can be viewed as a form of the omitted variable problem (Heckman) and is
therefore discussed in that section.
When the cov(HR,e)¼ 0 assumption is violated, OLS is no longer unbiased and
a search for alternative estimators may be undertaken. While other estimators, in
theory, are capable of providing less biased (or in large samples, more consistent)
estimates,^8 it is important to remember that these estimators make their own
assumptions, which if not suYciently met, can lead to no better, or even worse
estimates than provided by OLS.
- 1 Measurement Error and Construct Validity
- 1 Random Measurement Error
The standard paradigm for studying HR and performance has been to deWne and
measure HR in terms of HR practices, by either asking the percentage of employees
covered by a practice or by using a likert-type rating scale that asks the degree to
which a practice is used, important, etc. The strength of implementation of
practices has not generally received much attention (Bowen and OstroV 2004),
nor have employee descriptions of practices been used. Rather, for better or worse,
a single managerial respondent typically provides the HR practice information.
For some reason, the HR and business performance literature has beenWxated
on internal consistency (e.g. Cronbach’s alpha) reliability, which provides an
estimate of how much error in responses stems from the sampling of items from
the universe of items in a particular construct domain, given the traditional
reXective measurement model. Although this aspect of reliability is important,
recent evidence indicates that the far more serious source of measurement error in
measuring HR is the sampling of raters. This fact has been either (a) ignored, or
(b) mistakenly addressed using an index, rwg, that is not a reliability and which gives
an overly optimistic view of measurement error (Gerhart 1999 ). Gerhart, Wright,
and colleagues (Gerhart et al. 2000 a, 2000 b; Wright et al. 2001 ) have documented
that measurement error due to raters in studies using theWrmlevel of analysis is
substantial. For example, in theirWrst study, Gerhart et al. reported an interrater
reliability of no more than. 30 and probably more like. 20.
What are the consequences? First, such low interrater reliabilities mean that the
HR practice scores a researcher obtains for a particularWrm depend more on the
variable that is explained within a set of equations, or in a path model, any variable that has a
unidirectional arrow leading into it. Likewise, an exogenous variable is one that is not explained in the
system, and in a path diagram, has no unidirectional arrows leading into it. To avoid confusion, I use
the terms exogenous and endogenous in the latter sense.
(^8) Roughly speaking, consistency means that in large samples, an estimator’s probability density
function (or distribution) approaches that of the population parameter as the sample size increases.
modeling hrm and performance linkages 557