Fornell and Larcker (1981) note that constructs demonstrate discriminant
validity if the variance extracted for each is higher than the squared correlation
between the constructs. Hair, Anderson, Tatham, and Black (1998) propose that,
for discriminant validity, the composite reliability should be above the 0.70
threshold and extracted variance above the 50% threshold. All our constructs met
these criteria. To further control for potential biases stemming from the method of
data collection, we ensured that there were no differences between web-based
(69%) and paper-based respondents (31%). We also controlled for potential biases
on account of age and gender. T-tests indicated no significant differences.
As Table 1 indicates, all reliability scores were above the generally
accepted norm of 0.70. Further, a second wave of questionnaires, targeted at two
other members of sampled firms, was sent out after the initial responses had been
obtained (15% and 5% second and third respondents respectively). We calculated
an inter-rater agreement score (rwg) for each variable (James, Demaree, and Wolf
1993), which ranged from 0.67 to 0.85, which reflects “substantial” to “almost
perfect” agreement on Landis and Koch’s (1977) scale. The examination of intraclass
correlations also revealed a strong level of inter-rater reliability: correlations
were consistently significant at the 0.001 level (Jones, Johnson, Butler, and Main
1983). Finally, to check for common method bias, we conducted a Harman singlefactor
test, which did not suggest any cause for concern. Further, our confirmatory
factor analysis (CFA) model performed well in terms of fit (GFI: 0.981; RMSEA:
0.034; p ≤. 0.00; cf. null model GFI: 0.519; RMSEA: 0.22; p ≤ 0.00). All factor
loadings were well above the 0.40 level recommended by Ford, MacCallum and
Tait (1986). The CFA solution thus replicated our proposed operationalizations,
attesting to the reliability and dimensionality of the items and operationalizations