MASTER THESIS Rev: final
Complementarity and inter-firm absorptive capacity:
The mediating role of relationship‟s characteristics
Sunil I. Vaswani
Complementarity and inter-firm absorptive capacity:
The mediating role of relationship‟s characteristics
Master Thesis
BySunil I. Vaswani
Student Number: S2244179 Email: s.vaswani@student.rug.nl
Course: MSc. Business Administration (Marketing Research) Faculty of Economics and Business
University of Groningen The Netherlands
First supervisor: Prof. P. Leeflang Second supervisor: Drs. J. Berger
3
ACKNOWLEDGMENT
I am extremely grateful to my first supervisor Prof. P. Leeflang for granting me the permission for working on an analytically inclined thesis. I appreciate your critical feedback and efforts to facilitate the entire process of combining my master thesis along with the PhD. research of my second supervisor.
To my second supervisor Drs. J. Berger, thank you for the guidance and support provided by you from the beginning of my work. It made my job easier. I could not have hoped for a better guide.
I would also like to extend my sincere gratitude to the teaching and non-teaching staff of the marketing department at University of Groningen and BI Norwegian Business School for their dedication to the noble field of education. Their depth of knowledge and willingness to be there for students made the double masters academically enriching.
To my friends in Oslo and Groningen, their support and friendship made it easy to integrate and perform in a challenging environment.
Management summary
Although complementarity amongst firms in a vertical relationship is a necessary condition for an alliance to be successful and develop inter-firm absorptive capacity (ACAP), its effect can vary from relationship to relationship based on the characteristics of the partnership. This study successfully tries to demonstrate the multidimensional nature of inter-firm ACAP and analyzes the mediating role of relationship characteristics like compatibility, connectedness, idiosyncraticity and nature of co-ordination mechanism on the link between complementarity and the two constructs of inter-firm ACAP.
The hypotheses were tested applying partial least squares – structural equation modeling method to survey data from 151 Dutch executives (buyers) from knowledge intensive industries like automotive, machinery, chemicals, pharmaceuticals and electronics.
This study focuses on quantitative analysis technique and confirms that all the mediators partially mediate the relationship between complementarity and realized absorptive capacity (RACAP) while only connectedness, idiosyncraticty and relational norms had an effect on the relationship between complementarity and potential absorptive capacity (PACAP). Compatibility and contractual norms did not show any significant effect on PACAP.
5
Table of Contents
Acknowledgement……… 3 Management Summary……… 4 1 Introduction 1.1 Background………... 7 1.2 Relevance……….. 8 1.3 Aim……….. 9 2 Theoretical background 2.1 Frame work and Hypotheses……….. 102.1.1 Concept of Inter-firm Absorptive Capacity……….... 10
2.1.2 Complementarity as an Antecedent to Inter-firm ACAP…………. 11
2.1.3 Mediating role of relationship’s characteristics……….. 12
2.1.3.1 Compatibility……… 12
2.1.3.2 Connectedness……….. 13
2.1.3.3 Idiosyncraticity………. 13
2.1.3.4 Contracting……….. 14
2.1.3.5 Relational Norms……….. 14
2.1.4 Explorative and Exploitative learning……… 15
2.2 Conceptual Model………. 16
3 Methodology 3.1 Data collection procedure……… 17
3.1.1 Method………... 17
3.1.2 Instrument………... 18
3.1.3 Variables………... 18
3.2 Data processing procedure………. 23
3.3 Analysis Technique……….. 24
3.3.1 Structural Equation Modelling………. 24
3.3.2 Partial Least Squares………. 25
3.4 Analysis Strategy……….. 28 3.4.1 Choice of algorithm……… 28 3.4.2 Resampling method……….. 29 3.4.3 Hierarchical analysis……….. 30 3.4.4 Mediation analysis………. 31 3.4.5 Missing values……….. 31 3.4.6 Outliers………. 32 4 Model Assessment 4.1 Descriptive Statistics……… 34 4.2 Measurement Model………. 35 4.2.1 Indicator reliability………. 36 4.2.2 Construct reliability……… 39 4.2.3 Convergent validity……… 41 4.2.4 Discriminant validity……….. 42 4.3 Structural Model……… 44 5 Results 5.1 Model……….. 46
5.2 Non Linearity Analysis………. 48
5.3 Results Overview……….. 52
5.4 Effect Size………. 55
5.5 Ad-Hoc analysis……… 56
6 Discussion 6.1 Findings and Implications……… 58
7 Conclusion 7.1 Summary……… 61
7.2 Limitations and Future Research……… 61
References 63 Appendix 1: Reliability analysis for construct – Connectedness….. 69
7
1
INTRODUCTION
This thesis is a subset of the research being carried out by Drs. J. Berger of the University of Groningen. The background research, hypotheses development and data collection were carried out by him and made available for this thesis. The main focus of this report is the statistical analysis of the collected data.
The report is presented over seven chapters. In chapter 2, the hypotheses and conceptual model is introduced making use of theory. In chapter 3, the research methodology is presented, where details and rationale regarding various decisions taken during the data collection and analysis are explained. In chapter 4 and 5, we first assess the model in terms of its validity and reliability and then present the results of the empirical research. These results and its implications are discussed in chapter 6. In the last chapter we conclude with summary, limitations, and directions for future research.
But first let us briefly look into the background of the research area, its academic and managerial relevance and our aim.
1.1
Background
Research in the field of strategy has shifted its focus from external sources of profit and recognized the importance of a resource-based view (RBV), where the firm‟s resources and capabilities are considered as its main source of competitive advantage (Grant, 2010). According to Matusik & Hill (1998) a firm‟s capability to gain, manage and exploit knowledge is one such critical resource. This is because effective knowledge management can drive innovation, which in turn can provide a significant competitive advantage (Cohen and Levinthal, 1990).
(ACAP) which is a bundle of knowledge-based capabilities (Zahra & George, 2002) is considered as the most significant determinant of its knowledge transfer capabilities by many studies (e.g., Lane and Lubatkin, 1998; Gupta and Govindarajan, 2000).
Thus ACAP as a resource satisfies Barney‟s (1991) criteria of being competitively advantageous. It is valuable, is an unobservable process so is difficult to copy, and is non-substitutable because it moves along a learning curve (Jap, 1999).
This RBV approach focusing on ACAP is also applicable to an exchange between firms, (Palmatier et al., 2007). According to Dyer and Singh (1998) and Jap (1999) when partners combine exchange or invest in idiosyncratic assets, knowledge, and resources/capabilities, and /or they employ effective governance mechanisms it generates superior performance. As such, we presuppose that next to the partners‟ prior knowledge, the interaction patterns and coordination efforts transpiring across organizational boundaries, along with dedicated investments are the mediators of inter-firm ACAP, which create the differential advantages that the partners in the relationship strive for.
