Involvement (INV) .592 .609 .595 .567 .682 .588 .738 The diagonal values in bold presented in Table 4.9 are the square root of the AVE of the respective constructs, while other values are the correlation between the respective pair of constructs. The Discriminant Validity of the particular construct is achieved if the square root of its AVE exceeds its correlation value with other constructs in the model. In other words, the discriminant validity is achieved if the diagonal values (in bold) are higher than any other values in its row and column. Based on these criteria, and looking at the tabulated values in Table 4.9, the study meets the threshold of discriminant validity. Thus, the study concludes that the Discriminant Validity for all constructs is achieved. (iii) Reliability Reliability is the extent of how reliable the said measurement model is in measuring the proposed latent construct. The assessment for reliability for a measurement model uses the following criteria: - 175 a) Composite Reliability (CR) – The CR indicates the reliability and internal consistency of a latent construct. A CR value of 0.6 or higher is required in order to achieve composite reliability. Table 4.6 shows the CR value for the constructs perceived behavioural control (0.902), subjective norm (0.754), attitude (0.892), governmental influence (0.778), project viability (0.836), behavioural intention (0.840) and involvement (0.877). As such, all the constructs achieve the required CR. b) Average Variance Extracted (AVE) – The AVE value indicates the average percentage of variation explained by measuring the items for a latent construct. An AVE of at least 0.5 is required for every construct. Table 4.6 shows the AVE for constructs perceived behavioural control (0.569), subjective norm (0.506), attitude (0.543), governmental influence (0.538), project viability (0.506), behavioural intention (0.514) and involvement (0.545). Based on this, all the constructs have achieved the requirement of AVE. 4.4.2 Structural Model This section explains the process of analysing data using the Confirmatory Factor Analysis (CFA) method. This includes the processes of assessing measurement validity, uni-dimensionality and reliability. Prior to modelling the structural model and executing SEM, the study is first required to validate all latent constructs involved in the model (Zainudin, 2015). This validation procedure is known as CFA, which can 176 be executed using two methods – the single construct CFA and the Pooled-CFA for all constructs. The section illustrates the process of answering the research objectives and research questions that have been earlier presented in Chapter 1, mainly those related to the use of SEM analysis. After confirming the validity of the measurement model, the structural model is specified as by assigning relationships between the constructs based on the conceptual framework developed in Chapter 2 (Figure 4.4). The following hypotheses were then tested: H1: The decision makers’ attitude has a positive effect on their behavioural intention towards involvement in PPP toll expressway projects. H2: Subjective norms have a positive effect on the behavioural intention of the decision makers in the private sector towards invovement in PPP toll expressway projects. H3: Perceived behavioural control has a positive effect on the behavioural intention of the decision makers in the private sector towards involvement in PPP toll expressway projects. H4: Governmental influence has a positive effect on the behavioural intention of the decision makers in the private sectors towards involvement in PPP toll expressway projects. H5: Project viability has a positive effect on the behavioural intention of the decision makers in the private sectors in PPP toll expressway projects. 177 H6: The behavioural intention of the decision makers in the private sector has a positive effect on their involvement behaviour in PPP toll expressway projects. Figure 4.4: Structural Model The results displayed within Table 4.9 affirms that attitude, perceived behavioural control, government influence contribute significantly to behavioural intention and behavioural intention significantly impact on involvement. Meanwhile, subjective norm and project viability are not significant to behavioural intention. 