Section 5. Data Analysis Techniques

Data analysis can take the form of simple descriptive statistics or more sophisticated statistical inference, which are among those shown in Figure 3.

Figure 3. Type of Research, General Research Approaches, Data Collection Techniques, & Data Analysis Techniques

Data analysis techniques include univariate analysis (such as analysis of single-variable distributions), bivariate analysis, and more generally, multivariate analysis. Multivariate analysis, broadly speaking, refers to all statistical methods that simultaneously analyze multiple measurements on each individual or object under investigation (Hair et al., 1995); as such, many multivariate techniques are extensions of univariate and bivariate analysis. The diagram presented below as Figure 4 proposes an approach to decide when a specific type of data analysis technique is appropriate. Each data analysis technique is later defined in the glossary (Section 7).

Figure 4. Decision Tree for Choosing Best Form of Data Analysis (Adapted from Hair et al., 1995)

It should be noted that when selecting a data analysis technique, a researcher should make sure that the assumptions related to the technique are satisfied (i.e., normal distribution, independence among observations, linearity, and lack of multi-collinearity between the independent variables, etc.). Structured Equation Modeling, SEM, can be applied as a preferred substitute for many of the techniques in the diagram, in some cases providing additional statistics and examinations. For a detailed discussion refer to Gefen, Straub, and Boudreau (2000) and to Gefen (2003).

Examining interaction effects with LISREL when the interaction involves a continuous variable is problematic because of high degrees of shared variance. In such cases, PLS should be used. For a discussion of applying interaction effects in PLS, see Chin et al. (2003). When the interaction involves a nominal item, such as gender or group number, LISREL can be applied but measuring the interaction requires comparing samples. See discussion in Bollen (1989).

SEM are typically applied with reflective measures. On this, please see the discussion in Gefen, Straub, and Boudreau (2000) and to Gefen (2003). PLS is especially adept to handling formative measures (indicators) in addition to reflective measures (indicators). Although formative measures are, in theory, not applicable in LISREL Gefen, Straub, and Boudreau (2000), recent research (Diamantopoulos and Winklhofer, 2001) suggests that it might be possible to apply formative measures also in LISREL. Still such attempts are extremely rare in the literature and in the experience of the authors here seldom show appropriate fit indexes.