1.2
Relevance
Managers can only develop and implement performance enhancing strategy, when they understand what drives the relationships performance. In this study we develop a multidimensional construct of interfirm ACAP, which is specified for vertical relationships between independent firms operating at successive stages in the production chain. Explorative learning refers to recognizing, acquiring and making sense out of external knowledge, and it corresponds to the notion of potential absorptive capacity (PACAP) (Zahra & George, 2002). Exploitative learning relates to transforming and applying acquired knowledge, and it reflects the concept of realized absorptive capacity (RACAP) (Zahra & George, 2002).
9 Lubatkin, 1998; Kale et al., 2000; Lane et al., 2001; Selnes & Sallis, 2003), or only takes a part of the construct into consideration (e.g., Johnson et al., 2004).
1.3
Aim
The first aim of this study is to analyze the mediating effect of relationship‟s characteristics on the link between complementarity between partner firms and all the dimensions of inter-firm ACAP. Our point here is that the complementarity is a necessary but insufficient condition. Characteristics of relationship like compatibility, connectedness, idiosyncraticity, and existing co-ordination mechanisms may influence ACAP in different ways, depending on which characteristic is being analyzed. For example existence of legal contract between both partners may limit the acquisition and assimilation of new knowledge (PACAP) (Van den Bosch et al., 1999), but may facilitate in transformation and exploitation (RACAP) of the knowledge exchanged (Jansen et al., 2005).
2
THEORETICAL BACKGROUND
The main focus of this thesis is the analysis of the collected data; hence it does not include a classical literature review and as a consequence the theoretical backing of the relevant concepts is fairly limited. In this chapter the hypotheses are introduced using the underlying relevant theoretical concepts involved in the overall research. This is followed presenting the conceptual model under consideration.
2.1
Frame Work and Hypotheses
2.1.1 Concept of Inter-firm Absorptive Capacity
As touched upon briefly in the introduction, the absorptive capacity of a firm is a dynamic organizational capability and source of competitive advantage over rival organizations. Our focus here is not the individual, but the inter-firm ACAP between two organizations in a vertical buyer – seller relationship. We define it based on Zahra and George (2002) as a set of routines and procedures by which a corporate relationship identify and obtain important external information (acquire), make sense of the externally gathered data (assimilate), integrate the newly acquired and already existing knowledge (transform) and refine , extend or leverage this knowledge (exploit).
In order to understand the interrelation of four dimensions of inter-firm ACAP can be compartmentalized into two groups viz. potential absorptive capacity (PACAP) and realized absorptive capacity (RACAP) (Zahra & George, 2002). PACAP consists of acquisition and assimilation and RACAP consists of transformation and exploitation.
11
2.1.2 Complementarity as an Antecedent to Inter-firm ACAP
Useful prior knowledge forms the core of a firm‟s ACAP. This needs to be supplemented by combining with external resources of knowledge gained through alliances with other firms (Zahra & George, 2002). This is because “the ultimate purpose of alliances is to leverage firm resources along with the complementary resources of partner firms to create synergy effects” (Lin et. al.,2009). Synergy is not just addition of resources, it said to be created when the combined resources of the partners give them both an extra edge that they individually did not possess i.e. when 1+1>2. (Lin et. al.,2009). This is further explained by Sarkar et. al. (2001) that; when partners bring in unique and valuable strengths and resources not only does the performance of the project for which alliance is formed increase but also the learning aspects of the alliance helps the partners to individually incorporate the new knowledge for other ventures.
This interactive learning potential that help firm add to their existing capabilities and know how are likely to be greater when there is diversity and non-redundancy in the knowledge base of the two partners (Sarkar et. al., 2001).
Thus for two firms to be engaged in a relationship, one of the critical reasons is the complementarity between the alliance partners (Kale et. al, 2000). It refers to existence of related but dissimilar core competencies of the partners. The extent of overlap between the partners is inversely proportional to the complementarity between them. The higher the complementarity, the more diverse resources it brings into a relationship which can be valuable for both partners (Kale et. al, 2000). The chances of partnerships to succeed are much higher, as there is mutual benefit and a potential for each firm to learn from its partner (Larsson & Finkelstein, 1999).
2.1.3 Mediating role of relationship’s characteristics
Complementarity of resources indicates potential to gain knowledge. But is only potential enough? What are the factors that cause this potential to be realized enough? Can the same potential be realized to different extent by different firms ? To understand the process of inter-firm ACAP and gain more clarity we check the mediating effect of relationship characteristics on the link between knowledge potential (complementarity) and ACAP ( PACAP and RACAP).
2.1.3.1 Compatibility
Geringer (1988) advocates that partner compatibility and alliance success are correlated. Compatibility between partnering firms refers to the chemistry, similarity and understanding between them, i.e. how smooth is the relationship between them. Along with complementarity, they both together form the knowledge context of inter-organizational relationship.
The difference between complementarity and compatibility is that complementarity refers to the differences in useful resources between the two organizations in the relationship which forms the basis of potential learning opportunities. Compatibility between the two firms in the relationship on the other hand provides the basis for actually capitalizing on this knowledge potential (Kale et al., 2000).
Thus we propose that complementarity is a necessary but insufficient condition for building a relationship‟s ACAP and compatibility between the partners affects the ability to influence the actual benefit derived from the existing source of external knowledge, thereby mediating the relationship between complementarity and ACAP. Partners can demonstrate compatibility in different aspects of businesses like operating strategy, corporate cultures, management styles, nationality and firm size (Parkhe, 1993) hence the mediation effects both PACAP and RACAP in a positive way.
Hypothesis 1a: Compatibility positively mediates the relationship between complementarity
and PACAP.
Hypothesis 1b: Compatibility positively mediates the relationship between complementarity
13
2.1.3.2 Connectedness
Connectedness refers to the structure of social interactions between firms, i.e. how easy, smooth and convenient are the interactions between them (Sheremata, 2000). According to Jaworski and Kohli (1993) it facilitates knowledge exchange. It provides common platform for transferring new thoughts and integrating them with previous knowledge (Hansen 2002).
Complementarity provides the source of different external knowledge, but only when there is high connectedness between firms is the interaction and exchange of information is facilitated. Thus, connectedness affects the ability and motivation of firms to integrate and recombine different source of external knowledge (Jansen et. al., 2009), thereby positively mediating the relationship between complementarity and inter-firm ACAP
However, when the connectedness crosses a certain high threshold, the dependency on the partner increases and reduces the ability to spot other external sources of knowledge. Therefore, it can be expected that the positive relationship between connectedness and PACAP will not be linear and will cease to exist after the threshold and we expect a non- linear degressive curve.
Hypothesis 2a: Connectedness positively mediates the relationship between complementarity
and PACAP.
Hypothesis 2b: Connectedness positively mediates the relationship between complementarity
and RACAP.
2.1.3.3 Idiosyncraticity
increase with a trusted partner. The idiosyncratic resources themselves make it easier to actually transform and implement the knowledge which are obtained from the relationship. Therefore, it is expected that a positive relationship exists between idiosyncratic resources and RACAP.
Hypothesis 3a: Idiosyncraticity positively mediates the relationship between complementarity
and PACAP.
Hypothesis 3b: Idiosyncraticity positively mediates the relationship between complementarity
and RACAP.
