The knowledge strategy orientation scale: individual perceptions of firm-level phenomena.

Exploratory Factor Analysis. In order to pre-test our items and explore the underlying factor structure of our knowledge strategy orientation subscales, we used principal axis factoring in an Exploratory Factor Analysis (EFA), with a promax rotation, on the data from Sample One. While there is some disagreement on minimum sample size requirements, many methodologists suggest at least five to ten respondents per item are needed for EFA (Comrey, 1988; Hair et al., 1998). We have more than 12 respondents per item. We used the latent root criteria of an eigen value greater than one and a scree plot for the determination of factor extraction. Additionally, we considered items with loadings of greater than .40 to be "substantial" (Floyd and Widaman, 1995) and loadings above .50 to be "very significant" (Hair et al., 1998). Because we have a theoretical basis to support our belief that these constructs are correlated, we used the promax form of oblique rotation. It must be noted that EFA tends to capitalize on the chance characteristics of a sample. Because the purpose of this article is to examine the factor structure of responses to our scale items, we later used confirmatory factor analysis to cross-validate the results of our Sample One EFA.

Confirmatory Factor Analysis. Confirmatory factor analysis allows a confirmatory, rather than an exploratory, approach to determining the underlying structure of observed variables (Harris and Schaubroeck, 1990), and provides a means of assessing the relationships between constructs without the bias commonly introduced by measurement error (Judd et al., 1986). Confirmatory factor analysis is used to determine the extent to which alternative models explain the relationships between items in a scale. Two competing measurement models of strategic orientation were evaluated in this study. The alternative CFA models were: (a) a one-factor model that forced all items designed to measure Explorer and Exploiter onto a single factor of Knowledge Strategy Orientation and (b) a two-factor model that forced the Explorer items and the Exploiter items onto separate factors. In each model, error variances for the items were not allowed to correlate. Should the one-factor model provide a fit of the data equivalent to the two-factor model, it would indicate a single underlying latent construct (i.e., Explorer and Exploiter as opposite ends of an unidimensional knowledge strategy continuum). If the two-factor model should provide the better fit than the one-factor model, then our conceptualization of Explorer and Exploiter as distinct and independent constructs will be supported.

As suggested by Thompson and Daniel (1996), CFA is most useful when the researcher tests a priori models, because more effective decisions can then be made about the viability of the target model. Because the a prior/models above are nested, the chi-square difference can be used to test for significant differences between the models. If the chi-square difference is significant, it indicates that the more complex two-factor model fits the data significantly better than the simpler one-factor model.

Hu and Bentler (1998, 1999) recommend that several goodness-of-fit tests be conducted and that their resulting indices be reported. These indices are of two types: absolute and incremental. An absolute index tests how well the model covariance matrix reproduces the sample covariance matrix while an incremental index tests the fit of the hypothesized model as compared to a baseline model. The most commonly-used absolute fit index is the chi-square test that assesses the discrepancy between the implied covariance matrix of the hypothesized model and the sample covariance matrix. A non-significant chi-square is the desired result of this test as it suggests the model may be a reasonable approximation of the data. However, many researchers (c.f. Fan et al., 1999; Hu and Bentler, 1995) have cautioned that using the chi-square test as an assessment of fit can be confounded by sample size because as sample size increases, the chance of the chi-square test supporting a fit of the data decreases. Thus, small differences between the sample covariance matrix and the reproduced covariance matrix may be determined to be statistically significant and lead to rejection of the model. With this in mind, supplemental absolute indices were employed.

Another absolute index, the standardized root mean square residual (SRMR) is reported as a summary statistic based upon residuals between the elements of the implied and observed covariance matrices. The standardized root mean square residual ranges from 0 to 1 and values close to 0 are preferred. In fact, Hu and Bentler (1998, 1999) suggest that researchers always use the SRMR to assess model fit because of its sensitivity to simple model misspecification (misspecified factor correlations). They suggest that target values of the SRMR should be less than .08 in order to indicate adequate model fit. Another absolute fit index, the root mean square error of approximation (RMSEA), is reported in this study as well. The RMSEA assesses lack of fit based upon model misspecification and provides a measure of this discrepancy per degree of freedom (Browne and Cudeck, 1993). This fit index is quite sensitive to complex misspecification (i.e., misspecified factor loadings; Hu and Bentler, 1998). It ranges from 0 to 1, with target values of less than .08 indicating adequate fit (Browne and Cudeck, 1993).

