Validation of Mass Customization Capability as Second-Order Construct

To establish the dimensional structure of strategic capabilities for mass customization, we specified various alternative measurement models at the first-order and second-order levels and assessed their relative fits (Law et al. 1998). The fit statistics for these models are reported in Table 6.

Models χ2 (df) Normed χ2 RMSEA TLI CFI

Strategic capabilities

Model 1 (one-factor model) 166.053 (35) 4.744 0.181 0.470 0.588

Model 2 (3 uncorrelated factors) 51.786 (35) 1.480 0.065 0.932 0.947

Model 3 (3 correlated factors) 38.197 (32) 1.194 0.041 0.973 0.981

Model 4 (one second-order factor) 38.197 (32) 1.194 0.041 0.973 0.981

Performance measures

Model 5 (one-factor model) 66.960 (14) 4.783 0.182 0.689 0.793

Model 6 (two-factor model) 20.128 (13) 1.548 0.069 0.955 0.972

Table 6: Measurement Models and Fit Statistics

Model 1 has a unidimensional factor that accounts for the variance among all 10 items, which is also known as Harman’s single-factor test (Podsakoff et al. 2003, p. 889). Not surprisingly, model 1 has a very poor fit. In model 2, we conceptualize that the 10 items form three distinct and uncorrelated first-order factors, corresponding to SSD, RPD, and CN. Comparison of the fit indices for model 1 and model 2 shows that model 2 is the better-fitting model, indicating that a multidimensional model composed of three uncorrelated first-order factors is superior to a unidimensional first-order factor model. The chi-square difference (Δχ2 = 114.267, p < 0.01) across the two models is significant, providing further evidence in support of model 2.

Model 3 conceptualizes that the three first-order factors are free to correlate with each other. A comparison between the fit measures of models 2 and 3 indicates that model 3 represents the data considerably better than model 2; the chi-square difference between the two models relative to their degrees-of-freedom difference is also significant (Δχ2 = 13.589, Δdf = 3, p < 0.05). Moreover, we examined additional models that are similar to model 3 because they have the 10 items forming two, four, and five correlated first-order factors. As model 3 exhibited a better fit than these additional models, they are not described in further detail. Finally, model 4 posits mass customization capability (MCC) as a reflective second-order construct that accounts for the relationships between the three strategic capabilities. However, when two nested models have exactly the same chi-square and degrees of freedom, as do models 3 and 4, comparing goodness of fit statistics for the two models is not meaningful. In this case it is also not possible to calculate the target coefficient, which is the percent of variation in the first-order factors that can be explained by the second-order construct (Marsh and Hocevar 1985). The superiority of one model is instead established by examining the significance of the second-order factor loadings in the measurement model (Venkatraman 1990; Tippins and Sohi 2003) on one hand and significance of the structural links that link the measurement model to a criterion variable of interest such as company performance on the other (Venkatraman 1990). All the second-order factor loadings in model 4 are significant (p < 0.05). Further, as will be discussed in Section 6.3, only the second-order factor model has a significant impact on company performance. Collectively, these results suggest that the second-order factor structure is a better statistical specification for modeling mass customization capability, supporting our hypotheses H1, H2, and H3 of mass customization capability encompassing SSD, RPD, and CN.

The dimensional structure for the performance measures was assessed in a similar manner by comparing two measurement models. Model 5 consists of seven measures forming a unidimensional factor, whereas model 6 consists of two distinct yet correlated factors representing market growth (MG) and customer success (CS). Table 3 presents the fit statistics for these two models. We retain model 6 because of its superior fit to the data; its fit measures surpass the fit statistics associated with model 5 and exceed the critical cut-off values. The chi- square difference across the two models was also found to be significant relative to the corresponding change in degrees of freedom (Δχ2 _{= 46.832, Δdf = 1, p < 0.01).}

To further assess the content validity of the second-order mass customization capability construct (MCC second-order) comprising three strategic capabilities, we compare it with a predefined direct measure of mass customization capability (MCC direct). For this purpose, the adapted MCC measure from Tu et al. (2001) is used as a possible criterion. A positive and significant path coefficient between the two measures would suggest that the indirect measurement of MCC through SSD, RPD, and CN is a valid representation of the direct MCC measure. The corresponding model is shown in Figure 18. Within the model, we conducted confirmatory factor analysis (CFA) for the direct MCC measure. MCC1, MCC2, and MCC5 were removed from further analysis because they showed indicator reliabilities below the recommended value of 0.4. The resulting factor reliability was 0.67 and the average variance extracted 0.50, and thus above the cut-off values of 0.6 and 0.5, respectively (see 8.2.9).

SSD CN RPD MCC second- order MCC direct SSD2 SSD3 SSD4 SSD5 RPD1 RPD2 RPD3 CN1 CN3 CN4 0.866*** 0.520*** 0.411** 0.606*** MCC3 MCC4

χ2/df = 1.368, RMSEA = 0.057, TLI = 0.938, CFI = 0.953 _{***p < .001}**p < .01

R2_{= 0.749}

Figure 18: Relatedness of Strategic Capabilities to Mass Customization Capability

Overall, estimation of the model produced a good fit (χ2/df = 1.368, RMSEA = 0.057, TLI = 0.938, CFI = 0.953). In line with our reasoning, the path coefficient between the indirect and the direct measurement of MCC is positive and highly significant (0.866, p < 0.001). Moreover, the MCC second-order construct explains 75% (R2 = 0.749) of the variance in the direct MCC measure, meaning that it captures the major facets of observable mass customization capability. Finally, according to Fornell and Larcker’s (1981) test, the discriminant validity between the latent construct and the direct measurement of MCC is not

sufficient (1.49). That implies that the MCC second-order construct measures the same content as the intended direct measurement of MCC.