27.3 Estimating and Interpreting
Effect Sizes
.........................................................................................................................................................................................
In the model above, the unstandardized regression coeYcient is the primary eVect
size estimate. However, HR research draws from multiple disciplines, and its
methodology reXects this fact. Likewise, the approach to eVect size varies across
disciplines. For many in HR, the main foundation is in psychology. Psychology-
based research has traditionally focused, sometimes exclusively, on statistical sign-
iWcance testing. However, psychologists have increasingly recognized the problems
with this approach (Cohen 1994 ; Schmidt 1995 ). Statistical signiWcance tests alone
tell us whether our statistical estimates meet some minimum level of precision,
which is important, but it is only a binary index. Knowing that a relationship is non-
zero is, by itself, not terribly interesting. With a suYciently large sample size, even
relationships that are so small as to be trivial with respect to practical relevance can
be statistically signiWcant.
In studies of HR–performance, where large sample sizes are diYcult to obtain,
given the unit orWrm level of analysis, the opposite problem, that practically
important relationships may be missed because of inadequate statistical power, is
more common. Standard practice in statistical signiWcance-testing is toWx the Type I
error rate (usually at p¼. 05 ), thus forcing an increase in Type II error (i.e. a
decrease in statistical power) rates as sample size decreases. Rosnow and Rosenthal
( 1989 ) show that in the case of a population r¼. 10 (a reasonable estimate for the
HR-performance correlation)^3 and using a signiWcance level of p¼. 05 , the ratio of
Type II to Type I error is 2 with N¼ 1000 , 10 with N¼ 400 , and 17 with N¼ 100.
Likewise, Cohen ( 1992 ) shows that with a population r¼. 10 and p¼. 05 ,tohave
a. 80 probability of detecting the eVect (i.e. statistical power) would require N¼ 783 ,
a sample size much larger than typically used in the HR–performance literature.
Thus, statistical signiWcance-testing is not enough. What we really need in an
area like HR and performance, which aspires to have practical implications, is (a)a
meaningful index of eVect size (Becker and Gerhart 1996 ), typically an unstan-
dardized regression coeYcient, if the dependent variable is measured on a ratio
scale (e.g. proWtability, shareholder return, turnover, uptime), and (b) a conWdence
interval placed around the eVect size that conveys the precision of the estimate.
Both are important because investing in an HR system having a mean eVect that is
smaller, but less variable/uncertain, could, for example, be preferred to investing in
an HR system having a larger mean, but also a higher variance, eVect. Smaller
variance (i.e. the standard error of the regression coeYcient) is achieved as
(^3) Based on Gerhart et al. ( 2000 a: 809 ), r¼. 10 is a realistic estimate of the HR performance
correlation when examining the Huselid ( 1995 ) study.
554 b a r r y g e r h a r t