178 According to Hair et al (2016), the significance level for p value is below than 0.05 and the critical ratio (CRa) is above 1.96. Governmental influence report the highest contribution (β=0.415, t-value = 5.780 >1.96) followed by perceived behavioural control (β=0.329, t-value = 4.131 > 1.96), and finally attitude (β=0.205, t- value = 2.841 > 1.96) which has a very little bearing on involvement. On the other relationships, behavioural intention contributed significantly to involvement (β=0.657, t-value =11.080 >1.96). Table 4.10: The Hypotheses Result based on Bootstrapping H Estimate (β) SE CRa t=value P Result H1 Attitude  Behavioural Intention .205 .072 2.841 .004 Supported H2 Subjective Norm  Behavioural Intention .030 .123 .243 .808 Not Supported H3 Perceived Behavioural Control  Behaviour Intention .329 .080 4.131 *** Supported H4 Governmental Influence  Behavioural Intention .415 .072 5.780 *** Supported H5 Project Viability  Behavioural Intention .094 .101 .934 .350 Not Supported H6 Behavioural Intention  Involvement .657 .059 11.080 *** Supported Therefore, H1, H3, H4 and H6 are supported and H2 and H5 are not supported. Based on these results, the relationship between behavioural intention and involvement is significant. Subsequently, this study proceeded to analyse the moderating effect to address H7, as presented in the following sections. 179 However, before proceeding to analyse the moderating effect under H7 hypothesis, the author extended an additional exercise of H2 and H5 due to its insignificant result. In clarifying further on these insignificant results of H2 and H5, the interview exercise with few Chief Executive Officer (CeO) of the toll concessionaires’ companies have been conducted. Nevertheless, this interview exercise is only to seek clarification in justifying the result, thus content analysis is not required as this interview process is not part of the study. Due to insignificant result of H2 (subjective norm) and H5 (project viability), the interview with CeO of X, Y and Z company (undisclosed real name due to privacy reason) have been conducted and every insignificant construct have been asked a question as follow; Question 1: “Why do you think subjective norm or significant other is not important element for the company in considering involvement in PPP toll expressway project?” Answered of Question 1 by CeO of X toll concessionaire: “In PPP toll expressway business, the third parties neither give much impact nor influence on the consideration of participation, as management team in the companies would work very hard to secure this huge project and looking forward to have long-term contract with the government.” Answered of Question 1 by CeO of Y toll concessionaire: “In developing country, such as Malaysia, the influences of local society such as non-governmental organisations or association bodies do not jeopardise 180 much on the company’s decision making particularly in deciding to get involve in government contract.” Answered of Question 1 by CeO of Z toll concessionaire: “PPP is between the company and government, thus the company should take full consideration and cautioned more on fulfilling obligation between partners rather than too focus on others. Thus, besides stakeholders, others do not much influence in our decision making in PPP.” Question 2: “Why do you think project viability is not important element for the company in considering involvement in PPP toll expressway project?” Answered of Question 2 by CeO of X toll concessionaire: “Normally the government already took full study on the viability of the PPP project before call for any tender or RFP (request for proposal), thus, in deciding to participate or not, the company more focus on financial feasibility in term of projected revenue rather than overall of project viability.” Answered of Question 2 by CeO of Y toll concessionaire: “By having a concession agreement in PPP, the viability of project would not be the main concern for the company as the profit guaranteed and projected future profit are clearly stipulated in the contract.” 181 Answered of Question 2 by CeO of Z toll concessionaire: “We have a very good government and our government policy is a business- friendly policy particularly in lure of private sector’s participation and involvement in the public infrastructure projects. Hence, project viability should not be an issue as the company could rely on full government support and assistance especially when the nature of project involves national interest.” 4.5 Multiple Group Analysis – Moderation Analysis Zainudin (2015) emphasises the complexities in analysing the moderating effect for a model with latent construct. The multi-group CFA suggests an alternative method for this process, with two separate models – one is constrained model and the other is an unconstrained model. The path of a constrained model would have a parameter = 1. This study has adopted the moderation process for multi-group CFA by Zainudin (2015), looking at trust as the moderator variable in the relationship between behavioural intention and involvement behaviour of the private sector in PPP toll expressway projects (H7). This is depicted in Figure 4.5. In testing trust as a moderator, the researcher divided two datasets – high data trust (scale 3.51 to 5.00) and low data trust (scale 1.00 to 3.50). Each dataset was tested using the two models – constrained model and unconstrained model. 