Note also that Schneider et al.: (a) rely exclusively on bivariate correlations,
meaning no control variables are used, (b) provide no eVect sizes in dollar terms, so
it is diYcult to interpret the policy relevance of theirWndings, (c) do not explain
why the correlation of satisfaction withWnancial outcomes is sometimes stronger
the longer the lag until theWnancial outcomes are measured, (d) use only short-
termWnancial performance, rather than long-term measures, which might better
capture employee attitudes/relations as a source of sustained competitive advan-
tage (e.g. Fulmer et al. 2003 ), and (e) few of the sat!perf and perf!sat
correlations discussed above actually diVered statistically.
- 4 Summary
It is unlikely that causation runs in only one direction between HR and performance.
Simultaneous equation models oVer the prospect of a more accurate and nuanced
description of this relationship. However, the challenging assumptions of these
models, for example, regarding instrumental variables, may be diYcult to satisfy.
- 4 Causal Inference and Mediators
Becker and Gerhart ( 1996 ) emphasized the importance of specifying and testing
intervening causal mechanisms as a way to better understand how HR inXuences
performance, as well as its importance in assessing the credibility/causality of the
relationship. However, as Purcell ( 1999 : 29 ) observed, theWeld has tended to ‘take
for granted’ that ‘often very crude’ measures of HR practices are mediated by things
like ‘worker eVort, morale, cooperation, attitudes, and behaviour.’ More recently,
workforce ability, motivation, and opportunity to contribute (AMO) have been
hypothesized as key mediators (Appelbaum et al. 2000 ; Boxall and Purcell 2003 ).
Again, however, this hypothesis is largely untested (Gerhart in press).
The classic approach to testing and estimating mediation, described by Alwin
and Hauser ( 1975 ), begins with the reduced form of the model, which refers to an
equation for each endogenous variable that includes only purely exogenous vari-
ables on the right-hand side.^17 This is contrasted with the structural model, which
includes all right-hand-side variables, exogenous or otherwise. The total eVect is
deWned as the coeYcient on the exogenous variable in the reduced form equation.
The direct eVect is deWned as the coeYcient on the exogenous variable in the
structural equation. The indirect eVect is deWned as the total eVect minus the direct
eVect.
A later paper by Baron and Kenny ( 1986 ) uses a similar logic in testing for
mediation, but with greater emphasis on statistical signiWcance testing. However,
(^17) Recall that endogenous variables have determinants in the model, whereas exogenous variables
have no determinants (i.e. are unexplained) in the model.
572 b a r r y g e r h a r t