Chapter 3: Methodology
3.4 Analytical Methods
This study uses several analytical methods to test all hypotheses. First, in subsections 4.2, 4.3, 4.4, and 4.5, the structural equation modeling (SEM) technique is conducted to test the proposed model.8 The data are analyzed by using the LISREL 8.54 software, and the maximum likelihood estimation (MLE) method is used to estimate the factor structure of the proposed model. A standard two-step process is followed, in which CFA is firstly performed to assess the measurement model, and the structural model is then constructed when the measurement model is upheld (Anderson
& Gerbing, 1988). The model fit is assessed by using χ2/df, goodness-of-fit index (GFI), comparative fit index (CFI), normal fit index (NFI), and root mean square error of approximation (RMSEA). The threshold for χ2/df should be less than 3.0, or less than 2.0 in a more restrictive sense (Premkumar & King, 1994). Values of GFI, CFI and NFI should be over 0.90, while the value of RMSEA should be less than 1.0. In order to confirm the validity of the measurement model, both the convergent and discriminant validity are further tested (Venkatraman, 1989).
Second, to confirm the robustness of the moderating effect, a multi-group analysis in SEM technique is conducted in subsection 4.6. Following Brockman and Morgan (2006), a two-step approach is used to test the moderating effect. First, the appropriate structural parameters are constrained to be equal across groups, thereby generating a
8 In contemporary studies, the measurement (that is, factor analysis) and structure (that is, path analysis) have been integrated into SEM since the 1970s (Aryee, Budhwar, & Chen, 2002; Bagozzi, 1988;
Moreno & Casillas, 2008).
- 86 -
covariance matrix for each group and an overall chi-square value for the sets of sub-models as part of a system of structural models. Second, the parameter equality constraints must be removed, resulting in the chi-square value with fewer degrees of freedom. The moderating effect is identified by comparing difference of two chi-square values.
Third, to confirm whether the proposed model fit the data well, comparison of alternative models by using SEM technique is conducted in subsection 4.7. This study examines three alternative models (direct model and two partially models) by comparing them with our hypothesized models.
Forth, the ordinary-least-squared (OLS) hierarchical regression analysis is then used to examine the possible mediation and moderating effect of resource attributes on the relationship between EO and firm performance. With respect ct to the mediating effect, following Baron and Kenny (1986), three processes are used to test the mediation effect in subsection 4.8: (a) regression models are constructed by using only the mediated variable (that is, resource-capability combinations) as the regressor; (b) regression equations are constructed by using only the independent variable (that is, EO) as the regressor; and (c) regressions are conducted by introducing both the independent (EO) and mediated variables (resource-capability combination) into the models.
With respect to the moderating effect, hierarchical linear regressions are used again in subsection 4.9 when analyzing multiple terms in the regression equations. Following
- 87 -
Cohen and Cohen (1983), three processes are used to test the main-effect models and two-way interaction models. The two-way interaction models are constructed with the interaction of EO and environmental dynamism included in the equations. This study expects that each interaction term makes a significant contribution to the value/rareness of resource-capability combinations.
Finally, to provide considerable insight into the issue being studied and form the theoretical propositions in this study, the pilot case study is conducted and shown in Appendix B. Following Yin (1994), based on the existing literature, the information of a pilot case allows this study to develop relevant questions, producing research propositions and overall framework.
- 88 -
Chapter 4
Research Results
This chapter shows the research results by following statistical-analysis methods.
First, the results of descriptive statistics present correlations among all variables.
Second, the assessments of measurement model and comparisons of alternative measurement models are conducted in sections 4.2 and 4.3. Third, the mediating and moderating models conducted by SEM techniques showed in sections 4.4 and 4.5 respectively. Fourth, to confirm the moderating effect, the multi-group analysis is conducted in section 4.6. Fifth, section 4.7 provides the results of the comparisons on alternative models to examine whether the hypothesized model fits our data. Next two sections, to further confirm the mediating and moderating effects, regression analysis is used in sections 4.8 and 4.9 respectively. Finally, bivariate analysis is used in section 4.10 to examine again the moderating effect.
4.1 Descriptive Statistics
Table 4-1 reports the mean, standard deviations, and correlation coefficients of all the variables in this study. The degree of EO shows a high mean of 4.232, and the
- 89 -
degree of environmental dynamism also shows a high mean of 3.90. In addition, the degree of resource attributes shows a higher mean of 5.537 in the value and a higher of 5.444 in rareness. These results imply that public firms in Taiwan are more EO, and these firms usually possess higher value/rareness of resource-capability combinations in the high environmental dynamism.
