above the mean on HR practices and two standard deviations below the mean on a
diVerentiation business strategy before including such a point in theWgure.
- 6 MultiLevel Models and Concepts
Sometimes the goal of a paper is to study the eVect of HR policy on performance,
but it uses a design where there is no organization-level or unit-level variance in
HR. The typical approach in such cases seems to be to regress a perceptual measure
of eVectiveness (e.g. employee attitude) on a perceptual measure of HR (e.g.
opportunity to contribute). However, this analysis is a person-level analysis and
any relationship between HR and performance in this design is likely due to
common method variance if a single unit or organization has a single HR policy.
The exception would be if it can be shown that what appears to be a single policy is
implemented diVerently in diVerent parts of the organization, perhaps by super-
visors. But, to do this, one must show that at this supervisor/work group level of
analysis, there is suYcient between relative to within group variance (using the
appropriate ICC index), and then conduct the study at that level of analysis, not at
the person level of analysis. Whatever the level of analysis, it must be shown that
self-reports are not idiosyncratic to the respondent, but rather reXect a higher level
(unit orWrm) property. It is very diYcult to publish a paper that uses a single
source for all measures, but many such papers continue to be submitted to
journals.
Hiearchical linear modeling (HLM, Raudenbush and Bryk 2002 ) is increasingly
used for multilevel data and has application to the HR and performance literature
where individual data nested within organizations/units is used (OstroV and
Bowen 2000 ). Typically, data at the person level is nested within units or organ-
izations and thus is not independent, contrary to the assumption made by the
classical regression/OLS model. HLM has the advantage of incorporating ICC
analyses and of estimating standard errors that are corrected for the dependence
of observations nested within units/organizations. However, many statistical pack-
ages accomplish the same thing by allowing estimation of robust standard errors.
27.5 Conclusion
.........................................................................................................................................................................................
My goal here has been to identify challenges in estimating eVect sizes and drawing
causal inferences in research on HR and performance and to consider possible
solutions to these problems. In each of the areas discussed, researchers regularly
modeling hrm and performance linkages 575