Either type of error, of course, causes problems. Research may incorrectly deem
an eVective HR system as ineVective or an eVective HR system as ineVective.
Incorrect research inferences may stiXe research in a fruitful area or create a cottage
industry in an area not worth the bother. If, as we hope, research has some eVect on
managerial policy and practice, these errors of inference could likewise lead to
errors in the adoption or retention of HR systems in organizations.
I have two general goals in this chapter. First, I hope to help readers better
evaluate the contribution of published research on HR and performance. Second,
I hope to help authors in preparing their work for publication and avoid problems
that may otherwise lengthen the review process or adversely aVect the publication
decision. I begin with a simple model.
27.2 A Simple Model of HR
and Performance
.........................................................................................................................................................................................
The typical approach in HRM and performance is to use a model like the following:
Perf¼â 0 þâperfhrhrþåwhere population parameters are:â 0 , the intercept,âperfhr, an unstandardized
regression coeYcient representing the performance–HR relationship, andå, the
error or disturbance term, which captures all unspeciWed causes of Perf. This is the
classical linear regression model. Estimation using sample data, by ordinary least
squares (OLS), for example, yields:
Perf¼b 0 þbperfhrhrþe
Under standard assumptions, the OLS/classical regression estimator is unbiased
and eYcient among the class of linear estimators (the Gauss–Markov theorem).
Unbiased means that the mean (or expected value) of the distribution of b across
repeated samples is equal to the population parameter,â.EYcient (or best) means
that b has minimum variance (relative to other estimators in the class) across the
distribution of repeated samples.^2 Adding the assumption that the disturbance
term is normally distributed turns the model into the classical normal regression
model, which facilitates hypothesis-testing and formation of conWdence intervals
(e.g. Greene 1993 : ch. 10 ). In empirical work on HR and performance, this basic
model can and should, of course, be expanded to include other determinants of
performance.
(^2) EYciency is only meaningful in conjunction with bias because any constant (e.g. b¼ 4 , the
uniform number for hockey players Bobby Orr and Chris Gerhart) will be an eYcient estimator, but
of course, is likely to be a biased estimator.
modeling hrm and performance linkages 553