Incremental fit indices are also recommended (Hoyle and Panter, 1995; Hu and Bender, 1999) to assess model fit. The comparative fit index (CFI) developed by Bender (1990) is reported here. It is sensitive to misspecified factor loadings (Hu and Bentler, 1998) and assesses the improvement of fit of the hypothesized model over the null model. The null model is an independence model in which variables are hypothesized to be uncorrelated. The CFI ranges from 0 to 1, and values greater than .95 have recently been advocated (Hu and Bender, 1999) as an increase from earlier target values greater than .90 (Hoyle and Panter, 1995).

RESULTS

Sample One EFA Results

In Sample One, our principal axis analysis resulted in two factors with eigen values of 3.409 and 1.086 being extracted that explained 56.15% of the variance. As we envisioned, our promax oblique rotation resulted in each Explorer item loading more highly on one factor than the other and each Exploiter item loading more highly on the other factor. Three of four Explorer items showed "very significant" loadings greater than .60. Two of four Exploiter items showed "substantial" loadings greater than .40, while another item showed "very significant" loading. See Table 1 for the resulting pattern matrix. With this factor structure in mind we then cross-validated these results on the data from Samples Two and Three using CFA.

Item Level Statistics for Samples Two and Three

Each CFA measurement model was estimated in this study using LISREL 8.71 software (Joreskog and Serbom, 2004). A component of the LISREL software, PRELIS 2.30, was used to assess univariate normality and to generate the covariance matrix upon which the CFA was conducted. Kline (1998) advocates upper boundaries of 3.0 for skewness and 8.0 for kurtosis as indicators of univariate normality.

Sample Two. The univariate data for the Explorer and Exploiter scales were approximately normally distributed with skewness for the eight manifest indicators ranging from -0.88 to 0.39, and kurtosis ranging from -1.56 to 1.45 (see Table 2). Based upon the descriptive statistics for the sample, it appears that the data were approximately normally distributed. Therefore, the maximum likelihood (ML) method of estimation was employed in CFA.

Sample Three. The univariate data for the Explorer and Exploiter scales were approximately normally distributed with skewness for the eight manifest indicators ranging from-1.10 to 0.44, and kurtosis ranging from -1.39 to 1.31 (see Table 2). Based upon the descriptive statistics for the sample, it appears that the data were approximately normally distributed. Therefore, the maximum likelihood (ML) method of estimation was employed in CFA.

Confirmatory Factor Analysis Results

The eight items comprising the Explorer Orientation and Exploiter Orientation scales were subjected to CFA. Two models were compared: a one-factor model forcing all eight items onto the same factor and a two-factor model forcing the Explorer items and Exploiter items onto their respective factors. Error terms were not allowed to correlate in either model, and in the two-factor model items were not allowed to cross-load. See Table 3 for the fit indices of these two models in our samples.

Sample Two. The most complex model was the two-factor model, which resulted in CFI = 0.92, RMSEA = 0.097, and SRMR = 0.065. The SRMR indicates good fit of the data to the model, and the RMSEA and SRMR are only slightly outside the recommended thresholds. The more parsimonious one-factor model resulted in CFI = 0.79, RMSEA = 0.16, and SRMR = 0.098. None of these indices meets the criteria for good fit. Additionally, the [chi square] for the two-factor model was 39.60 (p < .001), while the [chi square] for the one-factor model was 73.24 (p < .001), resulting in a [DELTA][chi square] of 33.64 (p < .001). The significant [DELTA][chi square] indicates that the two-factor model fits the data significantly better than the one-factor model, providing evidence of the superior fit of the two-factor model.