longitudinal data can also be used to remove person-speciWc, time-invariant
omitted variables of this sort. Finally, when conducting aggregate-(e.g. organiza-
tion-) level analyses, OstroVet al. ( 2002 ) have developed a method where diVerent
subsamples are used to measure diVerent constructs, eliminating within-person
correlations.
- 3 Construct Validity
The empirical methods discussed above are important in assessing construct
validity. However, construct validity cannot be established by empirical methods
alone. Rather, they must be interpreted in the context of a careful deWnition of the
construct and its role in a nomological network (Schwab 1980 ). Indeed, Schwab (p. 34 )
notes that ‘Since the criterion in construct validity is conceptual, direct tests are not
possible.’ Therefore, one mustWrst decide whether the conceptual deWnition of the
construct is adequate, then consider the empirical evidence. One could argue that
early studies of HR and performance did a reasonable job of deWning the HR
construct, but the correspondence of the chosen measures with the deWnition was
not always what one might have liked. For example, Huselid ( 1995 ) essentially
conceptualized HR as having an impact on ability, motivation, and opportunity to
contribute (see also Appelbaum et al. 2000 ). (I would add cost as a fourth
mediating variable, Gerhart in press.) However, his two HR scales were only
indirectly related to these sub-constructs, or perhaps more accurately, mediating
variables. Consider, for example, that the two highest-loading items on hisWrst HR
scale, ‘employee skills and organization structures,’ were: ‘What is the proportion
of the workforce whose job has been subjected to formal job analysis?’ and ‘What is
the proportion of the workforce who are included in a formal information sharing
program (e.g. a newsletter)?’ The question is whether these items are critical
components of the HR domain and are they the major drivers of ability, motiv-
ation, opportunity to contribute, and cost.
- 4 Summary
To accurately describe the magnitude of the HR–performance relationship, one
must report a practically meaningful eVect size estimate, typically an unstandard-
ized regression coeYcient, as well as its precision (e.g. the conWdence interval).
Further, the deWnition and measurement of ‘HR’ (and performance, for that
matter) can greatly inXuence observed eVect size estimates.
- 2 Omitted Variable Bias
Many factors presumably inXuence performance, but most empirical studies
include a short set of right-hand-side variables beyond those related to HR.
Huselid and Becker ( 2000 ) argued that omitted variable bias was likely the major
560 barry gerhart