than full-information/system estimators such as maximum-likelihood because the
latter (Bollen 1996 ; Curran et al. 1996 ; Kennedy 1992 ): (a) assume multivariate
normality, (b) while having superior asymptotic/large-sample properties, may not
perform as well inWnite samples, which empirical research uses, and (c) allow a
speciWcation error in one equation (e.g. an incorrectly speciWed zero path) to bias
parameter estimates for other equations.
- 3 Longitudinal Data and Time Precedence
As noted earlier, Cook and Campbell ( 1979 ) identiWed time precedence of cause
and eVect as a necessary condition for causal inference. Longitudinal data is
necessary to satisfy this condition. Interestingly, a recent reviewWnds that this
condition is not only rarely met in the HR and performance literature, but worse,
HR is typically measured after performance (Wright et al. 2005 ). So, there is clearly
a great deal of room for improvement on this front.
Wright et al. ( 2005 ) also raise the possibility of reciprocal causation, especially
when HR is deWned in terms of employee attitudes. They do not estimate a model
with reciprocal causation, but do present correlations of an HR practice index and
an employee attitude measure with performance outcomes collected both before
and after the HR/attitude measures in 45 to 62 units of an organization. They then
seek to determine whether (a) research design (causal ordering of HR and per-
formance) and (b) controlling for concurrent performance diminishes the correl-
ation between past HR practices and subsequent performance (i.e. in a design
having correct time precedence if interested in the HR!performance path).
For some reason, their results do not seem to show any eVect of design (i.e.
whether correct time precedence exists) on the correlations. As one might expect,
when concurrent performance is controlled, the relationship of HR/attitude with
later performance is substantially diminished. The question is why. Wright et al.
consider a number of explanations: the relationship is non-recursive, the relation-
ship is spurious, or that it is due to temporal stability in both sets of variables.
TheWrst and third explanations seem quite similar. As Gerhart and Milkovich
( 1990 ) noted, if the causal model is something like:
hrt 2 !perft 1 !hrt!perftþ 1then, yes, controlling for perf att 1 will reduce the relationship between hr at time
tand perf attþ 1 , especially if hr and perf are stable over time and have stable
reciprocal eVects on one another over time. But, by controlling perf att 1 , one is
also removing the earlier eVect of hr att 2. So, they warned against over-control.
Controlling for the lagged value of performance almost never yields empirical
estimates that correspond to the conceptual model. Indeed, it is a mis-speciWcation
to control for the lagged value of performance in a model that seeks to explain
diVerences in the level of performance acrossWrms or units. By including a lagged
570 b a r r y g e r h a r t