median. In an analysis (results not shown) that controlled for all of the
other HRM practices, we found this variable to be consistently positive but
insignificant for each dependent variable. In addition, various transforma-
tions of this scale (including individual dummy variables indicating whether
each practice was above the sample median, and a spline function to capture
any nonlinearities in the relationship) did not improve the explanatory power
of this model for either dependent variable.
The second complementarity test we conducted focused on interactions
among the HRM practice variables. The extreme degree of multicollinearity
among the interactions precluded our simultaneously evaluating them all.
Moreover, given the exploratory nature of this research, we experimented
extensively with subsets of interactions, both within each HRM practice
category (employee skills, motivation, and the structure of jobs and work)
and between the categories. Coefficients on interactions were insignificant
in nearly all of the models we evaluated. For example, models 4 and 5 in
Table 3 and 9 and 10 in Table 4 report two-way interactions within conceptual
groupings of variables. The interaction between training and selective staffing
has a positive and marginally significant coefficient in model 5, but all other
interactions are insignificant. Joint-F tests indicate that addition of the group
of interactions did not significantly raise the amount of variance explained
in the overall model in any case. We found little evidence of complementarity among HRM practices in the NOS data.