2.1.3.4 Contracting
One of the generic type of co-ordination mechanisms between partners is the formal contracting. Contracting refers to establishing authority and equivalent of organizational hierarchy in a relationship (Stinchcombe, 1985). In such a case the acquisition of knowledge is on a market level competitive basis (Haour, 1992). The flow of knowledge gets restricted to only what is stated in the contract, because additional knowledge has an inherent value (it costs money) that the owner would like to exploit. The contracting firm does not get full access or control over the knowledge asset which is in place (Fey, 2005). Hence, it is expected that a negative relation exists between contracting and PACAP. However, contracting focuses on the output of a vendor. It sets guidelines and frameworks for achieving results, so the information which is shared is easily integrated and implemented. Hence contracting is expected that a positive relation between contracting and RACAP exists.
Hypothesis 4a: Contracting negatively mediates the relationship between complementarity
and PACAP.
Hypothesis 4b: Contracting positively mediates the relationship between complementarity
and RACAP.
2.1.3.5 Relational Norms
15 solidarity between organizations involved in informal relationships reduces barriers to information exchange (Zahra & George, 2002). Hence there are more opportunities to obtain information as they encourage interaction between different members of the group and it is expected that informal relational norms will positively influence PACAP. Also with no restrictions on the support to be provided to partners and the solidarity existing between them it is expected that the knowledge that is transferred is actually utilized. Therefore, it can be expected that a positive relationship exists between relational norms and RACAP.
Hypothesis 5a: Informal relational norms positively mediates the relationship between
complementarity and PACAP.
Hypothesis 5b: Informal relational norms positively mediates the relationship between
complementarity and RACAP.
Even though contracting and relational norms are different forms of governance they both have a positive effect on exploitation and transformation (RACAP). This is contradicting, when seen in isolation, however when seen in conjunction with PACAP it makes sense. In contracting, only specific knowledge agreed between two partners legally is made available. The exploitation of this limited knowledge is also part of the contract. Whereas in case of informal relation, a wide resource base is available and whatever help is required by partners to exploit the knowledge is provided.
2.1.4 Explorative and Exploitative learning
16
Hypothesis 6: PACAP has a positive association with explorative learning performance. Hypothesis 7: RACAP has a positive association with exploitative learning performance. Hypothesis 8: Explorative learning performance has a positive association with exploitative
learning performance.
2.2
Conceptual Model
The twelve research hypotheses presented in the previous section are depicted by the conceptual model presented in Figure 1. The model shows all the constructs and the hypothesized relationships between them.
Figure 1: Conceptual model with hypotheses
17
3
METHODOLOGY
In this chapter, the method, instrument and variables used in the data collection procedure are introduced. It also gives a brief insight into how this data is processed. Since analysis is the core of this report, the chapter provides the details of the chosen statistical method and the software used for data analysis. It also explains in details the choices made during the use of the selected software.
3.1
Data collection procedure
13.1.1 Method
The study focuses on analyzing vertical relationship between independent firms operating at successive stages in the production chain. The relationship is real but not directly observable. Such an abstract phenomenon presents a large potential for measurement error (Selnes and Sallis, 2003). A common measurement practice in such cases is to use key informants from both sides of the dyad (John & Reve, 1982; Lane and Lubatkin, 1998; Selnes and Sallis, 2003).
Firms operating within the Dutch automotive/machinery, chemicals/pharmaceuticals, and semiconductors/electronics industry were approached for participation. The choice of industry was influenced by prior work reporting different knowledge strategies in these industries (Lichtenthaler & Ernst, 2007).
Heads of purchase or a high-ranking technology manager were asked to supply contact data of four people within their companies who were central to their customer relationships. To overcome selection bias, the four relationships were selected following a 2 x 2 design, with relationship duration greater or less than two years on one dimension, and average or crucial importance of products or components on the other dimension (Johnson et al., 2004). After the buyer informants were recruited, they were asked to supply the names of their contacts in the supplier organization. The researcher then contacted the identified informant in the supplier organization and asked him or her to participate in the study. Both customer and supplier informants were then interviewed, using a standardized questionnaire.
3.1.2 Instrument
The survey was conducted using the „Inter-firm absorptive capacity survey – 2010‟. The 5 section survey had a total 97 individual sub questions (indicators) grouped under 16 overall questions.
For this thesis the environmental dynamism and customer diversity data collected in section III was not used. Part of section IV focusing on exploring, who between the buyer and seller benefitted more from the relationship was also omitted for this thesis. Thus limiting the total number of indicators used in this research to 75.
Section I of the questionnaire comprised one question seeking general information using 7 indicators (e.g. length of relationship, industry, respondent designation etc.). Section II consisted of 23 indicators, across five questions about explorative and exploitative capabilities of the relationship. 12 out of the 24 Section IV indicators we use refer to the explorative and exploitative learning performance of the relationship. The last section, Section V captures 33 indicators over six questions specific to the influencers of inter-firm absorptive capacity (ACAP).
The actual survey was made available in Dutch and English to prevent any respondent bias arising from preference for either of the languages. See Appendix 2 for the full English version.
3.1.3 Variables
Our research model presented in the previous chapter composed of 17 latent constructs. They differ in type (reflective and formative), order (first and second), function (independent, dependent and mediator) and number of indicators used to measure them. This is shown in figure 2.
19 Latent constructs that are estimated using directly observed measurement indicators are called first order constructs. Second order constructs are those that are measured using the first order constructs (Chin, 1998a). As suggested by Chin and Gopal (1995), second order construct can be either molecular (arrows pointing to its first order constructs) or they can be molar (arrows pointing from its respective first order constructs).
Let us look at the constructs and their indicators in more detail.
Legend:
Order of variable – Type of variable – Number of indicators R – reflective F - formative
Figure 2: Variables used in model.
In the model complementarity, connectedness, idiosyncratic, contracting, flexibility, information exchange, solidarity, assimilation and exploitation are all first order reflective constructs. The first order formative constructs are compatibility, acquisition, transformation, explorative and exploitative learning process.
The constructs of PACAP (acquisition and assimilation), RACAP (transformation and exploitation) are second order molar constructs and relational norm is a second order molecular construct with flexibility, info exchange and solidarity as its first order factors.
Apart from the dimensions of ACAP the study uses existing scales from value chain alliance, inter-firm and intra-firm learning literature to measure the latent constructs. Appropriate scales for dimensions of ACAP were not available and were specifically generated for this study by Berger (2013). Table 1 presents the measurement scales. Respondents are asked to answer on a seven point likert scale, ranging from strongly disagree to strongly agree (e.g. Jap, 1999). Likert scales, ranging from “strongly disagree” (1) to “strongly agree” (7) are used for all questions, except for the items measuring explorative learning process (Lane, Salk & Lyles, 2001), which ranges from “to no extent” to “to a great extent”.
Table 1: Measurement items
Construct Item Source
Complementarity
(COMPL) Var 10
We both contribute different resources to the relationship that help us achieve mutual goals.
We have complementary strengths that are useful to our relationship.
