whenever possible in a regression/ANCOVA approach. However, as Blalock ( 1961 :
871 ), building on work by Simon ( 1954 ), observed, ‘the question of when and when
not to control for a given variable seems to be more complex than is often
recognized.’ Duncan ( 1975 : 22 ) likewise cautions against adding control variables
‘not for any clearly deWned purpose, but simply because it is a ‘‘good idea’’ to look
at partial’ relationships.
The problem is that whether to control for a variable depends on the causal
model. Consider, for example, two of the causal models that can be used in a three-
variable case. Two variables are HR and performance. Let the third variable be a
composite of ability, motivation, and opportunity to contribute (AMO). A ‘partial
mediation’ model is:
HRAMO performance.Another model (‘control’) is
HR
AMO
performanceUse the hypothetical correlation matrix:HR AMO Performance
HR 1. 00
AMO 0. 40 1. 00
Performance 0. 20 0. 40 1. 0
With the partial mediation model, the direct eVect of HR on performance is. 048
and the indirect eVect (via AMO) is. 152. Thus, the total eVect of HR is. 200 .In
contrast, using the control speciWcation, the total eVect of HR is. 048 (direct
eVect¼. 048 , no indirect eVect). Thus, using the mediation speciWcation leads to
the total eVect of HR being estimated as roughlyWve times larger in magnitude
than when using a control speciWcation. (The estimates here can be obtained using
LISREL or by using the Alwin and Hauser 1975 approach.)
- 5 Summary
There are several approaches that may reduce omitted variable bias. However, some
of these approaches (e.g. Heckman’s procedure) rest on assumptions that may not
always be supported by the data. It was also noted that including more control
variables does not always make for better estimates, but rather depends on the
theoretical model.
566 b a r r y g e r h a r t