As expected, the main effects are significantly correlated with the dependent variables and mediating variables. First, when examining the correlation between EO and firm performance variables (i.e., CA, satisfaction, ROA, and Tobin’s q), the positive correlations are found. Second, when examining the correlation between EO and mediating variables (i.e., value and rareness), the positive correlations are found. Third, as predicted, when examining the correlation between mediating variables (i.e., value and rareness) and firm performance (i.e., CA and satisfaction), the positive correlations are found. However, it is found that both value and rareness variables are not significantly correlated to firm performance (i.e., ROA, and Tobin’s q). In addition, another purpose of this study is to examine the relationships between EO, environmental dynamism, and resource attributes (i.e., value and rareness). It is fount that there exists a significant and positive relationship between EO and environmental dynamism. The correlations between environmental dynamism and resource attributes are positively significant.
The correlation coefficients among independent, dependent and control variables are very low, with the highest correlation of 0.649 between value and rareness. If a high
- 90 -
correlation is found among the variables, high multicollinearity problems may exist in this study. Chatterjee, Hadi and Price (2000) suggest that multicollinerarity is not serious when the largest variance inflation factor (VIF) is not greater. Therefore, the VIF values in the regression models is assessed and significant multicollinearity problems are not found (VIF＜2.0). This implies that no serious multicollinearity problems exist in this model.
In addition, this study uses self-reported data collected from CEOs or top managers (single respondent), so it may be vulnerable to common method variance (CMV). Using ex ante preventive methods, this study guarantees anonymity and mailed the questionnaires directly to the managers. To avoid respondents guessing the relationship between variables, this study also reduces item ambiguity and separated related items (Podsakoff et al., 2003). For the ex post testing methods, this study uses Harman’s single-factor test, a widely adopted post hoc remedy, to estimate whether the data have a CMV problem (Podsakoff and Organ, 1986). The result showed that the first factor accounted for only 10.34% of variance among variables. Therefore, the data do not have a serious CMV problem and single response bias.
91
Table 4-1 Descriptive Statistics and Correlation Coefficients of Variables (n = 201)
Variables Mean S.D. 1 2 3 4 5 6 7 8 9 10 11
1.ROA 6.704 8.217 1.000
2.TQ .681 .561 .443*** 1.000
3. Satisfaction 26.956 7.624 .349*** .270*** 1.000
4. CA 5.594 .777 -.028 .001 .239*** 1.000
5. EO 4.232 1.072 .138** .198** .287*** .195** 1.000
6. Value 5.537 .736 -.045 .017 .270*** .679*** .228*** 1.000
7. Rareness 5.444 .762 -.008 -.015 .288*** .669*** .242*** .649*** 1.000
8. Age 29.295 13.575 -.150** -.196*** -.018 -.032 -.080 -.033 -.068 1.000
9. Size 2.510 .562 .157** .073 .114* .097 .165** .046 .059 .245*** 1.000
10. Dynamism 3.902 1.153 -.003 .121* .118* .064 .496*** .130** .102 -.329*** .063 1.000
11. DEMKT .698 .990 -.351*** -.309*** -.144** .093 -.169** .075 .109 .107 -.190*** -.084 1.000
Note:
The VIF values are less than 2.0, implying that our model contains no significant multicollinearity problems.
* P < 0.1, ** P < 0.05, *** P < 0.01.
92
4.2 Measurement Assessment Procedures
Following the study of Anderson and Gerbing (1988), a two-step analytical approach is used to test the hypothesized model. To test the construct validity of the measures, the study firstly employs a confirmatory factor analysis (CFA) using LISREL 8.54 (Bentler and Wu, 1995) and then conducts a SEM based on the measurement model to estimate the fit of the hypothesized model to the data. Unlike the traditional and more commonly-used EFA (exploratory factor analysis), the CFA contains inferential statistics that allow for a stricter and more objective interpretation of validity (Gerbing and Anderson, 1988). Moreover, SEM has certain advantages: (1) It offers a simultaneous test for an entire system in the proposed model. (2) It can assess whether or not the model is consistent with the data (Byrne, 1994).
A measurement model represents that the measure items (i.e., observed variables) are posited to underlying constructs (i.e., latent variables) (Bollen, 1989). This confirmatory assessment approach comprises both convergent validity and discriminant validity. First, the significant of factor loading and the average variance extracted (AVE) are used for the verification of convergent validity. The results of Table 4-2 confirm the convergent validity of the scales because the estimated coefficients of all indicators are significant on their posited underlying constructs (t＞1.96) (Anderson & Gerbing, 1988). All the AVEs are above 0.5, implying that the indicator variables can respond to the constructs (Bagozzi & Yi, 1988).
