they are careful to note that mediation is not an all or nothing phenomenon and
thus statistical signiWcance testing is not suYcient for assessing the degree of
mediation. Rather, one must focus also on the percentage change in the regression
coeYcient when the mediator is added to the equation (moving from the reduced
form to the structural equation), consistent with Alwin and Hauser. Recent work
comparing tests of mediation further demonstrates the drawback of relying exclu-
sively on statistical signiWcance tests, in part, because of their very poor statistical
power in most mediation tests (MacKinnon et al. 2002 ).
- 5 Testing for Fit or Moderation
Few seem to believe that the ‘best practice’ or ‘universal’ model of HR is valid. Most
of usWnd that unlikely and can make plenty of persuasive arguments for why HR
practices must display (a) internal/horizontalWt among themselves, (b) external/
verticalWt with strategy, and (c)Wt with the institutional (including country)
environment (Boxall and Purcell 2003 ; Dowling and Welch 2004 ; Paauwe and
Boselie 2003 ).
The fact, however, is that there is precious little (formal research) evidence that
(a), (b), or (c) make much diVerence to business performance (Dyer and Reeves
1995 ; Gerhart in press; Gerhart et al. 1996 ; Wright and Sherman 1999 ). This is not to
say there is no evidence on any of these three aspects ofWt. For example, with
respect to (c) above, the regulatory environment may preclude or mandate certain
HR practices, depending on the country.
In the case of horizontalWt, however, a case can be made that there is evidence,
but it has been misinterpreted as supportive evidence (Gerhart in press). For
example, an important study by Ichniowski et al. ( 1997 ) used monthly observations
on thirty steelWnishing lines.^18 Their dependent variable was line uptime and their
independent variables were HR practices, either alone, or combined via cluster
analysis into HR systems. Their key conclusion was that ‘Systems of HRM policies
determine productivity. Marginal changes in individual policies have little or no
eVect on productivity. Improving productivity requires substantial changes in a set
of HRM policies’ (p. 37 ).
For each HR individual practice, Ichniowski et al. ( 1997 ) estimated a separate
equation. Each HR practice’s coeYcient was then compared, with and without HR
system dummy variables in the equation. The individual HR practice coeYcients
were smaller when the HR system dummy variables were in the model. This
formed the basis for Ichniowski et al.’s conclusion that changes in sets of HR
policies are necessary. However, as Gerhart et al. ( 1996 ) noted, the fact that the
(^18) Note that a working paper version of this article was available in 1993 , meaning the inXuence of
the study was felt before 1997.
modeling hrm and performance linkages 573