We each have separate abilities that, when combined together, enable us to achieve goals beyond our individual reach.
(Jap, 1999) (Lambe et al, 2002)
Compatibility
(COMPA) Var 11
The organizational cultures of the two partners are compatible with each other. The management and operating styles of the partners are compatible with each other. In this relationship, a clear strategic fit between the partners‟ objectives is evident.
(Kale, Singh & Perlmutter, 2000) (Geringer, 1988)
Connectedness
(CONEC) Var 12
In the relationship, there is ample opportunity for informal “hall talk” among employees. In the relationship, employees from different departments (partner‟s departments included) feel comfortable calling each other when the need arises.
In this relationship, it is easy to talk with virtually anyone you need to, regardless of rank or position.
In the relationship, managers discourage employees from discussing work- related matters with those who are not their immediate superiors or subordinates. (Rev C) People around here are quite accessible to each other.
21 Idiosyncratic
(IDIO) Var 13
Both of us have created capabilities that are unique to this alliance.
Together we have developed a lot of knowledge that is tailored to our relationship. Together we have invested a great deal in building up our joint business. Both of us have made a great deal of investments in this relationship.
If this relationship were to end, we both would be wasting a lot of knowledge that is tailored to our relationship.
If either company were to switch to another partner, we both would lose a lot of investments made in the present relationship.
(Lambe et al, 2002)
Contracting
(CTRAC) Var 14
We do not have specific, well detailed formal agreements with our partner. (Rev C) Rules and procedures take a central role in the relationship.
In the relationship, partners‟ compliance to the formal contract is monitored extensively. Written procedures regulate virtually all aspects regarding the parties‟ tasks and influences in our joint operations.
Formal agreements stipulate all aspects concerning the exchange of information between both partners.
In the relationship, formal agreements are more important than informal ones.
(Deshpande & Zaltman, 1982) (Cannon & Perreault, 1999) (Buvik & Reve, 2002) Flexibility (FLEX) Var 15 (1-3)
Flexibility in response to requests for changes is a characteristic of this relationship. The parties expect to be able to make adjustments in the ongoing relationship to cope with changing circumstances.
When some unexpected situation arises, the parties would rather work out a new deal than hold each other to the original terms.
(Heide & John, 1992)
Information Exchange
(INFO EX) Var 15 (4-7)
In this relationship, it is expected that any information that might help the other party will be provided to them.
Exchange of information in this relationship takes place frequently and informally, and not only according to a pre-specified agreement.
It is expected that the parties will provide proprietary information if it can help the other party.
It is expected that we keep each other informed about events or changes that may affect the other party
(Heide & John, 1992)
Solidarity
(SOLID) Var 15 (8-10)
Problems that arise in the course of this relationship are treated by the parties as joint rather than individual responsibilities.
The parties are committed to improvements that may benefit the relationship as a whole, and not only the individual parties.
The parties in this relationship do not mind owing each other favors.
(Heide & John, 1992)
Acquisition
(ACQUIS) Var 1
In the relationship, data on the state-of-the-art of relevant external technologies are accessible when needed.
Relevant changes in the industrial environment catch us (the joint parties in the relationship) by surprise. (Rev C)
In the relationship, technological advancements that are critical to our operations are identified before they enter our markets.
In the relationship, our companies put a lot of effort into acquiring new external knowledge.
In the relationship, we have relevant, continuous and up-to-date information on our current and potential competitors.
Assimilation
(ASSIM) Var 2
In the relationship, we are not sufficiently capable of processing newly acquired information on new technologies and innovations, which are useful or have proven potential. (Rev C)
New developments are well understood due to the shared interpretation efforts of members of our companies.
In the relationship, we quickly and professionally analyze and interpret changing market demands.
We effectively make use of our joint employees‟ level of knowledge, experience and competencies to analyze and interpret external information.
The atmosphere in the relationship stimulates productive discussion encompassing a variety of opinions. (Jansen et al. 2005) (Camison & Fores , 2010) Transformation (TRANSF) Var 3
In the relationship, we easily integrate newly acquired external knowledge into our common understanding of business-related affairs.
For us it is not easy to see the connections among the pieces of external knowledge held by different members of our companies. (Rev C)
In the relationship, an apparent incongruity between existing knowledge and newly acquired knowledge leads to ambiguity. (Rev C)
In the relationship, we are able to transcend traditional mind-sets if newly acquired knowledge demands it.
(Bontis et al., 2002) (Cadiz et al., 2009)
Var 4
Employees of both partners get acquainted well enough to know who knows what. Employees of both companies know where critical expertise resides within the relationship.
In the relationship, employees actively exchange newly acquired knowledge.
In the relationship, it is well known to which employee, or group of employees, specific new knowledge needs to be transferred.
Exploitation
(EXPL) Var 5
In the relationship, both companies constantly discuss how to better exploit knowledge. Our customers can immediately benefit from new technological knowledge learned in the relationship.
New knowledge is easily integrated into our common operations.
The relationship‟s capacity to use and exploit new knowledge allows us to respond quickly to changes in the environment.
We are proficient in transforming new technological knowledge into product and process patents (Jansen et al. 2005) (Cadiz et al., 2009) (Camison & Fores , 2010) Explorative learning process (E‟PLOR LP) Var 8 (1-5)
Joint creation of new manufacturing and production expertise? Joint creation of new product development expertise? Joint creation of new technological expertise? Joint creation of new marketing expertise? Joint creation of new managerial expertise?
(Lane, Salk & Lyles, 2001)
Exploitative learning process
(E‟PLOI LP) Var 9 (1-7)
The relationship with the other company has resulted in lower logistics costs.
Flexibility to handle unforeseen fluctuations in demand has been improved because of the relationship.
The relationship with the other company has resulted in better product quality.
Synergies in joint sales and marketing efforts have been achieved because of the relationship.
The relationship has a positive effect on our ability to develop successful new products. The relationship helps us to detect changes in end-user needs and preferences before our competitors do.
Investments of resources in the relationship, such as time and money, have paid off very well.
(Selnes & Sallis, 2003)
Legend:
Name of variable – (Acronym of variable) – Corresponding number on the questionnaire
23
3.2
Data processing procedure
At first, preliminary checks after data collection were conducted to confirm that the organizations contacted for the research were in fact production companies from our chosen industry or not. Responses from different sector or services companies were excluded. The current research focuses on the buyer‟s perspective of the vertical relationship; hence for further analysis all the 151 buyer responses were extracted from the total usable responses. 2
Next, based on the aims and characteristics of research and limitations of the dataset the statistical method to be applied for analysis is chosen. This is followed by carefully considering the software to be used because all the different commercially available statistical packages have unique strength and weaknesses that make them appropriate for use in a particular research and inappropriate for the use in others. Once the software to be used is finalized the various analysis strategies to operationalize the study are decided.
