Proposed Theoretical Models, Research Questions and Hypotheses
Study 1: Development and Preliminary Validation of Occupational Stress Scale for Soldiers (OSSS)
4.5 Phase 3: Exploratory Factor Analysis (EFA)
relationship within the data. In other words it would indicate that a third of the variables share too much variance, and hence it becomes impractical and difficult to determine if the variables are correlated with each other or the dependent variable (multicollinearity) (Williams, Onsman, &
Brown, 2010). Hair et al. (1995a) categorized the correlation loadings in the following way: 0.30
= minimal, 0.40 = important, and 0.50 = practical. If the correlation is less than 0.30, then it should be reconsidered (Tabachnick & Fidell, 2001). If the correlation matrix is an identity matrix which signifies that there is no relationship among the items then EFA should not be applied (Kaiser, 1958).
Kaiser- Meyer-Olkin (KMO): Before the extraction of the constructs, there is a test which is conducted to examine the adequacy of the sample and the suitability of data for conducting EFA (Laura & Mazerolle, 2011). The sampling adequacy can be assessed by examining the Kaiser- Meyer-Olkin value (KMO) (Kaiser, 1970). KMO value is used when the cases to variable ratio is less than 1:5. It ranges from 0 to 1. According to Hair et al. (1995a) and Tabachnick & Fidell (2001), value of 0.50 is considered appropriate for EFA. Contrary to this Netemeyer, Bearden &
Sharma (2003) stated that a KMO correlation above 0.60 - 0.70 is considered adequate for analyzing the EFA output. The KMO value indicates sample adequacy. If the value comes in the acceptable range then the researchers can move forward with the next step of analysis (Netemeyer, Bearden & Sharma, 2003).
Factor Extraction: There are several ways to extract factors like principal components analysis (PCA), principal axis factoring (PAF), image factoring, maximum likelihood, alpha factoring,
unweighted least squares, generalized least squares and canonical (Tabachnick & Fidell, 2001;
Thompson, 2007; Costello & Osborne, 2005). Principal components analysis and principal axis factoring are the methods that are most commonly used in research (Tabachnick & Fidell, 2001;
Thompson, 2004; Henson & Roberts, 2006). Thompson (2007) stated that the reason why PCA is mostly used is because it is set as the default method in many statistical software. It is fundamentally used when there is no prior theoretical basis for the model under question (Gorsuch, 1990).
Factor Retention Methods: After the extraction phase is over, the researcher must decide how many constructs should be retained for rotation. Hayton, Allen & Scarpello (2004) state why this decision is important. The utility of exploratory factor analysis depends on being able to differentiate major factors from minor ones (Fabrigar et al.,1999). There is empirical evidence that both under extraction and over extraction are substantial errors that affect results, although specifying too few is traditionally considered more severe. (Velicer, Eaton & Fava, 2000 ; Hayton, Allen & Scarpello, 2004). This also affects EFA efficiency and meaning (Ledesma & Valero- Mora, 2007).
A number of methods are available to assist the researcher in making this decision, but they do not always lead to the same or even similar results (Zwick & Velicer, 1986; Thompson & Daniel, 1996). Most widely used factor retention methods are; Cumulative percent of variance extracted, Kaiser’s criteria (eigenvalue > 1 rule) (Kaiser, 1960), Scree test (Cattell, 1966) and Parallel Analysis (Horn, 1965). Hair et al (1995a) state that in a majority of cases multiple criteria are used.
According to Hair et al (1995a) in cumulative percentage of variance, factors should be retained when at least 95% of the variance is explained. Although in the humanities and social sciences, the explained variance in most cases is generally as low as 50-60% (Hair et al., 1995a; Pett, Lackey
& Sullivan, 2003).
According to the K1 - Kaiser’s method, only constructs which have the eigenvalues greater than one should be retained for interpretation (Kaiser, 1960). This approach is the best known and most widely used (Fabrigar et al., 1999). Another popular method used in this regard is the Cattell’s Scree test which involves the visual exploration of a graphical representation of the eigenvalues for breaks or discontinuities (Cattell, 1966). The number of data points above the break (which does not include the point at which the break occurs) decides the number of factors to be retained.
The break point divides the important or major factors from the minor or trivial factors (Ledesma
& Valero-Mora, 2007). Interpreting Scree plots is very subjective. Therefore, the number of factors to be retained and subsequent results can be different for different researchers (Zwick & Velicer, 1986; Pett et al.,2003). This disagreement and subjectivity can be reduced if the sample size is large, N:p ratios are (>3:1) and communalities values are high (Linn, 1968; Gorsuch, 1983;
Lackey, & Sullivan, 2003). A comparison between Scree test and K1 rule concluded that the Scree test performed better (Zwick & Velicer, 1986), although it was still correct only 57% of the times.
In most cases the problem of overestimate of factors was found (Ledesma & Valero-Mora, 2007).
Still according to Costello & Osborne (2005) scree test is the best choice for researchers.
Selection of Rotation Method: Rotation methods help to produce an interpretable and simplified solution by maximizing high item loadings and minimizing low item loadings. Oblique and orthogonal rotations are two types of rotation techniques which are widely used. Oblique rotation
is more accurate where the data does not meet priori assumptions (Costello & Osborne, 2005).
This method produces the construct structures that are correlated. Quartimin, direct oblimin and promax are commonly available methods where oblique rotation is concerned. In contrast, orthogonal rotation produces factors that are uncorrelated. Orthogonal method has several options for rotation; quartimax, varimax, and equamax. According to Costello & Osborne (2005) orthogonal rotation produces more easily interpretable results and is slightly simpler than oblique rotation. Varimax rotation which was developed by Thompson (2004) is the most common form of orthogonal rotational method.
The following section describes the details of EFA conducted (method and results) on the 37 items of “Occupational Stress Scale for Soldiers” generated in the phase 2.