Therefore, convergent validity is confirmed. Finally, in Table 4-3, regarding discriminant validity, the results show that the confidence intervals of the correlations for the constructs excluded 1.0, implying the discriminant validity of inter-constructs. In addition, discriminant
93
validity is assessed by comparing the unconstrained model with the constrained model in which the correlation between the two constructs is constrained to 1.0 (Anderson & Gerbing, 1988; Jöreskog and Sörbom, 1989). The results show that each pair of constructs has a significant difference (see Table 4-3). Therefore, discriminant validity is also achieved.
Table 4-2 Parameters of Measurement Model Construct Standardized
Firm performance 0.75 0.66
CA1 0.84 0.71 19.00
Note: CR represents composite reliability; AVE represents the average variance extracted.
94
Table 4-3 Analysis of Discriminant Validity (n = 201)
Construct EO Value Rareness Environmental dynamism
Value 42.09***
(0.153, 0.309) Rareness 36.72***
(0.101, 0.257)
38.47***
(0.533, 0.807) Environmental dynamism 20.29***
(0.058, 0.214)
74.70***
(0.561, 0.898)
60.30***
(0.585, 0.899)
Firm performance 46.16*** 43.30*** 37.96*** 77.47***
(0.058, 0.214) (0.058, 0.214) (0.585, 0.899) (0.585, 0.899)
Note: The statistics compare the differences between the unconstrained model and the constrained model. The estimated confidence intervals are in
parentheses.
* P < 0.10, ** P < 0.05, *** P < 0.01.
95
This study estimates the first-order factor measurement models by dividing the constructs of the models into four theoretically plausible groups (i.e., Model 1: three dimensions of entrepreneurial orientation; Model 2: two constructs of resource attributes;
Model 3: performance and competitive advantage; Model 4: environmental dynamism). The CFAs provide an acceptable fit for the four measured models, and the results are exhibited in Table 4-4.
This study conducts a CFA for EO, value, rareness, competitive advantage, satisfaction, and environmental dynamism. Individual variables in this six-factor model are loaded on different factors. Because innovation, proactiveness, and risk-taking are regarded as the three dimensions of EO, this study averages items into the construct for EO and views the three dimensions as separate indicators. The CFA provides an acceptable fit for the full measurement model in which EO, value, rareness, performance, competitive advantage, and environmental dynamism are all included ( χ2 (214) = 365.06, GFI=0.86, CFI=0.97, NFI=0.94, RMSEA=0.059). Figure 4-1 illustrates the results. ROA and Tobin’s q are not included in the measurement model because construct validity of these two measures has been considered by prior literature. In addition, it is also widely acknowledged that ROA and Tobin’s q has been defined and calculated by academic literature and TEJ.
96
Table 4-4 Analysis of Measurement Model
Model Items of measurement models GFI AGFI NFI CFI RMSEA Model 1:
Entrepreneurial Orientation
Innovation: 3 items Proactiveness: 3 items Risk-taking: 3 items
0.95 0.90 0.97 0.98 0.077
Model 2:
Resource attributes Value: 4 items
Rareness: 3 items0.96 0.92 0.98 0.99 0.075
Model 3:
Firm performance Satisfaction: 5 items
Competitive advantage: 3 items
0.97 0.94 0.98 0.99 0.057
Model 4:
Environmental dynamism
Environmental dynamism: 5 items
0.99 0.93 0.98 0.99 0.081
97
χ2 (214) = 365.06; GFI=0.86; CFI=0.97; NFI=0.94; RMSEA=0.059 Figure 4-1: Full Measurement Model in a CFA
98
4.3 Comparison for EO Measurement Models
To confirm whether EO measurement model fits the data well, this study compares the efficacy of several alternative models (Aryee, Budhwar, & Chen, 2002). In addition to a first-order factor measurement model, this study may consider a second-order factor as a form of aggregation. Aggregation is useful since it can represent the relationship between variables more parsimoniously. For example, several observed variables can be represented by a single first-order latent variable, and a second-order factor encompasses the meaning of several first-order latent variables.