182 Figure 4.5: Trust as a Moderator The results of the tests are significant if the difference in Chi-Square value for the constrained and unconstrained model is found to be above 3.84. Two types of moderators, full moderator and partial moderator, were seen. Full moderator occurs when high data or low data is significant, while, partial moderator occurs when both high and low data are significant (Zainudin, 2015). Trust was analysed as a moderator on both tests, high and low data, using constrained and unconstrained model. 4.5.1 Test Moderation for High Data - Trust In elaborating this test, Figure 4.6 High Trust Data: Outputs for the Constrained Model, Table 4.11 Chi-Square Value and DF for the Constrained Model for High Data Trust, Figure 4.7 High Trust Data: Outputs for the Unconstrained Model, and Table 4.12 Chi-Square Value and DF for the Unconstrained Model for High Data Trust are referred. Further explanation is provided in the following paragraphs. Involvement Behaviour Behavioural Intention Trust 183 Figure 4.6: High Data Trust: Outputs for the Constrained Model Table 4.11: Chi-Square Value and DF for the Constrained Model for High Data Trust Model NPAR CMIN DF P CMIN/DF Default model 32 200.056 88 .000 2.273 Saturated Model 120 .000 0 Independence Model 15 795.408 105 .000 7.575 184 Figure 4.7: High Data Trust: Outputs for the Unconstrained Model Table 4.12: Chi-Square Value and DF for the Unconstrained Model for High Data Trust Model NPAR CMIN DF P CMIN/DF Default model 33 170.587 87 .000 1.961 Saturated Model 120 .000 0 Independence Model 15 795.408 105 .000 7.575 Table 4.11 shows the output of chi-square values for the constrained model, where the CMIN is 200.056 and DF is 88. Meanwhile, Table 4.12 shows the output of chi-square values for unconstrained model, where the CMIN is 170.587 and DF is 87. 185 For the moderators to be significant, the difference in Chi-Square value between constrained model and unconstrained model must be higher than 3.84 with 1 degree of Freedom (Zainudin, 2012). Table 4.13 shows the moderation test for High trust is significant as the difference in Chi-Square value between the constrained (200.056) and unconstrained (170.587) models is 29.469 (200.056-170.587) while the Degree of Freedom is 88-87 = 1. Table 4.13: The Moderation Test for High Data Trust Results Constrained Model Unconstrained Model Chi-Square Difference Result on Moderation Chi-Square 200.056 170.587 29.469 Significant DF 88 87 1 The test carried out for H7 on moderation found that the moderator variable for High trust does have a moderating effect on the relationship between behavioural intention and involvement behaviour as the difference in Chi-Square value is 29.469, which is above the requirement of 3.84. 4.5.2 Test Moderation for Low Data - Trust In elaborating this test, Figure 4.8 Low Data Trust: Outputs of the Constrained Model, Table 4.13 the Chi-Square Value and DF for the Constrained Model for Low Data Trust, Figure 4.9 Low Data Trust: Outputs for the Unconstrained Model and Table 4.14 the Chi-Square Value and DF for the Unconstrained Model for Low Data Trust are referred. Further explanation is provided in the following paragraphs. 186 Figure 4.8: Low Data Trust: Outputs for the Constrained Model Table 4.14: Chi-Square Value and DF for the Constrained Model for Low Data Trust Model NPAR CMIN DF P CMIN/DF Default model 32 291.141 88 .000 3.308 Saturated Model 120 .000 0 Independence Model 15 2766.463 105 .000 26.347 187 Figure 4.9: Low Data Trust: Outputs for the Unconstrained Model Table 4.15: Chi-Square Value and DF for the Unconstrained Model for Low Data Trust Model NPAR CMIN DF P CMIN/DF Default model 33 205.258 87 .000 2.359 Saturated Model 120 .000 0 Independence Model 15 2766.463 105 .000 26.347 Table 4.14 is output of chi-square value for the constrained model, where the CMIN is 291.141 and DF is 88, while, Table 4.15 shows the output of chi-square value for the unconstrained model, where the CMIN is 205.258 and DF is 87. For the test moderation to be significant, the difference in Chi-Square value between the constrained and unconstrained models must be higher than 3.84 with 1 degree of freedom (Zainudin, 2012). 