However, before the actual analysis, several preparatory procedures are executed. The first precautionary check is to scan for problems in the dataset caused by missing values and outliers. This is because high missing values render the conclusions regarding data fit inappropriate (MacCallum & Browne, 1993). If the missing values are random in nature and below the allowable threshold of 10%, they are substituted by the column average. Outliers caused by measurement error distort the result and can be deleted but a non-erroneous outlier on the other hand gives new insights, hence outlier data points should be carefully considered. Since an adequate sample size is necessary for any statistical result to be valid; an adequacy check on the final sample size after applying the data corrections is carried out. Indicators that refer to the same construct but are coded in the opposite direction will have negative correlation and will lead to improper internal consistency check. Hence care should be taken to identify the reverse coded indicators. In our research 6 indicators (Var 1.2, Var 2.1, Var 3.2, Var 3.3, Var 12.4 and Var 14.1) are recoded in order to align them.
The next step is the actual analysis. Here the quality of the research model is evaluated. Since our study includes latent construct that are measured using multiple indicators we follow the stepwise guidelines formulated by Gotz et al. (2009) for evaluation of the measurement and structural model. Based on the results obtained, inferences are made regarding the proposed hypotheses.
3.3
Analysis Technique
3.3.1 Structural Equation Modeling
Structural Equation Modeling (SEM) is a second generation data analysis technique that combines factor analysis and multiple regression analysis to simultaneously answer a set of researchers interrelated questions on a particular field of interest (Hair et al., 2010). In other words SEM has the ability to test a hypothesized model that depicts relationships between several independent and dependent variables in one combined model (Schumacker & Lomax, 2004; Bagozzi & Fornell, 1982; Gefen et al., 2000).
For researchers another advantageous characteristic of SEM is that it allows usage of latent constructs – unobserved concept, in the analysis (Hair et al., 2010). Unlike first generation techniques that focus on directly measured variable and ignore the measurement error, SEM reduces the concept measurement error and improves statistical estimation of the relationship between unobserved concepts (Hair et al., 2010).
Since we are using unobserved latent variables and are testing multiple hypotheses involving these latent constructs in one model, SEM is the method of choice for us.
However SEM techniques are of two types, covariance-based (CBSEM) which are analyzed using popular software like LISREL, M-plus etc., or the variance-based Partial Least Squares (PLS) analysis. As per Chin (2010) both these analysis methods are complementary to each other.
25
3.3.2 Partial Least Squares
PLS is a variance based SEM analysis technique developed by Wold (1975) to estimate the parameters of a structural model for cases where the restrictive assumptions linked with using CBSEM are not fulfilled (Fornell & Bookstein, 1982).
The basic structure of model involves the outer measurement part which links the latent construct to its measurement indicators (manifest variables) and the inner structural part which links latent constructs to each other is the same as CBSEM. The difference lies in estimating algorithm and use of resampling techniques which do not require parametric assumptions to be met (Lohmoller, 1989).
PLS disintegrates the total model into smaller blocks and solves them simultaneously (Chin, 1998) and can be compared to running several multiple regressions (Hair et al., 2010). CBSEM on the other hand is a full information approach. It solves the model as one unit and mistakes from one part tend to influence the other part. Therefore strong theoretical base is required to avoid misspecification, and this makes it inappropriate for use in exploratory studies (Chin & Newsted, 1999).
According to Gefen et al., (2000) other requirements imposed by CBSEM include constraints regarding distribution (normal), sample size (more than 100) and type of variable used (reflective). Garza (2011) summarizes several differences between CBSEM and PLS SEM in a table. Its extract is presented in Table 2 below.
Table 2: Comparison of Partial Least Squares and Covariance-Based SEM (adapted from Chin & Newsted, 1999, Gefen et al., 2000 & Garza 2011).
Criterion PLS CBSEM
Approach: Variance-based Covariance-based
Basis : Prediction oriented Parameter (covariance – fit ) oriented
Assumptions: Predictor specification
(nonparametric).Robust to deviations from a multivariate distribution
Multivariate normal
distribution and independent observations (parametric)
Parameter estimates Consistent as indicators and sample
size increase
Construct scores Explicitly estimated by weighting sums of underlying items
Indeterminate.
Latent variables: Both formative and reflective can
be used
Only reflective indicators can be used
Higher order model Can accomplish either molar or
molecular model.
Limited to 2nd order molecular model
Model complexity: Can handle high complexity Can handle moderate complexity
Model specification and interpretation
Easy – only requires specifying measurement items and construct relations
Additional considerations like measurement scale adequacy, model identification etc. need to be
addressed. Sample size
requirements
Smaller in comparison Larger in comparison
Nature: Supports exploratory and
confirmatory research
Confirmatory - strong theoretical base required.
As can be seen from above, due to its flexibility there are cases where PLS SEM technique is better suited. According to Ringle, C., Sarstedt, M., & Straub, D. (2012) the reasons presented by researchers for using PLS is:
Sample size is small and data is non-normal.
Formative measures are to be used.
Model is complex with large number of latent variables and indicators.
Research is exploratory in nature and theory has to be developed.
Predictions are to be made.
Categorical variables are to be used.
Our model also consists of formative latent constructs operationalized at first and second order. Thus we confirm choice of PLS based SEM for our analysis.
3.3.3 Software
Application of PLS path modeling in marketing and management/organizational research has gathered momentum recently due to increase in the required use of formative constructs. This has led to substantial improvement in the number and quality of software solutions at a researcher‟s disposal e.g. PLS – GUI, Visual PLS, PLS-Graph, Smart PLS, SPAD – PLS, Warp PLS (Temme et. al. 2010).
27 WarpPLS is ideal for our analysis because it has a number of unique features which the other PLS path modeling software do not support. Based on Knock (2010), these differences can be summarized as follows:
(1) Ability to fit both linear and non-linear SEM models and provide graphical output.
In our research we have hypothesized positive and negative mediation, however we expect some relations between constructs for e.g. connectedness and PACAP to be non-linear. Only WarpPLS is able to estimate coefficients of association taking nonlinear relationships (U-curve and S-(U-curve) between latent variables into consideration. All the other available software can estimate only non-linearity caused due to consideration of moderating effects. WarpPLS also provides visual 'scatter plots' for depicting the fit of the regression lines to the data, making it easier to comprehend.
(2) Calculation of indices like f 2 and Q2 in addition to R2
All available software calculate the value of R2 for the endogenous latent variables to assess the structural model quality. WarpPLS however also calculates Stone-Geisser Q2 coefficients (Geisser, 1975; Stone, 1974) - a nonparametric measure calculated via resampling for the assessing the predictive validity and Cohen‟s (1988) f 2 - effect size coefficients for all path coefficients. It determines practical effect (small, medium, or large) of statistically significant path coefficients and is particularly useful when dealing with a large dataset.
(3) Calculation of variance inflation factor (VIF) scores for all latent variable
(4) Calculation and output of p-values for path coefficients
WarpPLS shows p-values for all weights and loadings to check for significance. This makes it user friendly in model assessment. The other software however calculate only t-values and requires the researcher to manually check the significance using statistical tables.
(5) Ease of dealing with outliers, mediators and moderators and performing hierarchical analysis.
WarpPLS is very user friendly and makes analysis of a complex model easy. WarpPLS allows the option of ranking the data and restricting the range of indicator value in order to deal with outliers.
