error is the use of instrumental variables, usually using two-stage least squares
estimation (discussed later). In the multivariate case, LISREL is very useful.
In our own recent work (Fulmer et al. 2003 ), we have talked about employee
relations as a source of competitive advantage, conceptualized and measured in
terms of (multiple) employee (not a managerial informant’s) views. A methodo-
logical advantage of this approach is that multiple responses from each organiza-
tion are averaged, which, with enough responses, virtually eliminates measurement
error due to sampling of raters. What employees perceive, think, and feel about HR
and employee relations may also have more theoretical credibility as a cause of
business performance than what Purcell ( 1999 ) described as ‘crude’ measures of HR
practices.
- 2 Non-random Measurement Error
Single-respondent reports of HR practices can cause additional problems, espe-
cially if the same respondent also serves as the source of performance data. This
design may result in measurement errors in HR and performance being correlated.
For example, a respondent may consistently have positive or negative response
errors across scales. Or,Wrms having better performance may over-report levels of
HR eVectiveness or even use of practices thought to be desirable (Gerhart 1999 ).
There has been much debate about whether this common method variance
makes a diVerence in management researchWndings. While method variance
may not make much diVerence in binary tests of statistical signiWcance (Doty
and Glick 1998 ), it does inXuence eVect size. SpeciWcally, Doty and Glick’s meta-
analysis found that correlations between an array of measures used in organiza-
tional behavior were 25 percent higher, on average, when both measures were based
on single source self-reports than when they were not. They also found that this
diVerence was larger still when the study design was cross-sectional.
Strategies for control of common methods variance include the use of the
multitrait-multimethod matrix (Campbell and Fiske 1959 ) to assess convergent
and discriminant validity, with later approaches applying structural equation
modeling (e.g. Alwin 1974 ). In the absence of multiple methods (and thus the
ability to identify and remove method variance this way), a more recent suggestion
is to use a marker variable (Lindell and Whitney 2001 ), which theory says should
have no relationship with other constructs in the study. Any observed relationship
is thus assumed to be due to common method variance only. This relationship is
then partialed from the relationships of substantive interest. As discussed later,
eliminates any measurement error by (apparently) permitting no disagreement and then once any
disagreement in interviewee data is eliminated or reduced, giving this homogenized data to two raters.
Not surprisingly, their ratings of the same, homogenized information are highly similar. It should also
be noted that Study 1 in Wall et al. was at the plant level of analysis (where the second Gerhart et al.
2000 bstudy found that reliabilities were generally higher).
modeling hrm and performance linkages 559