particular person completing the survey than on what practices are actually used in
the Wrm because the HR practice scores are idiosyncratic, rather than being
consistent across diVerent managerial respondents.
Second, the correction for attenuation in an unstandardized regression coeY-
cient, though little discussed, diVers from the correction for attenuation in a
correlation (Gerhart 1999 ; Gerhart et al. 2000 a). The corrected regression coeY-
cient is equal to the observed regression coeYcient divided by the reliability of HR,
whereas the corrected correlation is obtained by dividing by thesquare rootof HR’s
reliabilityandperformance’s reliability. Therefore, unreliability in the independent
variable, HR has a larger impact on the unstandardized regression coeYcient than
on the correlation.^9
Thus, the 20 percent eVect of HR onWrm performance (for an increase in 1 SD in
HR practices) found by Huselid and others (see Gerhart 1999 ), once corrected,
becomes a 67 percent (. 20 /. 30 )eVect size using the. 30 reliability estimate. That is
the eVect size compared to the mean. If comparing low ( 1 SD) and highWrms (þ 1
SD), the highWrms are 167 percent of the mean and the lowWrms are at 33 percent
of the mean, which implies that the highWrms have 167 / 37 ¼ 4. 5 times higher
performance. Is this eVect size credible? (Keep in mind that other inXuences on
Wrm performance, e.g.Wnance, marketing, operations, etc. are typically not in-
cluded in obtaining such estimates.) If not, it may indicate a need to re-examine
the entire approach. With a more positive assessment, it is still clear that method-
ology (reliability estimation in this example) matters a great deal in quantifying the
impact of HR.^10
In research on HR and performance, there are multiple sources of measurement
error (both items and raters in the case here). This is best addressed by estimating a
generalizability coeYcient (Cronbach et al. 1972 ), which is equivalent to a reliability
that recognizes multiple sources of error. Gerhart et al. ( 2000 a) provide a tutorial
and example.^11 One standard econometric approach to correcting for measurement
(^9) In the case of more than onexvariable, the eVects of measurement error are more complex. In
fact, in the multivariate case, it is mathematically possible for measurement error to result in upward
bias in a regression coeYcient. We discuss two predictor scenarios later in this chapter.
(^10) When using the plant/facility as the level of analysis, which is the typical unit of analysis in
industry level studies (e.g. Batt 2002 ; Hunter and Lafkas 2003 ; MacDuYe 1995 ), evidence suggests that
reliability may be signiWcantly better (Gerhart et al. 2000 b). ThisWnding makes sense because plant
practices are easier to observe, both because plants are smaller on average thanWrms, and because
there is less likely to be variation in actual HR practices within a plant compared to within aWrm with
multiple sites (and many employees).
(^11) A nicely done study by Wall et al. ( 2004 ) that might (incorrectly in my view) be interpreted as
showing that reliability is less of an issue than we have suggested. Wall et al.’s approach in measuring
HR practices included the following steps: (a) three diVerent people per company were interviewed,
(b) ‘evidence from each interviewee wascross checked with that from others(emphasis added), and
with additional information from company documents and a tour of the manufacturing facility’
( 2004 : 101 ). Finally, (c) ‘based on this information, two researchers, unaware of company perform
ance, independently rated the extent of use of each of the practices’ (p. 101 ). This process essentially
558 b a r r y g e r h a r t