Information Technology Reference
In-Depth Information
According to this perception, a causal relation can only be verified in experimental settings where a
response can be observed after a cause has been manipulated and the system can be isolated from other
(unknown) influences. This is necessary because there exist situations where the mere observation of
cause and effect misleadingly suggest causality between two events where there is none. For example
the influence of an exogenous third variable
U
on two endogenous variables
C
and
E
can lead to corre-
lation as well as to a temporal precedence relation between
C
and
E
although they are not causing each
other. This example demonstrates the inability of mere temporal precedence to explain the asymmetry
in causal relations quite graphically.
The restriction of interventionistic approaches to analyze the concept of causality only experimentally
is the main reason for their failure to explain many real-life problems as it is also the case for manage-
rial cause-and-effect relations: In an enterprise there is hardly any situation where an experiment-like
situation can be created because the trial-and-error manner of these settings usually would inflict losses
for the business and the response time of basic cause-and-effect relations can be rather long. As a con-
sequence it seems to be too costly to prove causality in this way.
There has been a broad discussion of causal theories in economics - as in many other sciences. In
the first half of the 20
th
century economists mostly neglected causal concepts as described above. The
standard solution was to identify additional determinants which discriminate between otherwise simul-
taneous relationships (Hoover 2008, p. 6). One important momentum for the discussion of causality in
economics was the conception of the Cowles Commission in the 1950s (Koopmans 1950). According
to this, the solution for finding causal relationships was the a priori knowledge from economic models.
This implies that - according to this notion of causality - causal structures cannot be inferred merely
by empirical data but have to be built upon well-known hypotheses which can only be falsified.
One major critique of this approach comes from the work of Lucas (1976) who argues that a change
in economic policy renders the parameters of the causal structures unstable. Lucas' critique led to the
postulate of invariance (i.e. the independence of policy changes) of economic causal structures and
therefore was the most important reason why the discussion shifted more towards so-called inferen-
tial approaches (cf. Granger 2003, p. 70). These techniques infer causal structure only form empirical
observations without the need for a priori knowledge. One of the most important contributions in this
area originates from the works of Granger who proposed a generic definition of inferential causality.
He argued that “a (time series) variable A causes B, if the probability of B conditional on its own past
history and the past history of A (besides the set Ω of the available information) does not equal the prob-
ability of B conditional on its own past history alone.” (Granger 1980, p. 330; cited after Moneta, 2004,
p. 1). The asymmetry of cause and effect in this concept known as Granger causality is secured by the
requirement that A occurs before B in time (Hoover 2008, p. 12). This notion of causality led to numer-
ous applications within the field of economics and was awarded the 2003 Nobel Prize in economics.
Newer inferential approaches employ graphical models in order to describe the causal structures
which are then tested by conditional causality concepts. These graph-theoretic approaches were mostly
developed in other scientific areas like computer science (Pearl 2000) or psychology (Glymour 2001).
However, they were then transferred to economic issues (cf. e.g. Hoover 2001; Hoover 2008). The logic
behind these techniques is rather simple: It is possible to depict a simple causal relation between two
variables as a directed edge between two nodes representing the dependent and the independent vari-
ables. By combining more causal edges it becomes possible to model more complex causal structures
as for example shielded colliders which is a common term for two causes for one indicator variable.
By analyzing the conditional correlations between these variables it is possible to introduce undirected
Search WWH ::
Custom Search