We have used SPSS 12.0, AMOS 6.0 and NLOGIT 3.0 softwares to analyze the collected data. The current research conducted the following data analysis.
3.5.1 Descriptive Analysis
To better understand the characteristics of each variable, we use descriptive statistics to explain the structure of sample data and show the distribution of our sample.
3.5.2 Factor Analysis
Factor analysis takes a large number of variables, and puts them into a small number of factors, within which all of the variables are related to each other. Factor analysis can identify the basic underlying variables which account for the correlations between actual test scores. The purpose of factor analysis is to explore the underlying variance structure of a set of correlation coefficients. Factor analysis can be used not only to summarize or reduce data but also to explore or confirm data.
Principal components method with a varimax rotation was employed to the measurement scales and to examine their dimensionality. The guidelines adopted in this study for identifying factors are (1) eigenvalues should be greater than 1; (2) cumulative explained variance is suggested to exceed 60%; and (3) factor loading is generally required above 0.50 (Wu, 2005). Then Cronbach`s α was used to confirm the reliability of each extracted factor.
6 www.ezship.com.tw.
37
In this thesis, we will conduct exploratory factor analysis on the logistics service quality items, using principal component analysis as the extraction method and virmax as the rotation method. We used the subgroup numbers criterion to determine the factors to retain. We then went through a series of iterations, each involving the elimination of items with low loadings on all factors or high cross-loadings on two or more factors, followed by the factor analysis of the remaining items. The iterative process resulted in the final LSQ Scale, consisting of 13 items on four dimensions, which we labeled as timeliness, information quality, convenience and personnel contact quality as shown in Table 4.6.
3.5.3 Reliability
Reliability means the trustworthiness of measurement, like accuracy or precision.
It also represents the stability or consistency of a result. Reliability is depended on error of measurement. It reflects the degree of trustworthiness of measuring tools or procedures.
There are three kinds of reliability: equivalence, stability, and consistency.
Equivalence is divided into alternate forms and split-half, stability concludes the test-retest, and is divided into split-half, Kuder-Richarson and Cronbach`s α. They are suitable for different proposes and situations. In this thesis, we use Cronbach`s α value to test the consistency of measurements of each factor, it is most suitable for testing reliability under Likert scale.
Cronbach`s α is proposed by Cronbach(1951). Cronbach proposed a principle to determine reliability. The rationale for Cronbach`s α is that the individual scale items should all be measuring the same construct and thereby the highly intercorrelated.
Cronbach`s α is a measure of squared correlation between observed scores and true scores. The lower limit for Cronbach`s α is generally agreed on 0.70, although it may decrease to 0.60 in exploratory research (Hair, nderson, Tatham, & Black, 1998). If a scale has a Cronbach`s α below 0.60, it should be considered for any roots of measurement errors. In practice, as long as α >0.60, we can claim an acceptable reliability.
𝑁 𝑁 − 1
σ𝑋2 − 𝑁𝑖=1σ𝑌2𝑖 σ𝑋2
N = the number of components
σX2= the variance of the observed total test σY2i= the variance of component i
38
3.5.4 Validity
Validity is a scale to examine the degree of measurement. In this thesis, we use construct validity to measure the effectiveness of model. Construct validity means that the construct can reflect actual situation. Construct validity divides into convergent validity and discriminate validity. Convergent validity refers to the extent to which each measure correlates with other measures of the same latent construct. If measurement items of each construct have individual factor loadings of at least 0.50 (Grandon & Pearson, 2003; Lee, Kim, & Lee, 2004), and all measurement items are significant with a t-value greater than 1.96, we can conclude that the scale has convergent validity.
Discriminate validity, in contrast, refers to the extent to which the measure of a conduct does not correlate with the measures of other constructs. Discriminate validity can be assessed based on a confidence interval of the correlation between any two constructs, namely, to test H0: ρ = 1 versus H1: ρ ≠ 1 (ρ is the correlation between paired constructs). If none of the correlations includes 1, discriminate validity is reached, and we can conclude that the two constructs differ (Anderson &
Gerbing, 1988).
In this thesis, we will do the t-test for factor loading of every indicator variable after finishing confirmatory equation analysis. If the t value of the factor loading comes from every indicator variable and its construct is higher than 1.645, it means that every measured variables can effectively measure the common construct.
About testing discriminate validity, we take variance extracted estimate as the indicator. Usually, variance extracted estimate of measurement construct should be higher than 0.5 to fit in with the standard proposed by Fornell and Larcker (1981).
3.5.5 Structural Equation Modeling (SEM)
Structural equation modeling is usually categorized as advanced statistics. It belongs to a part of multivariate statistics and integrates factor analysis and path analysis. SEM concludes the relationships between manifest variables, latent variables, error variables and further obtains direct effects, indirect effects and total effects caused from independence variables to dependence variables. SEM encourages confirmatory rather than exploratory modeling. Therefore, it is critical to all construct of SEM modeling that must be directed by theory for model development and modification.
SEM is characterized by two basic components: (1) the measurement model, allowing the researcher to use several variables for a single independent or dependent variable; (2) the structural model, relating independent to dependent variables (i.e.,
39
the path model). The superiority of structural equation model over other statistical techniques owes to its ability to estimate multiple and interrelated dependence relationships, and also to represent unobserved concepts, or latent variables, in those relationships and account for measurement error in the estimation process. The conceptual model is used and shown in Figure 3.1 to explain the relationships among the logistics service quality, perceived value, customer satisfaction, switching cost and customer behavioral intention.
