CHAPTER 5 QUALITATIVE COMPARATIVE ANALYSIS
6.5 ANALYSIS STRATEGY
The research project adopted three separate statistical analyses briefly outlined in table 8. Table 8 below shows the relationship of the analyses to each other and the phases for each separate analysis. The methods are described in more detail later for each analytical phase. Principal components factor analyses were used for preliminary analysis to investigate the appropriateness of the measures for the main analyses. This was conducted to ensure that the measures used are valid and reliable. The first set of analyses uses hierarchical cluster analysis in order to answer the research question, 1a Are there specific configurations that are different from each other?
Hierarchical cluster analyses are useful to explore whether groups exist in the data that are different from each other on key attributes, followed by K-means cluster analysis to validate the groups with a guide from the research framework for the domains being investigated (Ketchen Jr. & Shook, 1996).
They are subsequently assessed with an ANOVA to determine statistically distinct and significant groups based on the cluster analysis (This is presented in the appendix table 27). It is important to ascertain whether there are different types of venture for the configuration analysis. Secondly, the configuration analysis, which is the main focus of the study, used fuzzy sets Qualitative Comparative Analysis (fsQCA), which is a mixture of quantitative and qualitative analysis methods to answer the research questions (Kuckertz, Berger, & Allmendinger 2015, Deutscher, Zapkau, Schwens, Baum and Kabst, 2016 and Kuckertz, Berger, and Mpeqa 2016). As the ventures are at different stages and phases of development entering the survey, previous studies by van Gelderen et al., (2005) have found that one way to control for this effect is to assess the entrepreneurs in a subsequent year. At this point the nascent entrepreneurs would have all had one year of attempting to start their venture, hence wave 2, year 2006, is used for the initial wave. Wave 6, 2011, is used for the analyses as Reynolds et al. (2007) have found that it takes approximately seven years to start a new venture (refer to table 9).
Other design options were also considered, such as the true longitudinal design and the true experimental design. However, this was not possible for two reasons. Firstly, software developments using QCA with a time dimension were only in their infancy at the time of conducting the research, which limited the option for full longitudinal analysis. An article by Hak, Jaspers, and Dul (2013) considered identifying temporal sequences of gestation activities to study configuration changes over time. They have outlined how this could be achieved. The option to study each wave of data was considered but the research focus was on finding interactions based on early starters versus later starters and considering causal differences based on configurations. The performance differences from the initial conception to a realistic time that most ventures would be expected to have started based on prior research findings was one of the foci (Bruderl & Schussler, 1990; Reynolds, 1997). The design by Peltoniemi et al., (2014) was considered, although in their research, the focus was on considering survival which is not the current research focus and the configuration changes from
wave to wave were not as pronounced in their study which limited the findings. Other methods were considered for the current study such as the use of longitudinal growth models which was helpful in explaining and accounting for changes over time in the study by Samuelsson and Davidsson (2009). They examined the process differences between innovative and imitative start-ups with the application of longitudinal growth modelling (LGM). The LGM makes full use of the longitudinal data in relation to both independent variables and dependent variables being evaluated at several different points in time. The method is useful for assessing growth over time, but is limited in its assessment of interactions of variables leading to performance from configurations. From the literature review, methods such as correlation based methods such as regression analyses were considered for the current analysis, however, it was found to be inappropriate as a result of limited ability to explain the effects of interactions between the domains and the outcome. It does not account for the causal complexity of the domains (Ragin, 2008, Fiss, 2011). The other limitation is based on the sample size. When the size is reduced beyond a point it makes it difficult to assess the effects accurately (Aiken et al., 1991; Byrne & Ragin, 2009), hence the current fsQCA method was chosen (refer to appendix C p. 304 for the detailed steps of the fsQCA for this study). Thirdly, the last analysis phase investigated the sensitivity of results in order to determine the validity of the fsQCA analyses in explaining the outcome. The analyses consider whether the method used was robust and can be cross-validated to enable explanations of findings.
Table 10 Outline of the Analysis strategy Wave 1 Year 2005/6 Wave 2 Year 2006/7 Wave 6 Year 2010/11 Method Descriptive Statistics Validate Measures using Principle Components Analysis
Validate Measures using Principle Components Analysis
Phase 1 Cluster Analyses
Phase 2 Fuzzy Sets Qualitative Comparative Analysis Fuzzy Sets Qualitative Comparative Analysis Phase 3 Sensitivity Analyses Sensitivity Analyses