Confirmatory Factor Analysis

While EFA methods can be used in a confirmatory manner, CFA has become synonymous with the use of structural equation modelling (SEM) based approaches (DeVellis, 2003; Spector, 1992). SEM allows for the comparison of the variance-covariance matrix of a set of data with the pre-specified parameters of the model being tested (Byrne, 2001; Schumaker & Lomax, 2004). It also allows for variant models to be tested and evaluated.

The aim of Study 2 was to evaluate the fit of the factor structure identified in Study 1 with (a) an employee sample, and (b) with a retirement sample. Browne and Cudeck (1993) have suggested that no a priori model will fit the data in a population perfectly, because all models are approximations. While this claim remains contentious

(e.g. Hurley et al., 1997), Browne and Cudeck suggested that an alternative way of assessing fit is to examine the relative fit of a number of alternative a priori models.

Accordingly, four models were proposed for the current study and these are illustrated in Figure 2.2. Specifically the models were:

Model 1: A unidimensional scale of organizational memory.

Model 2: Five unrelated factors.

Model 3: Five correlated first-order factors as per Study 1 (EFA).

Model 4: Five first-order factors related to an over-arching second-order factor, organizational memory.

According to Marsh and Hocevar (1985) Model 3 was most likely to represent the benchmark for all other models in the case of multidimensional measures. This model represented a confirmation of the EFA model in Study 1. It was also expected that Model 4 while not achieving the same fit as Model 3, would also surpass Models 1 and 2,

underlining the utility of the overall scale as well as use of the separate subscales in future analyses. Therefore it would be expected that:

Hypothesis 1: Model 3 will achieve greater fit than Models 1, 2 and 4.

Hypothesis 2: Models 3 and 4 will achieve greater fit than Models 1 and 2

Model 1. A unidimensional Model 2. Five un-correlated Model 3. Five correlated Model 4. Five first-order factors and scale. first-order factors. first-order factors one second-order factor.

as per the EFA model in Study 1.

Figure 2.2. Alternative factor structure models of the organizational memory scale for comparison by confirmatory factor analysis.

Goodness of Fit Indices

The SEM-based CFA approach allowed for the evaluation of model fit according to a number of goodness of fit statistics. These fit indices each express how far the sample covariance matrix differs from the estimated covariance matrix of the hypothesised model (Blunch, 2008). Again while there is much discussion over the relative merits of each of these, most have agreed on the best practice of using several fit indices, (e.g. Hurley et al., 1997). Blunch (2008) suggested selecting the best measures from a variety of groups, but in general recommended reporting the chi-square statistic (x²), with degrees of freedom (df), and significance (p-value), along with RMSEA with confidence level and PCLOSE.

If seeking a relative fit measure, Blunch recommended the use of CFI but cautioned that the weakness of this statistic is that the baseline model used is the independence model (zero correlation among manifest variables) and that this is not a realistic model. Bearing these issues in mind, the following fit statistics were employed for the current CFA.

Absolute Fit Measure

CMIN or x² represents the distance between the unrestricted sample covariance matrix and the restricted covariance matrix (Byrne, 2001). A significant x² value indicates that the observed (sample) and hypothesised variance-covariance matrices differ

(Schumacker & Lomax, 2004) and so a nonsignificant result is indicative of fit. However a weakness of this statistic is its vulnerability to sample size; if a sample is sufficiently small, the model will be accepted, if large enough, the model will be rejected (Byrne, 2001; Blunch, 2008). Commonly, findings show a large x² relative to degrees of freedom (Joreskog & Sorbom, 1993). This limitation led to the development of other fit indices, including those that follow, and ideally the x² should not be used exclusively.

Relative Fit Measure

The comparative fit index (CFI, Bentler, 1990) can range from 0 to 1.0 and in effect shows the improvement in non-centrality going from the independence model to the hypothesised model (Schumacker & Lomax, 2004). A value over .90 is considered indicative of good fit (Bentler, 1992), but more recently a revised cut-off value of .95 has been recommended (Hu & Bentler, 1999). However, there are differing opinions here, and Brown and Cudeck (1993) for example, have suggested that a CFI of .80 and above may not necessarily indicate poor fit.

Fit Measure Based on the Non-central Chi-square Distribution

The root mean square of approximation (RMSEA, Browne & Cudeck, 1989), considered one of the most informative of the fit indices, focuses on the error of

approximation in the population (Byrne, 2001). The RMSEA effectively asks how well the model would fit the population covariance matrix if it were available (Browne &

Cudeck, 1993). RMSEA values less than .05 indicate good fit, although up to .08 are considered to represent reasonable errors of approximation in the population (Browne &

Cudeck, 1993). MacCallum, Browne, and Sugawara (1996) added that values ranging from .08 to .10 can be interpreted as mediocre fit, and that those values over 1.0 indicate poor fit. A 90% confidence interval is given in AMOS to aid interpretation of the

RMSEA, with a small interval reflecting more precision (MacCallum et al., 1996).

AMOS also gives the p-values (PCLOSE) as a test of the null hypothesis that RMSEA in the population is less than .05 (Byrne, 2001), and so again a non-significant value is considered favourable.

Reliability and Validity

Internal consistency estimates were sought to evaluate the reliability of the subscales and correlations between study variables were of interest to evaluate construct

validity. Specifically, a consistent positive relationship between organizational tenure and the organizational memory subscales was sought, along with correlations between job tenure and job knowledge, and industry tenure with industry knowledge.

Samples

While the data for Study 3, Study 4 and Study 5 were to test the assumptions outlined in the introduction to this thesis, they were also suitable for the purpose of CFA for the organizational memory scale. There were two separate samples: Study 2(a) an employee sample consisting of participants from Study 3 and Study 4 that follow, and Study 2 (b), a group of retirees from Study 5.

Study 2 (a) Employee Sample