Mediation analysis can also be carried out using the already available direct and indirect effect analysis or applying the traditional Baron and Kenny (1986) method easily. In our research we are testing the multiple parallel mediation effect of a relationship‟s characteristics on the link between complementarity and dimensions of interfirm ACAP, so it is a very important feature. It also allows the user to define a reflective or formative latent variable as a moderator.
Hierarchical analyses can be conducted easily by adding standardized latent scores of first order variables to the dataset as new set of indicators. In our research we have the constructs of relational norms, PACAP and RACAP operationalized as a second order variable, hence the use of WarpPLS is justified.
In the next section we look at the analysis strategy applied while using WarpPLS.
3.4
Analysis Strategy
3.4.1 Choice of algorithm
29 Our analysis uses Warp3 PLS Regression algorithm. This algorithm first tries to find an S-Curve (function whose first derivative is a U-curve) relationship between latent variables. If identified, the algorithm transforms (or “warps”) the scores of the predictor latent variables to reflect the S- curve relationship in the estimated path coefficients in the model (Knock, 2012). If not identified it tries to fit a U-curve, and if still required a straight line in the end (Knock, 2012).
It is the most advanced of the regression algorithm which actually encompasses two of the remaining three choices that the software provides. It is also the default option of the software.
The two other regression based options are the Warp2 PLS Regression algorithm that first tries to identify a U-curve relationship between latent variables followed by a straight line and The PLS Regression algorithm, that only tries to find a linear relation between constructs and does not perform any warping (Knock, 2012).
The warping takes place after the estimation of all weights and loadings in the model i.e. during the estimation of path coefficients. Primarily while calculating weights, all the three algorithms consider the factor score as exact linear combination (zero error) of its indicators.
The final fourth option is a non-regression based Robust Path Analysis algorithm, where factor scores are calculated by averaging all of the indicators associated with a latent variable; and the P values are calculated through resampling (Knock, 2012).
3.4.2 Resampling method
PLS modeling does not constrain the researcher with distributional assumption for the data. So the p-values generated for the parameter estimates (β- values) which are based on normal theory are unsuitable and have to be generated using resampling procedures (Temme et.al. 2010).
The bootstrapping technique creates a user determined number of re-samples. The sample size of the re-samples is the same as the original dataset but there can be repetition of the rows (“re-sampling with replacement”) because the rows are selected at random from the original dataset (Knock, 2012). It tends to generate more stable resample path coefficients with samples larger than 100 (Nevitt & Hancock, 2001) and with samples where the data points are evenly distributed on a scatter plot.
Jackknifing, on the other hand, creates a number of resamples that is equal to the original sample size. One row from the original sample is deleted for each resample therefore the individual sample size of each resample is one less than the original sample size (Knock, 2012). Jackknifing tends to generate more stable resample path coefficients with small sample sizes lower than 100, and with samples containing outliers (see, e.g., Chiquoine & Hjalmarsson, 2009).
Blindfolding creates a user determined number of resamples that contain a certain number of rows replaced with the means of the respective columns (Knock, 2012). This number equals the sample size divided by the number of resamples decided by user. Blindfolding tends to perform somewhere in between jackknifing and bootstrapping (Knock, 2012).
Since our sample size is 151 we select bootstrapping. The default number of resamples for bootstrapping is 100. Even setting the number of resamples at 50 is considered enough for a fairly reliable P value estimates (Efron et al., 2004) but conversely, increasing the number of resamples well beyond 100 leads in complexity that may make the software run very slowly. Hence we select 100 resamples.
3.4.3 Hierarchical analysis.
31
3.4.4 Mediation analysis.
Mediation analysis can be carried out by using the indirect and total effects calculated by WarpPLS (Knock, 2012) or simply applying techniques such as the causal steps approach popularized by Baron and Kenny (1986)
The software output aggregates the indirect effects for paths between two latent variables with same number of segments (Knock, 2012). In our analysis we need to check mediation effects of 5 variables on the link between complementarity – PACAP and complementarity – RACAP, so the software will provide a combined analysis of the five effects. In order to check out single mediator effects we will have to isolate the effects by modeling 5 individual models. This is a very lengthy procedure and hence we use Baron and Kenny‟s (1986) approach, and adapt it to test for multiple mediating effects between one independent latent variable (X) and a dependent latent variable (Y). The advantage is that we do not need to separately isolate the effect of each mediator (M).
The procedure outlined in Knock (2011), requires two models to be built and analysed using WarpPLS. The first model should only have X pointing at Y and the other model should have X pointing at Y, X pointing at M, and M pointing at Y.
The three criteria‟s for confirming mediation require, the path between X and Y to be significant in first model and the path between X and M plus M and Y to be significant in the second model.
In our first model we include complementarity, PACAP, RACAP, explorative learning and exploitative learning and check for direct effects. We then include all the five mediators together in the second model because Preacher & Hayes, (2008a) suggest it is convenient, precise, and parsimonious to do so.
3.4.5 Missing values
Hair et al. (2010) also suggests that there are different methods of handling missing values in a dataset. List-wise deletion requires completely deleting the respondent‟s answers, even if a single question is not answered. Pairwise deletion allows using the non-missing data for the particular respondent and ignoring the missing values. Model based approach or imputation techniques require making suitable assumptions for the missing fields, like replace the missing values with column averages.
The software automatically replaces the missing values with column averages. Since the data is also standardized, it basically means the missing value is replaced as zero by the software. As the values are missing at random this approach is acceptable (Hair et. al., 2010)
3.4.6 Outliers
Outliers are observations that are distinctively different than other observations and should be treated with caution if it is not due to procedural error (Hair et. al., 2010). Once outliers are observed they can be addressed using three methods in WarpPLS i.e. using multiple re-sampling analysis, using ranked data and restricting the range of values that can be used.
Multiple analyses:
Since the warping algorithms are sensitive to the presence of outliers, one can estimate P values with multiple re-sampling techniques and use the P values associated with the most
Table 3: Missing Values
Sr.No. Indicator % Missing Sr.No. Indicator % Missing
1 Var 1.3 (ACQ) 1.99% 10 Var 8.3 (E‟PLOR LP) 0.66%
2 Var 2.1 (ASSIM) 1.32% 11 Var 9.4 (E‟PLOI LP) 0.66%
3 Var 2.2 (ASSIM) 0.66% 12 Var 9.6 (E‟PLOI LP) 0.66%
4 Var 3.4 (TRANSF) 1.32% 13 Var 12.1 (CONEC) 1.32%
5 Var 5.1 (EXPL) 1.32% 14 Var 12.4 (CONEC) 0.66%
6 Var 5.2 (EXPL) 1.32% 15 Var 12.5 (CONEC) 1.32%
7 Var 5.3 (EXPL) 0.66% 16 Var 14.3 (CTRAC) 0.66%
8 Var 5.4 (EXPL) 1.32% 17 Var 15.10 (SOLID) 0.66%
33 stable coefficients (Knock, 2012). Out of bootstrapping and jackknifing which are both complementary resampling methods, jackknifing provides more stable results in case of presence of outliers (Knock, 2012).
