In order to analyze the effect of knowledge sharing on KMS and innovation and further examine the relationships between organization context, knowledge sharing, KMS, and innovation, a mediation regression approach was adopted. The mediated regression analysis followed a three-step process as described by Baron and Kenny (1986). According to Baron and Kenny (1986) there’s mediation when the following conditions are met:
1) the independent variable (s) is/ are significantly related to the dependent variables (knowledge sharing) in the first test;
2) the independent variable (s) is/ are significantly related to the mediator (s) in the second test;
3) the mediator (s) is/are significantly related to the dependent variables ; and. the beta value of the independent variables on the dependent variables is less in the third test than in the second test.
According to Baron and Kenny (1986), full mediation applies when an independent variable becomes insignificant in the final test. Partial mediation exists when an independent variable is still significant in the final test for mediation procedure.
63
A series of validity and reliability tests were carried out to ensure the accuracy of the measurements. Reliability tests were conducted to evaluate the consistency of the participants’
answers to the items in the questionnaire. Reliability tests included Cronbach’s alpha item if deleted, Cronbach’s alpha coefficient, and Corrected item-total correlation.
Statistical Analyses in SPSS 19.0
Cronbach’s alpha item if deleted was used to identify what items in a scale were not homogenous and did not load on the same scale. This was achieved by checking the “Cronbach’s alpha item if deleted” column and examining whether any entries in the column were greater than the overall Cronbach’s alpha coefficient. Bartee et al. (2004) explained that if Cronbach’s alpha increased as a result of deleting any particular items, this would signal that these items measure something different than what the other items measure. The authors further added that these items might not belong in the same scale with others.
Item total correlation was examined to determine how well the items in a scale were related to one another. In addition it also helped the researcher to evaluate the performance of the questions. Bartee et al. (2004) stressed that item total correlation values should be positive. It is recommended that values for item total correlation should be greater than three. Values for an item total correlation between 0 and .19 may suggest that the question is not discriminating well, and values .40 and above indicate very good discrimination.
The research instrument construct validity was assessed through factor analysis. At least four criteria were taken into account when examining the appropriateness of the factors’ scales.
First, individual item loading should exceed the factor loading cutoff point respective to the final sample size (see table 3.1, p.58) recommended by Hair et al. (2006). For example, according to Hair et al. (2006) sample of 350 requires that the factor loading for each construct’s item should to be greater than .30 while the threshold for a sample size of 50 is .75. Second, the construct’s eigen value should be greater than 1 (Hair et al., 2006). Third, each factor should explain at least 50 % of the variance of its indicator and finally Kaiser-Meyer-Olkin measure of sampling adequacy values should be between .50 and 1 (Krizman, 2009). Kaiser-Meyer-Olkin is useful to determine the appropriateness of the data set for the factor analysis.
64
Following reliability and validity analyses, descriptive statistics and hierarchical regression analysis were conducted. As a first step in conducting hierarchical regression, correlation analysis was carried out to diagnose the strength of the relationship between the independent variables, the mediator and dependent variables. In addition, Pearson correlation tests were performed to fulfill Baron and Kenny (1986)’s primary requirement for mediation analysis.
Based on their guidelines, the correlations among the independent variables (i.e. organization contexts) should be significantly associated to the mediator variable (i.e. knowledge sharing).
Authors such as Staples & Webster, 2008; Zhi-Hong, Li, Bo, & Shu, 2008 indicated that structural equation modeling (SEM) is appropriate for a mediated research model. Zhi-Hong et al.
(2008) further added that SEM is especially effective when testing models that 1) are path analytic with mediating variables; and 2) contain latent constructs that are measured with multiple indicators (p. 974). Since the research model fulfilled these characteristics, PLS-based SEM analysis of variance was used for data analysis.
Statistical Analyses in Smart PLS
In addition, PLS uses a combination of principal component analysis, path analysis and a set of regressions to simultaneously evaluate the theory and the data (Staples & Webster, 2008, p.
626; Yu, Kim, & Kim, 2007). Aside from being an effective tool to confirm both theory and predictions, PLS was used in this study for the following reasons. First, PLS is adequate for small sample analysis and complex models. PLS necessitates a sample size that is ten (10) times the total indicator for complex construct of the research. Since this study 6 indicators therefore a sample size of (6x10) =60 would be necessary.
Second, PLS is also useful for exploratory analysis investigating whether relationships might exist among variables (Yu, Kim, & Kim, 2007). In addition PLS have been applied in previous studies related to organization behavior (e.g. Higgins, Duxbury, & Irving, 1992), KMS success (e.g. Al-Busaidi et al., 2010), innovation and knowledge sharing (e.g. Bock et al., 2005;
Krizman, 2009). Third, PLS is essential for model including formative and reflective indicators as it is the case for this study. For these reasons and the ones mentioned above, PLS was used to test the hypothesized relationships among the study variables. Following the recommendations
65
indicated in (Al-Busaidi et al., 2010, Bock et al., 2005; Karim, 2009) studies, analyses in PLS were carried and interpreted in two stages: the measurement model and the structural model.
Measurement Model
The reliability of the constructs was determined by composite reliability and Cronbach’s alpha. Nunally and Bernstein (1994) suggested that Cronbach’s alpha value of .70 is satisfactory.
Similarly, to Cronbach’s alpha, composite reliability was considered as a measure for internal consistency. According to Lawson et al. (2009) a value of .70 or higher is recommended as evidence of composite reliability. Construct validity was assessed by examining the loadings of the constructs and the average variance extracted (AVE). Whereas Li et al. (2009) advanced that a loading of .70 is considered desirable, Hair et al. (2006) suggested that the threshold for the factor loading is determined by the size of the sample size (see Table 3.2). For AVE, a cut-off value above or equal to .50 or higher is considered acceptable (Al-Busaidi et al., 2010; Li et al., 2009).
Structural Model
The structural model was evaluated by examining the value of R2 of the latent variables, the path coefficients and the goodness of fit. The R2 values were used to examine the predictive relevance of a structural model for the dependent latent variable. As in indicated in Karim (2009), this study also considered R2 values of .67, .33, and .19 of latent variables as substantial, moderate, and weak respectively. With regards to the quality of the model, three indicators were taken into account: communality indexes and average communality, average redundancy and goodness of fit. The goodness of fit (GoF) of the model was used to determine the overall fit of the model and was assessed on the basis of Tenenhaus, Vinzi, Chatelin, and Lauro (2005) global fit measure. The corresponding formula is written as follow:
66
67