According to the theoretical model of EO, a second-order measurement model is developed, and EO is viewed as a second-order factor, which is measured by three first-order latent variables, including risk-taking, innovation and proactiveness (Bhuian et al., 2005). In order to identify the validity of first- and second-order measurement models, alternative models are examined in the CFA, and the patterns are shown in Figure 4-2.
99
Model 1: One first-order factor
Model 3: Three correlated first-order
factors
Model 2: Three uncorrelated first-order factors
Model 4: Second first-order factors
Figure 4-2: Alternative EO Models in CFAs
Model 1, a first-order factor, hypothesizes that one factor is measured by all items. This means that a uni-dimensional construct refers to the fact that EO can explain all the common variance among the 9 items in this model. Model 2 supposes that all 9 items form three uncorrelated first-order factors, namely, innovation, proactiveness, and risk-taking. The correlation between three first-order factors exists in Model 3, and these factors account for the covariance among 9 items. Model 4 hypotheses that all 9 items form three first-order
100
factors and these three factors are measured by a second-order factor (EO).
These four alternative models are examined by using SEM technology, and the results of the models are shown in Table 4-5. Models 1 and 2 are not reasonable because their fit indices do not achieve the threshold criteria. The results of both Models 3 and 4 are acceptable because all their fit indices meet the threshold criteria.
Table 4-5 Alternative EO Measurement Models
Model
Construct dimension χ2 df (χ2 /
df ) GFI AGFI NFI CFI RMSEA Model
1
EO (One first-order
factor) 491.95 27 24.40 0.65 0.41 0.75 0.76 0.293 Model
2
Three uncorrelated
first-order factors 200.71 26 8.35 0.82 0.68 0.87 0.88 0.183 Model
3
Three correlated
first-order factors 50.20 23 2.08 0.95 0.90 0.97 0.98 0.077 Model
4 EO (Second order) 50.20 23 2.08 0.95 0.90 0.97 0.98 0.077
101
4.4 Assessment of Model Fit and Path Significance
A three correlated first-order factors model analysis is used to examine the proposed hypotheses. The study averages items into dimensions for innovation, proactiveness, and risk-taking, and treats the dimensions as separate indicators of their corresponding construct (EO) in the SEM analyses. The study first tests the fully mediated model, the results of which are presented in Figure 4-3. The fit indices for this model are adequate: χ2 (60) =93.61;
GFI=0.93; AGFI=0.90; CFI=0.99; NFI=0.98; RMSEA=0.055. EO is found to be positively related to the value of resource-capability combinations (β= 0.28 for value, p ＜ 0.001) and is also positively related to the rareness of resource-capability combinations (β= 0.31 for rareness, p ＜ 0.001). Therefore, Hypotheses 1 and 2 are supported. In Hypotheses 3 and 4 (the association between resource attributes and competitive advantage), the value of resource-capability combinations is positively associated with a firm’s competitive advantage (β= 0.49, p ＜ 0.001), and the rareness of resource-capability combinations also shows a positive association (β= 0.46, p ＜ 0.001). Thus, Hypotheses 3 and 4 are thus supported.
With respect to the mediated effect of resource attributes (H5), the mediated model is found to be preferred, thus supporting Hypotheses 5a and 5b. The results of our analyses strongly support the mediated model proposed in this study. This means that the influence of EO on a firm’s competitive advantage is channeled through its attributes, specifically, value and rareness, of resource-capability combinations.
102
χ2 (60) = 93.61; GFI=0.93; CFI=0.99; NFI=0.98; RMSEA=0.055 Figure 4-3: Structural Model: Results of the SEM Model with CA (n = 201)
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01.
Figure 4-4 shows the results of the fully mediated model when indicators of performance are derived from ROA and Tobin’s q. The fit indices for this model are adequate:
χ2 (97) =194.62; GFI=0.90; AGFI=0.86; CFI=0.96; NFI=0.94; RMSEA=0.071. When examining all the hypotheses, the variety of the path coefficients is observed. First, EO has a positive effect on the value of resource-capability combinations (β= 0.27, p ＜ 0.001).
Second, EO has a positive effect on the rareness of resource-capability combinations (β= 0.31,
p ＜ 0.001). Therefore, Hypotheses 1 and 2 are supported. Third, the value of
resource-capability combinations is not positively associated with ROA and Tobin’s q. Finally, the rareness of resource-capability combinations is not related to ROA and Tobin’s q.Therefore, Hypotheses 3 and 4 are not supported.
103
χ2 (97) = 194.62; GFI=0.90; CFI=0.96; NFI=0.94; RMSEA=0.071 Figure 4-4: Structural Model: Results of the SEM Model with ROA and TQ (n = 201)
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01.
