reported that they used a Hausman test for endogeneity, but did not report speciWc
results.^15 None of the three studies reported results of theWrst-stage regression,
which as we have seen, are necessary to evaluate the IV/ 2 SLS procedure. Only Bae
and Lawler reported what they used (company years in operation, log of number of
employees, whether a joint ventureWrm) as instruments. We should note, however,
that this increased emphasis on the Wrst-stage results did not appear in the
literature until after some of these studies of HR and performance were already
completed.
Another potential use of IV/ 2 SLS would be to examine the relationship between
attitudes and performance discussed below. However, in the case of the perform-
ance equation, what variable would make sense as an instrument for attitude?
Many texts are silent on the question of where one might look for possible
instruments. Kennedy ( 1992 ) suggests some general candidates for instruments,
most relevant here being the lagged value of an independent variable. One possi-
bility would be to use attitude from an earlier time period as an instrument
for current attitude. Wooldridge ( 2002 : ch. 5 ) provides several examples of how
IV/ 2 SLS has been used in economics studies (e.g. on education and wages) and
the ingenuity required toWnd a good instrument.
Finally, the 2 SLS estimator will have a larger variance across samples than the
OLS estimator. For this and other reasons, it is important to test whether this loss
of eYciency is oVset by an increase in consistency. Wooldridge ( 2002 ) provides a
helpful discussion of how to proceed in this regard.
- 2 LISREL
A SEM approach such as LISREL (Joreskog and Sorbom 2002 ) can also be used to
estimate parameters of simultaneous equation models. LISREL is most useful when
there are (a) measurement error issues or signiWcant work to be conducted on the
dimensionality and validity of constructs, (b) simultaneity issues, or (c) where a
system-wide/full-information estimator (e.g. maximum-likelihood) is useful either
to increase eYciency (by recognizing cross-equation error term correlations) or to
provide for a goodness ofWt test for a system of equations.
If none of these conditions hold, then LISREL may be unnecessary and the
American Psychological Association Guidelines on statistical methods (Wilkinson
et al 1999 ) recommendation of ‘choosing a minimally suYcient analysis’ should be
kept in mind. In addition, econometricians have long relied more heavily on
limited information/single-equation estimators such as 2 SLS to a greater degree
published an article entitled ‘An Illustration of the Use of Instrumental Variables’ by Foster and
McLanahan, implying some novelty of IV/ 2 SLS.
(^15) The Hausman test (e.g. Wooldridge 2002 : ch. 6 ) compares the OLS and 2 SLS estimates. If the
estimates are the same, they should diVer only by sampling error. If the estimates are diVerent (not just
statistically, but practically), there may be reason for concern with endogeneity.
modeling hrm and performance linkages 569