HR system variables (see 1996 : 28 ) were derived on the basis of the individual HR
practices using a clustering algorithm likely means that the HR system variables
were collinear with the individual practices. As such, the fact that the coeYcients
on individual HR practices are diminished when the HR system variables could just
show collinearity. It is not clear why such evidence indicatesWt or complementar-
ities between individual HR practices.^19
The Ichniowski et al. ( 1997 ) method of testing internalWt continues to be used.
For example, like Ichniowski et al., the research reported in Appelbaum et al.
( 2000 ) is well done and interesting. However, their approach to testing forWt has
similar problems. They conclude that ‘The synergies created by bundling these [HR
and work] practices together have a stronger eVect on performance than do the
individual practices’ ( 1997 : 142 ). My re-examination of their results, however, turns
up nothing that really supports this claim.^20
The concern here is that unwarranted claims regardingWt/synergy have un-
wanted consequences for research and policy. From a policy point of view, we do
not want to tell companies that they have to ‘buy the whole package’ of HR system
practices to obtain improvement if, in fact, that is not necessary. This general issue
is one that very much needs to be addressed more carefully.^21
Finally, there are three other areas where researchers sometimes do not follow
recommended practice. First, when testing for interactions by entering a cross-
product term, it is necessary to have all lower-order eVects in the model that
involve those variables. For example, to test a three-way interaction between HR,
business strategy, and country, one must include not only the three-way cross-
product, but also the three main eVects and the three two-way cross-product terms.
Second, it is recommended that when testing an interaction, one rules out the
possibility that an observed interaction is due to a quadratic eVect (MacCallum
and Mar 1995 ). Third, when plotting an interaction (always recommended), care
must be taken to include only values that actually exist in the dataset. So, for
example, it is necessary toWrst verify that there areWrms two standard deviations
(^19) Ichniowski et al. do not report the R (^2) for the equations using (a) separate HR practices, and
(b) HR practices combined into clusters/systems. Thus, there is no means of comparing theWtof
(a) versus (b) based on their article.
(^20) When I simply add the linear eVects of the individual HR and work practices variables, I obtain
predicted values of their uptime dependent variable that are quite similar to those obtained using the
system variables (derived using cluster analysis). In addition, the adjusted R^2 for the performance
equation containing the separate HR practices is larger (. 81 in their table 8. 7 ) than the R^2 in the
corresponding performance equation containing the HR clusters/systems (. 75 in their table 8. 8 ). This
provides no support for synergy. 21
A broader issue is whetherWt is an overly static concept, which has received too much attention,
to the detriment of related issues, such as ‘sustainableWt,’ which ‘can be achieved only by developing a
Xexible organization’ (Wright and Snell 1998 : 758 ). Useful here would be longitudinal data on how
Wrms are able to respond in HR areas (practices, employee skills, employee behaviors, p. 758 ) over time
to changes in their competitive environments.
574 b a r r y g e r h a r t