188 Table 4.16 shows the moderation test for Low Data Trust is significant since the difference in Chi-Square value between the constrained and unconstrained model is 85.883 (291.141-205.258), while the Degree of Freedom (DF) is 88-87 = 1. Thus, the Chi-square value is above than 3.84 and therefore, the low data trust is significant. Table 4.16: The Moderation Test for Low Data Trust Results Constrained Model Unconstrained Model Chi-Square Difference Result on Moderation Chi-Square 291.141 205.258 85.883 Significant DF 88 87 1 The tests for H7 on moderation that has been carried out found that the moderator variable for High and Low trust does have a moderating effect on the relationship between behaviour and involvement. The next step is for this analysis to identify which data group (low trust or high trust) shows a more definite moderator variable (trust) effect. This process requires the standardised estimates for the path of interest for both datasets. Table 4.17: The Standardised Estimate for High Data Trust Construct Standardised Beta Estimate P Result Behavioural Intention  Involvement .260 .000 Supported Table 4.18: The Standardised Estimate for Low Data Trust Construct Standardised Beta Estimate P Result Behavioural Intention  Involvement .171 .001 Supported 189 Based on the results, the standardised beta estimate for high data trust is 0.260 with the P value 0.000 (Table 4.17) and for low data trust is 0.171 with the P value 0.001(Table 4.18). As such, it can be concluded that the effect of the moderation is supported in both high trust and low trust. The results also show that the type of moderation is partial moderation as the standardised estimates for both high data trust and low data trust are significant. In conclusion, H7 is supported where trust is confirmed as a partial moderator in the relationship between behavioural intention and involvement. 4.6 Final Structural Model Figure 4.10 shows the critical ratio and followed by p value for the Final Structural Model. The critical ratio indicates the level of factor loading, where a minimum critical ratio value of 1.960 is required for it to be considered significant (Byrne, 2010). Meanwhile, a p value of below 0.05 is considered to be significant. Figure 4.10: Final Structural Model 190 Table 4.19 shows the summary of seven hypotheses of this study. The results for each hypothesis are as follow: H1, H3, H4, H6 and H7 are ssupported and H2 and H5 are not supported. Table 4.19: Summary of Hypotheses Hypothesis Statement Result H1 The decision makers’ attitude has a positive effect on their behavioural intention towards involvement in PPP toll expressway projects. Supported H2 Subjective norms have a positive effect on the behavioural intention of the decision makers in the private sector towards invovement in PPP toll expressway projects. Not Supported H3 Perceived behavioural control has a positive effect on the behavioural intention of the decision makers in the private sector towards involvement in PPP toll expressway projects. Supported H4 Governmental influence has a positive effect on the behavioural intention of the decision makers in the private sectors towards involvement in PPP toll expressway projects. Supported H5 Project viability has a positive effect on the behavioural intention of the decision makers in the private sector towards involvement in PPP toll expressway projects.. Not Supported H6 The behavioural intention of the decision makers in the private sector has a positive effect on their involvement behaviour in PPP toll expressway projects. Supported H7 Trust has a moderating effect on the relationship between behavioural intention and involvement behaviour in PPP toll expressway projects Supported 191 4.7 Conclusion This chapter presents in detail the findings of this study. Overall, the response rate to the survey conducted was good and focused, with 88 per cent of respondents strictly from toll concessionaire companies in Malaysia. The respondents also show a balanced ratio between middle and top management employees, with the majority of them having over 11 years working experience. The data screening process was conducted to remove the outliers, confirm normality as well as to check for multi-collinearity. Confirmatory factor analyses were conducted to ensure construct validity of all variables. Reliability tests were conducted to determine the internal consistency between items and minimise random errors. On top of that, discriminant validity assessment was done to provide an additional measure for all remaining items. This chapter also presented the results of the hypotheses tests, effect of the moderator, and the outcome of the structural equation models. The results, in general, supported most of the hypotheses, except subjective norms and project viability. The study’s overall results and recommendations will be discussed in greater detail in the next chapter.