value, the model, in fact, becomes a change model. To see this, begin with the
model:
perftþ 1 ¼b 1 perftþb 2 hrtþetþ 1Then, impose the restriction,b 1 ¼ 1 , and rearrange terms to obtain:
perftþ 1 perft¼b 2 hrtþetþ 1Thus, by regressingWrm performance at timetþ 1 onWrm performance at time
tand HR at timet, the model estimates the eVect of HRlevelatton thechangein
Wrm performance between timestandtþ 1. This isWne if this empirical speciWca-
tion coincides with the conceptual model. For example, change in performance
would be the appropriate dependent variable if during the period between timet
and timetþ 1 , some companies changed their HR practices and the researcher was
able to measure HR at these two time periods. However, this is almost never the
case in HR and performance research. Rather, HR is usually measured at a single
point in time. In that case, it is not clear that one would expect an unchanging set
of HR practices to lead to a change in performance.
A recent study by Schneider et al. ( 2003 ) uses longitudinal data on bothWrm
performance and HR (measured as aggregate employee attitude, speciWcally job
satisfaction). For each facet of job satisfaction and overall job satisfaction, they
estimate (a) sat!perf, the correlation of job satisfaction measured at time 1 with
proWtability variables measured at time 2 , time 3 , and so forth, and (b) perf!sat,
the correlation of proWtability variables measured at time 1 with job satisfaction
measured at time 2 , time 3 , and so forth. TheyWnd that the correlations in the
(b) analyzes are generally larger than those obtained in the (a) analysis. They
conclude from this that proWtability is more likely to cause job satisfaction than
job satisfaction is to cause proWtability. Likewise, a look at their proposed research
model (that follows the empirical portion of their study) portrays job satisfaction
almost solely as an outcome variable, not as a key cause of behaviors orWrm
performance.^16
In my view, however, there is no need for a contest here. Consistent with the
simultaneous equation model discussed earlier, employee attitude can be both a
cause and a consequence ofWnancial performance. WhileWnancial performance in
yeartmay precede employee attitude for yeartþ 1 , it is also true that employee
attitude in yeart 1 precedesWnancials for yeart and so on. (See the above
discussion on whether to control for a lagged dependent variable.) To say that
causality runs in only one direction may not be accurate.
(^16) Their model also shows high performance work practices inXuencing production eYciency and,
in turn,Wnancial performance withnomediating variables (or moderators for that matter). This
seems to move us in the wrong direction. We should be focusing on opening the proverbial ‘black box’
(Becker and Gerhart 1996 ), not hypothesizing its existence.
modeling hrm and performance linkages 571