Figure 4-5 shows the results of the fully mediated model when indicators of performance are derived from firm performance. The fit indices for this model are adequate:
χ2 (85) =211.42; GFI=0.92; CFI=0.98; NFI=0.96; RMSEA=0.052. All the hypotheses are examined again, and the results are shown as follows. EO has a positive and significant influence on the value of resource-capability combinations (β= 0.30, p ＜ 0.001) and the rareness of resource-capability combinations (β= 0.34, p ＜ 0.001). In addition, the value and rareness are also positively associated with satisfaction. Therefore, Hypotheses 1, 2, 3, and 4 are supported. With respect to the mediated effect of resource attributes (H5), the mediated model is found to be preferred, thus supporting Hypotheses 5a and 5b.
104
χ2 (85) = 211.42; GFI=0.90; CFI=0.94; NFI=0.92; RMSEA=0.087 Figure 4-5: Structural Model: Results of the SEM Model with Satisfaction (n = 201)
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01.
105
4.5 High and Low Levels of Environmental Dynamism Models
This study examines whether the levels of environmental dynamism can influence the relationships between EO and the value/rareness resource-capability combinations, when facing a dynamic environment. This study splits the sample into two groups: high levels of environmental dynamism (n=95) and low levels of environmental dynamism (n=106).
Hypotheses 6a and 6b state that a dynamic environment strengthens the relationship between entrepreneurial orientation and the value of resource-capability combinations and strengthens the relationship between entrepreneurial orientation and the rareness of resource-capability combinations. To validate this hypothesis, the mediating models are tested in two different groups and with different indicators of firm performance (i.e., CA, ROA, TQ, and performance)
When firm performance is a competitive advantage, Figures 4-6 and 4-7 show the structural models with path coefficients. The results of Figure 4-6 show that EO has a positive and significant influence on the value (β= 0.41, p ＜ 0.001) and the rareness (β=
0.55, p ＜ 0.001) in high levels of environmental dynamism. However, the results of Figure 4-7 show that EO has no significant impact on the value (β= 0.15, p ＞ 0.1) and rareness (β=
0.15, p ＞ 0.1) in the low level of environmental dynamism. The comparison of the two models shows following results. First, EO has a stronger effect on value in the high level of environmental dynamism than the low level. Second, EO has a stronger effect on rareness in the high level of environmental dynamism than in low level.. Therefore, this study supports Hypotheses 6a and 6b.
106
χ2 (60) = 103.22; GFI=0.86; CFI=0.98; NFI=0.95; RMSEA=0.089
Figure 4-6: Structural Model: Results of the SEM Model with CA in High Levels of Environmental Dynamism
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01. (n=95)
χ2 (60) = 62.9; GFI=0.92; CFI=1.00; NFI=0.96; RMSEA=0.025
Figure 4-7: Structural Model: Results of the SEM Model with CA in Low Levels of Environmental Dynamism
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01. (n=106)
107
When firm performance is ROA and TQ, Figures 4-8 and 4-9 present the structural models with path coefficients. The results of Figure 4-8 show that EO has a positive and significant influence on value (β= 0.40, p ＜ 0.001) and rareness (β= 0.53, p ＜ 0.001) in a high level of environmental dynamism. The results of Figure 4-9 show that EO has no influence on value (β= 0.14, p ＞ 0.1) and rareness (β= 0.14, p ＞ 0.1) in a low level of environmental dynamism. A comparison of the two models shows the following results. First, the influence of EO on value in the two different groups is different. Second, this study finds that EO has a stronger effect on rareness in high levels of environmental dynamism than in low levels. Therefore, Hypotheses 6a and 6b are supported again.
χ2 (97) = 161.90; GFI=0.84; CFI=0.95; NFI=0.91; RMSEA=0.082
Figure 4-8: Structural Model: Results of the SEM Model with ROA and TQ in High Levels of Environmental Dynamism
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01. (n=95)
108
χ2 (97) = 161.90; GFI=0.86; CFI=0.96; NFI=0.91; RMSEA=0.070
Figure 4-9 Structural Model: Results of the SEM Model with ROA and TQ in Low Levels of Environmental Dynamism
Note: Standardized factor loadings and path coefficients are presented. The estimates of t-value are reported in parentheses. * P < 0.10, ** P < 0.05, *** P < 0.01. (n=106)
When firm performance is satisfaction, structural models with path coefficients are
When firm performance is satisfaction, structural models with path coefficients are