Ranked data:
Using this option one can choose to conduct their analyses with only ranked data. When data is automatically ranked, prior to the SEM analysis, it typically eliminates the outliers without decreasing sample size. This is possible due to reduction in value distances that classify an outlier (Knock, 2012).
Even though using ranked data has an effect on collinearity it can be very useful in assessing effects of outliers on path coefficients and respective P values, especially when outliers are not believed to be due to measurement error (Knock, 2012).
.
Range restriction:
The third option is using range restriction option. One needs to select either a standardized or unstandardized indicator as a range restriction variable and set minimum and maximum values. This restricts the analysis to the rows in the dataset within that particular range. Users can remove outliers by restricting the values assumed by a variable to a range that excludes the outliers, without having to modify and re-read a dataset (Knock, 2012).
For our analysis there is no outlier due to procedural error, there are some extreme values, but they are within the range of 4 standard deviations. However we have a few unstable β - values i.e. higher than 0.2 but still insignificant. This can be caused due to outliers (Knock, 2012). The method we use is conducting multiple analyses using second resampling technique of jackknifing. This is done because the sample size is not much greater than 100 and jackknifing tends to generate more stable resample path coefficients with small sample sizes and with samples containing outliers (see, e.g., Chiquoine & Hjalmarsson, 2009).
4
MODEL ASSESSMENT
This chapter begins with the descriptive analysis of the dataset. This is followed by the reliability and validity check of the measurement model. It also provides information regarding the quality of the structural model by providing the details of the determination coefficient (R2) and prediction relevance (Q2).
4.1
Descriptive Statistics
Correlations and cross-correlations between all the indicators are calculated. The correlations between indicators of the same reflective construct are higher than 0.5 for all indicators except Var 12.4 (CONEC4) This variable shows a relatively low correlation (0.01 – 0.133) with the four other indicators of the construct - „Connectedness‟ (Table 4). The removal of this indicator is discussed later in the chapter. Since formative indicators are not supposed to correlate with each other, the correlation among the indicators of the formative latent variables was not checked.
Appropriately no cross correlations were higher than 0.85 (Kline, 2009). The mean and standard deviation for all the measurement indicators is shown in Table 5.
Table 4: Correlation table for items measuring connectedness
Var 12.1 Var 12.2 Var 12.3 Var 12.4 Var 12.5
Var 12.1 1
Var 12.2 0.728 1
Var 12.3 0.577 0.78 1
Var 12.4 0.01 0.071 0.014 1
Var 12.5 0.523 0.654 0.569 0.133 1
35 Table 5: Mean and standard deviation of indicators
Indicator Mean S. Dev Indicator Mean S. Dev Indicator Mean S. Dev
ACQ 1 4.748 1.537 ELOR LP1 3.53 1.868 IDIO 1 4.311 1.541
ACQ 2 4.801 1.571 ELOR LP2 3.788 1.769 IDIO 2 4.358 1.634
ACQ 3 4.318 1.442 ELOR LP3 3.887 1.699 IDIO 3 4.119 1.766
ACQ 4 4.06 1.524 ELOR LP4 2.682 1.498 IDIO 4 4.391 1.575
ACQ 5 4.272 1.62 ELOR LP5 2.45 1.32 IDIO 5 4.212 1.828
ASS 1 4.497 1.603 ELOI LP1 4.179 1.682 IDIO 6 4.397 1.778
ASS 2 4.653 1.296 ELOI LP2 4.815 1.564 CONTR 1 4.119 2.163
ASS 3 4.225 1.396 ELOI LP3 5.013 1.51 CONTR 2 4.132 1.928
ASS 4 4.503 1.57 ELOI LP4 2.88 1.633 CONTR 3 3.667 1.799
ASS 5 5.57 1.339 ELOI LP5 4.205 1.771 CONTR 4 3.351 1.756
TRA 1 4.391 1.346 ELOI LP6 3.5 1.704 CONTR 5 2.921 1.512
TRA 2 3.927 1.357 ELOI LP7 4.834 1.555 CONTR 6 3.066 1.569
TRA 3 4.543 1.603 COMPL 1 5.199 1.27 FLEX 1 5.132 1.32
TRA 4 5.195 1.35 COMPL 2 5.305 1.238 FLEX 2 5.258 1.023
TRA 5 5.152 1.712 COMPL 3 4.98 1.577 FLEX3 5.424 1.186
TRA 6 5.397 1.41 COMPA 1 4.093 1.581 INFOEX1 5.397 1.132
TRA7 4.523 1.595 COMPA 2 3.901 1.769 INFOEX2 5.046 1.655
TRA 8 4.934 1.561 COMPA 3 4.417 1.702 INFOEX3 4.642 1.485
EXP 1 3.839 1.545 CONEC 1 4.215 1.853 INFOEX4 5.675 1.152
EXP 2 4.396 1.553 CONEC 2 4.987 1.693 SOLID 1 5.444 1.181
EXP 3 4.633 1.319 CONEC 3 5.007 1.711 SOLID 2 5.305 1.2
EXP 4 4.745 1.327 CONEC 4 5.473 1.379 SOLID 3 5.32 1.277
EXP 5 4.044 1.316 CONEC 5 5.047 1.511
ACQ - Acquisition ELOR LP - Explorative
Learning process IDIO - Idiosyncratic
ASS - Assimilation ELOI LP - Exploitative
Learning process CONTR - Contracting
TRA - Transformation COMPL - Complementarity FLEX - Flexibility
EXP - Exploitation COMPA - Compatibility INFOEX - Information
Exchange
CONEC - Connectedness SOLID - Solidarity
4.2
Measurement Model
4.2.1 Indicator reliability
Indicator reliability examines how each measuring indicator relates to its overlying latent construct. For indicators measuring reflective constructs the factor loadings from a confirmatory factor analysis (PCA) are used for assessment. Whereas for formative latent variable indicators significant P values associated with indicator weights are used. As per Mathwick et al. (2001) this method indicates the individual indicator‟s contribution to the respective formative construct.
Table 6-a shows the combined loadings (un-rotated) and cross loadings (rotated) for all indicators measuring reflective constructs. Latent variables are listed on top of each column, and indicators at the row level. Each cell refers to strength of an indicator-latent variable link using PCA. The maximum range of the loadings is -1 to 1 as they are from structure matrix and un-rotated (Knock, 2012). For a reflective measurement model to have acceptable indicator reliability the loadings should be greater or equal to 0.5 and its corresponding P value should be lower than 0.05; (Hair et al., 2010).
The factor loadings for all the indicators except Var 12.4; which measures connectedness satisfy this above mentioned criteria. In most cases the loading are even higher than the more stringent criteria of 0.7 (Gotz, Liehr-Gobbers, & Krafft, 2009). However Var 12.4 has a loading of 0.1 and P-value of 0.244 implying the question did not measure the construct as required. Since indicator with loadings of less than 0.4 warrant deletion (Churchill, 1979), removal of Var 12.4 is concluded after checking for construct reliability.