Goodness-of-fit measures the correspondence of the actual or observed input (covariance or correlation) matrix with that predicted from the proposed model. In other words, goodness-of-fit tests are used to determine whether the model should or should not be rejected. Jöreskog & Sörbom (1993) pointed out that concept of measurement model concludes measurement, reliability and validity. So, the complete analysis of structural model consist of (1)calculation of factor loading of each variable, (2)testing the fitness between data and measurement model of each factor, (3)calculation of the relationship between each latent variable, and (4) testing the fitness between whole model and data.
Table 3.3 Goodness-of-Fit index of model
Goodness-of-Fit Measurement Threshold value Absolute Indexes of Fit
Chi-square Statistic (2) P<0.05; (The smaller; the better) Normed Chi-square (2/df ) P <2.00 is perfect;
P <5.00 is acceptable Root Mean Square Residual (RMR) P<0.05
Goodness of Fit (GFI) P>0.9 Adjusted Goodness of Fit (AGFI) P>0.9 Root Mean Square Error of
Approximation (RMSEA)
P <0.05 is perfect;
P <0.08 is good;
P <1 is acceptable) Incremental Indexes of Fit
Normal Fit Index (NFI) P>0.9 Non-normal Fit Index (NNFI) P>0.9 Relative Fit Index (RFI) P>0.9 Comparative Fit Index (CFI) P>0.9
40
3.5.6 Logit Model
The cognitive process for decision-making is the mental mechanism that defines the cognitive task and the role of perceptions, affect, attitudes, motives and preferences in performing this task to produce a choice. Random utility (or discrete choice) models have been extensively used to analyze consumers‟ choice behavior.
These models used only the observed attributes and individual characteristics. In behavioral sciences, there are some concepts that cannot be directly measured.
Ben-Akiva et al. (1999) presented a theoretical framework which integrated choice and latent variable model to incorporate psychological factors and their influences on choices.
In this thesis, we will specify the joint choice and latent variable model, and we need the structural and measurement equations for both the latent variable component and the choice component. The parameters are assumed to be liner and the equations are constructed as:
Latent variable structural equations:
𝑈𝑖𝑛 = 𝑋𝑖𝑛𝛽 + 𝑋𝑖𝑛∗ 𝛽𝑋∗ + 𝜀𝑖𝑛 𝜀𝑖𝑛~𝐺𝑢𝑚𝑏𝑒𝑙(0, 𝜇) (11) 𝑋𝑖𝑛∗ = 𝑋𝑖𝑛𝜆 + 𝜔𝑖𝑛, 𝜔𝑖𝑛~𝑁(0,1). (12)
Latent variable measurement equations:
𝐼𝑖𝑛 = 𝑋𝑖𝑛∗𝛼𝑖 + 𝜐𝑖𝑛, 𝜐𝑖𝑛~𝑁(0, 𝜃𝜐𝑏). (13) 𝑦𝑖𝑛 = 1, 𝑖𝑓 𝑈𝑖𝑛 = 𝑚𝑎𝑥𝑗 𝑈𝑗𝑛
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (14) The likelihood function for the joint model is:
𝑃 𝑦𝑖𝑛, 𝐼𝑖𝑛|𝑋𝑖𝑛, 𝑋𝑖𝑛∗ ; 𝛽, 𝛽𝑋∗, 𝛼, 𝜆, 𝜃𝜀, 𝜃𝜔, 𝜃𝜐 =
𝑃 𝑦𝑖𝑛|𝑋𝑖𝑛, 𝑋𝑖𝑛∗ ; 𝛽, 𝛽𝑋∗, 𝜃𝜀 𝜑 𝑋𝑖𝑛∗ 𝑓𝑖 𝐼𝑖𝑛|𝑋𝑖𝑛∗ ; 𝛼, 𝐼𝜐 𝑑𝑋∗𝑑𝜐𝑑𝜀 ∙ (15) which describes the decision-maker (n) and the alternative (i), the latent variable (𝑋𝑛∗) ,the observable explanatory variables (𝑋𝑛 ), unknown parameters (𝜆, 𝛽, 𝛼), and covariances of random disturbance terms (𝜔𝑖𝑛, 𝜀𝑖𝑛, 𝜐𝑖𝑛).
The model fit is measured in terms of the fit between the estimated choice probabilities and the observed choices, and in terms of the ability of the model to forecast the observed response. In this study, we used 𝜌2 to measure the model fit, which is similar to R2 in the regression analysis. It is defined as:
41
0 ) 1 (
2
LL LL
0 ) )
( 1 (
_ _2
LL
K LL
where 𝜌 2 is adjusted likelihood ratio, LL(
) and LL(0) are the log-likelihood function values at convergence and at zero, K is the number of parameters estimated in the model. According to the literatures, the preferred procedure for evaluating a model is to use the log-likelihood value or transforms of it, such as 𝜌2 and 𝜌 2.
We will use AMOS 6.0 to obtain the parameters of the latent constructs and calculate the factor score. We put them in Eq. (11) to Eq. (13) to analysis which items will influence the respondents‟ choice behavior to choose an RD provider.
42