Confirmatory factor analysis

Harman’s (1970) single-factor test using exploratory factor analysis in SPSS, was used to examine the likelihood of common method bias, because it can be an issue in questionnaire-based studies (Podsakoff et al., 2003). Harman’s single factor test is one of the most widely used techniques to address the issue (Podsakoff et al., 2003). The unrotated factor analysis results show that one-factor solution does not explain the majority of the variance, suggesting that there is no common method bias affecting this study (Harman, 1970). The analysis was continued with a confirmatory factor analysis (CFA) which is commonly used to model the relationship between an underlying construct and the observed variables (Steenkamp &

Baumgartner, 1998). CFA is usually used when the structure of the underlying constructs is assumed based on a theory (Byrne, 2010). The factor loading presents the change in an item caused by a unit change in a latent construct (Steenkamp & Baumgartner, 1998). Based on the results, one item (V10) from the customer relationship performance scale was deleted due to a low (<0.5) factor loading. All other items got a rather good factor loading, ranging from 0.557 to 0.947. To evaluate the internal consistency of the factors, the level of Cronbach’s alpha was measured for each factor. Alpha levels of 0.5-0.6 are considered satisfactory yet levels greater

56

than 0.7 are preferable as they indicate high internal consistency (Hair, Black, Babin, &

Anderson, 2010). Cronbach’s alpha in all constructs is at a good level, ranging from 0.833 to 0.956, which ensures the reliability of constructs. Factor loadings and Cronbach’s alphas of the re-specified model are illustrated in Table 5.

Table 5. Measure items and reliability statistics

Unstandardized

factor loadings Standardized factor loadings Big data usage (Cronbach’s alpha= 0.956)

We use big data to develop customer profiles. 1.000 0.902

We use big data to segment markets. 1.132 0.947

We use big data to assess customer retention behavior. 1,000 0.885 We use big data to identify appropriate channels to reach customers. 0.942 0.854

We use big data to customize offers. 0.963 0.849

We use big data to identify our best customers. 1.053 0.878

We use big data to assess the lifetime value of our customers. 0.723 0.758 Customer relationship performance (Cronbach’s alpha= 0.833)

Achievement of customer satisfaction 1.000 0.692

Retaining present customers 1.532 0.943

Customer structure (e.g. stable customer relationships) 1.172 0.793 Quality of the products and services (e.g. greater customer benefit) 0.823 0.557

Firm financial performance (Cronbach’s alpha= 0.885)

Growth in sales 1.000 0.828

Realization of profits 0.896 0.715

Achievement of the target market share 0.946 0.882

Realization of return on investment 0.827 0.713

At first, the model fit figures showed a model fit that was poorer than the acceptable measures.

A correlation of error terms is a way in which model fit can be enhanced (Hooper, Goughlan &

Mullen, 2008). It is easier to justify a decision to correlate within-factor error than to correlate errors across latent variables (Hooper, Goughlan & Mullen, 2008). A strong theoretical justification is needed if error terms are allowed to correlate but it can be acceptable if the researcher finds support for their decision (Jöreskog & Long, 1993; Hooper, Goughlan &

Mullen, 2008). For this model, two within-factor errors were correlated in big data usage variable and two within customer relationship performance variable. In big data usage the items were “We use big data to identify our best customers” and “We use big data to assess the lifetime value of our customers” which are theoretically closely related, since the assessment of lifetime value is commonly used to determine which customers are best for the firm and thus the correlation can be justified. In firm financial performance the items were “Realization of return on investment” and “Realization of profits”, which again can be justified to correlate since these items are often closely related to each other because profits often derive from

57

investments. Even though these items are related to each other, they measure different things.

Such as weight and height of people often correlate, the measurements do not make another irrelevant, yet correlation of error terms can be allowed. The re-specified confirmatory factor analysis model is represented in figure 4.

Figure 4. Confirmatory factor analysis

The re-specified measurement model provides good fit since it indicated acceptable goodness-of-fit-measures (Table 6). Values of NFI (Normal Fit Index) range between 0.0 and 1.0 and values closer to 1.0 indicate good model fit. NFI in this model exceeds the 0.9 limit, which has been suggested to be necessary for establishing adequate fit (Hair et al. 2010). The Comparative Fit Index (CFI), a revised form of the NFI, performs well even when sample size is small (Tabachnick & Fidell, 2007) because it takes the sample size into consideration (Byrne, 1998).

A value greater than 0.90 is needed also in CFI in order to ensure that wrongly defined models are not accepted (Hu & Bentler, 1999). In fact, a value of CFI greater than 0.95 is presently recommended to ensure a good fit (Hu & Bentler, 1999). In light of these recommendations, this analysis indicates great model fit as the CFI is 0.960.

The RMSEA figure shows how well the model, with unknown but optimally chosen parameter estimates, would fit the covariance matrix of the population in question (Byrne, 1998). Today, an upper limit of 0.07 is generally considered to be adequate for RMSEA among researchers (Steiger, 2007). Despite that, multiple researchers argue that a RMSEA below 0.08 shows a good fit (e.g. Hooper, Goughlan & Mullen, 2008). The RMSEA was 0.075 in this model, indicating a reasonable fit, especially taking into account that the sample size for this study was rather small.

58 Table 6. Goodness of fit statistics

RFI NFI IFI TLI CFI RMSEA

0.901 0.920 0.960 0.951 0.960 0.075

In this study, CFA correlation matrix (Farrell, 2010) was used to evaluate discriminant validity, that is whether the measured constructs are truly unrelated. Hair et al. (2010) state that to ensure discriminant validity, the square root of average variance extracted (AVE) for each construct should be compared to the correlation between every construct and the AVE-values should be greater than the correlations measured. In this study, all AVE values exceeded the correlations between constructs. Furthermore, composite reliability values that were calculated to factors all were greater than the accepted limit of 0.80 (0.840-0.956) which further verifies validity of the measurement model. The results indicate that discriminant validity exists between constructs, and the model does not have validity concerns. The AVE values and correlations between constructs with the composite reliability figures are presented in Table 7.

Table 7. AVE values and squared correlations of the measurement model

Composite reliability AVE BDU CRP FFP

BDU 0,956 0,756 0,869

CRP 0,840 0,577 0,159 0,759

FFP 0,867 0,621 0,263 0,559 0,788