For indicators measuring formative constructs, Table 6-b shows the indicator weights. Each latent variable score is as an exact linear combination of its indicators, the weights are coefficients of multiple regressions between the indicators and the latent variable; all cross-weights are therefore zero (Knock 2012).
37 Indicators of formative latent constructs should not be redundant and should measure different aspects of the same construct. Variance inflation factors (VIF) is a measure to analyze indicator redundancy. Table 6-b also displays VIF values for all the indicators. Most stringent VIF threshold values of lower than 3.3 (Cenfetelli & Bassellier, 2009; Petter et al., 2007) or even 2.5 recommended by the WarpPLS software (Knock, 2012) are used to check for redundancy of indicators.
Since the 25 of the 32 VIF values in the Table 6-b are lower than 2.5 and the remaining 7 lower than 3.3, we can confirm indicator non redundancy for the formative measurement items in this study.
Table 6 - a : Principal component analysis – Reflective variable
COMPL CONEC IDIO CTRAC FLEX INFO
EX SOLID ASSIM EXPL
Continued…
COMPL CONEC IDIO CTRAC FLEX INFO
EX SOLID ASSIM EXPL
REL. NORMS P -Value Var 15.4 -0.071 0.02 0.073 0.118 0.162 0.692 -0.249 -0.038 -0.273 NA <0.001 Var 15.5 0.083 0.063 -0.109 -0.188 -0.079 0.785 -0.104 -0.157 0.022 NA <0.001 Var 15.6 -0.116 -0.055 -0.049 0.018 -0.055 0.803 0.068 0.191 -0.029 NA <0.001 Var 15.7 0.099 -0.024 0.096 0.066 -0.007 0.771 0.258 -0.005 0.253 NA <0.001 Var 15.8 -0.058 -0.101 0.051 -0.018 0.105 -0.009 0.911 -0.046 -0.012 NA <0.001 Var 15.9 0.003 -0.1 0.165 0.131 -0.028 0.164 0.855 0.094 0.079 NA <0.001 Var 15.10 0.063 0.22 -0.233 -0.119 -0.089 -0.164 0.806 -0.048 -0.07 NA <0.001 Var 2.1 0.04 -0.187 0.014 0.061 0.208 -0.412 0.056 0.684 -0.041 NA <0.001 Var 2.2 0.048 -0.007 -0.133 -0.092 0.06 0.162 -0.273 0.787 0.072 NA <0.001 Var 2.3 0.068 0.172 -0.221 0.002 -0.092 0.108 -0.172 0.813 0.242 NA <0.001 Var 2.4 -0.127 0.148 0.268 0.055 -0.137 -0.029 0.057 0.824 -0.155 NA <0.001 Var 2.5 -0.02 -0.165 0.069 -0.021 -0.002 0.117 0.346 0.779 -0.125 NA <0.001 Var 5.1 0.081 -0.131 0.127 0.1 -0.172 0.525 -0.136 -0.289 0.644 NA <0.001 Var 5.2 -0.066 -0.231 0.304 -0.231 0.02 0.004 0.147 0.061 0.722 NA <0.001 Var 5.3 -0.054 0.038 -0.382 -0.052 0.101 -0.067 -0.049 -0.034 0.832 NA <0.001 Var 5.4 0.035 0.091 -0.019 -0.034 0.041 -0.17 0.043 0.132 0.845 NA <0.001 Var 5.5 0.018 0.239 0.055 0.291 -0.037 -0.237 -0.023 0.098 0.597 NA <0.001 Second Order Flex 0 -0.151 -0.12 0.06 NA NA NA NA NA 0.77 <0.001 Info Ex -0.059 0.236 0.141 -0.037 NA NA NA NA NA 0.841 <0.001 Solid 0.057 -0.093 -0.03 -0.017 NA NA NA NA NA 0.878 <0.001
Table 6-b : indicator weights – Formative variable
COMPA ACQUIS TRANSF E’PLOR
39
4.2.2 Construct reliability
Irrespective of individual indicator reliability, it is important to check whether all the construct‟s indicators jointly measure the construct adequately. Testing for construct reliability is the means to confirm this (Gotz et. al., 2009). There are two coefficients of reliability Cronbach‟s α (Cronbach, 1951) and Composite reliability (ρ) (Werts, Linn, & Joreskog, 1974).
Cronbach‟s α is the most widely used coefficient that provides an estimate for the reliability based on the indicator inter-correlations. Cronbach‟s α assumes that all indicators contribute equally to overall reliability and uses equal weights, (Hensler et al., 2009).
Composite reliability (ρ) on the other hand includes the actual factor loadings of individual indicators (Gotz et. al., 2009). Hensler et. al. (2009) recommends the use of composite reliability in PLS path models.
Both reliability coefficients are interpreted in the same way, a value between 0.7 and 0.9 depending on the stage of research is considered satisfactory (Nunnally & Bernstein, 1994), whereas a value below 0.6 indicates a lack of reliability.
Continued...
COMPA ACQUIS TRANSF E’PLOR
As expected from previous analysis, a check on Cronbach‟s α coefficient for the construct „Connectedness‟ reveals that by excluding Var 12.4 the value increases more than 10 % from 0.79 to 0.876 (Annexure 2). This confirms that indicator Var 12.4 is not reliable and it is therefore deleted from the analysis.
Table 7 lists the coefficient of composite reliability and Cronbach‟s α for all the latent reflective constructs. It can be seen that the range of composite reliability score is 0.848 to 0.922 and the range of Cronbach‟s α is 0.76 to 0.897. As it is higher than 0.7, the reliability of reflective constructs in the model is appropriate. For formative constructs these coefficients of composite reliability and Cronbach‟s α cannot be used.
Full collinearity VIFs which enables the identification of vertical and lateral collinearity in a model are also shown for all the latent variables in Table 7.
Table 7: Construct reliability measures Indicator
Original
Indicator
After del. Type
Comp. Rel. Cr. Alpha AVE Full coll. VIF First Order COMPL 3 3 Reflective 0.874 0.783 0.699 1.522 CONEC 5 4 Reflective 0.916 0.876 0.731 2.675 IDIO 6 6 Reflective 0.921 0.897 0.661 2.658 CTRAC 6 6 Reflective 0.888 0.847 0.573 1.321 FLEX 3 3 Reflective 0.864 0.763 0.679 1.676 INFO EX 4 4 Reflective 0.848 0.76 0.583 2.696 SOLID 3 3 Reflective 0.893 0.82 0.737 2.115 ASSIM 5 5 Reflective 0.885 0.837 0.607 3.622 EXPL 5 5 Reflective 0.852 0.781 0.54 2.548 COMPA 3 3 Formative 1.755 ACQUIS 5 5 Formative 2.333 TRANSF 8 8 Formative 4.412 E'PLOR LP 5 5 Formative 1.997 E'PLOI LP 7 7 Formative 2.835 Second Order
REL. NORMS 3 3 Reflective 0.87 0.774 0.691 2.116
PACAP 2 2 Formative 0.919 2.67
