Measurement Model of Formative Indicators
To evaluate ? quality of ? measurement model, ? design of constructs (Diamantopoulos and Winklhofer 2001) and ? relevance of indicators (Chin 1998a) need to be analyzed. According to Diamantopoulos/Winklhofer and Chin, there are five serious issues determining ? quality of ? measurement model: (1) content specification, (2) indicator specification, (3) indicator reliability, (4) indicator co linearity, and (5) external validity. Content specification consists of defining ? scope of ? latent constructs to be measured.
In particular, “? breadth of definition is extremely significant to causal indicators” (Nunnally and Bernstein 1994, p. 484) ? research model presented in this research paper includes four latent constructs measured with formative indicators: financial hazard, performance hazard, strategic hazard and psychosocial hazard. These constructs were precisely defined and their domain intensively discussed (see section 4. 2), ensuring ? proper specification of ? applicable content of all ? constructs deployed. Indicator specification comprises ? identification and definition of indicators which constitute ? latent constructs.
? aggregation of all formative indicators forms ? specification of ? formatively measured construct. So, indicator specification is particularly significant for models using formative indicators. ? indicators used in this model were identified by intensive literature review and have been validated through many pre-tests with senior bank managers who were knowledgeable about ? topic of this research. ? indicators were presented in section 4. 2; all items were measured on ? seven-point Likert scale. Indicator reliability analyzes ? importance of each individual indicator that forms ? relevant construct.
Two quantitative arguments have to be accounted for: (1) ? sign of ? indicator needs to be correct as hypothesized and (2) ? weighting of ? indicator should be at least 0. 2 as proposed by Chin (1998b). ? model tested shows correct signs for all indicators used and all (except two) have ? weight of at least 0. 2 and are (with three exceptions) at least significant at ? 0. 05 level. As formative measurement models are based on linear equation systems, substantial indicator co linearity would affect ? stability of indicator coefficients.
Neither ? analysis of correlations of indicators nor ? calculation of variance inflation factors (all indicators fall far below ? threshold of 10 as suggested by (Cohen 2003) necessitated ? rejection of any indicators used. So, all indicators could be retained as no redundancy was identified. External validity aims at ensuring that all indicators which form ? construct are actually included in ? model. Following Diamantopoulos and Winklhofer (2001), external validity can be analyzed by creating ? phantom construct which is measured using reflective indicators.
If ? formatively measured construct strongly and significantly correlates with ? reflective measured construct, external validity is given. ? correlations of constructs within ? tested model were all strong and significant at ? 0. 001 level. Therefore, it is shown that ? formative indicators used in this research actually form their respective constructs. Measurement Model of Reflective Indicators ? quality of ? measurement model is determined by (1) convergent validity, (2) construct reliability, and (3) discriminant validity (Bagozzi 1979; Churchill 1979; Peter 1981).
Convergent validity (Bagozzi and Phillips 1982) (Peter 1981) Indicator reliability can be examined by looking at ? construct loadings. In ? model tested, all loadings are significant at ? 0. 001 levels and above ? recommended 0. 7 parameter value (importance tests were conducted using ? bootstrap routine with 500 resample’s (Chin 1998b)). ? estimated indices were all above ? threshold of 0. 6 (Bagozzi and Yi 1988). Discriminate validity of ? utilized indicators can be analyzed by looking at ? average variance extracted (AVE). ? calculated figures are all above ? recommended threshold of 0. 5 (Chin 1998b).
Discriminate validity can also be assessed by checking ? cross-loadings. These are obtained by correlating ? component score of each latent variable with both its respective block of indicators and all other items that are included in ? model (Chin 1998b). Structural Model ? adequacy of indicators in ? measurement model enables one to evaluate ? explanatory power of ? entire model as well as ? predictive power of ? independent variables. ? explanatory power is examined by looking at ? squared multiple correlations (R^sup 2^) of ? dependent variables. As can be inferred from Figure 2, 46% (R^sup 2^=0.
46) of ? variation in perceived hazard is explained by performance, financial, strategic and psychosocial hazard. Furthermore, 37% of ? variation in ? attitude towards outsourcing is explained by perceived hazard. ? R^sup 2^ value for ? intention to increase ? level of BPO (R^sup 2^=0. 56) is also encouragingly high. Predictive power is tested by examining ? magnitude of ? standardized parameter estimates between constructs together with ? corresponding t-values that indicate ? level of importance. All path coefficients exceed ? 0. 2 level except for psychosocial hazard.
In particular, bootstrapping revealed strong importance (at ? 0. 005 level) of ? dependent variables except for psychosocial hazard (at ? 0. 1 level) Analysis of ? overall effect size (f^sup 2^) of ? antecedents of perceived hazard reveals that all hazard facets have moderate effect except for psychosocial hazard (low effect). Though, small f^sup 2^ scores do not necessarily imply an unimportant effect (Cohen 1988). Therefore, all hypotheses have been proven to be correct, including that regarding ? effect of psychosocial hazard on perceived hazard. Figure 2 depicts ? findings graphically.