> 6. Slogan: "causal net before functional model".

As described in Spirtes, Glymour and Scheines (1993,2001) and Pearl
(2000), these formalisms are equivalent for the purposes of determining
the population distributions resulting from the implementation of
policies,.

A little more discussion follows:

It is true that traditionally functional relationships (e.g. linearity)
have been assumed in structural equation models, but more modern
approaches are non-parametric, in which case, at a conceptual level, there
is no difference between the two.

[As a side note: Structural equation models can be interpreted as relating
to individual causal effects - i.e. inferring causes from effects, "Would
John have a headache had he not drunk coffee this morning given that he
did drink coffee and does have a headache", but this has not been done by
econometricians - whereas this is not true of "causal nets". However, such
inferences are based on assumptions that cannot in general be tested even
in experimental contexts. Further Econometricians have typically not drawn
such inferences from their models, hence this is not directly relevant
here. See Pearl (2000) Ch.5 and 7 for further discussion.)


> 4. Consequently it would be better to use causal Bayesian networks instead,
> since in that case we only need to estimate the probability distribution of
> each variable conditional on its direct causes. Such models should be
> checked (one should check that the causal Markov condition holds for the
> model and that the model is robust for forcasting and modeling
> interventions) and refined as necessary.

If "only" we ever knew the direct causes and could directly check the
Markov property!

However, such conditional independence assumptions are equivalent to the
assumptions made by the coresponding structural equation model.

Note that in general there will be more than one "causal net"  obeying a
given set of conditional independence relations, hence the fact that our
data are compatible with the independence relations given by a given
causal model does not verify the model, to the extent that the data
confirms or supports the proposed model, it also supports every other
model in the given Markov equivalence class.

Even this inference requires the (untestable) assumption that conditional
independence relations present in the population distribution only arise
through causal structure and not through cancellation of parameters
(so-called violations of faithfulness or stability).

In general it is often unclear whether all of the relevant variables have
been included, which means that models with hidden variables (or
"correlated errors" or "semi-Markovian") must also be considered, thereby
expanding further the class of causal models that are not ruled out by the
data. This problem is equally present for both formalisms. (Again see the
above references for more discussion.)

In general, on finite samples, it is often unrealistic to hope to reliably
determine the conditional independence structure holding in the population
non-parametrically. For this reason functional forms are often assumed
(e.g.  "noisy-or" in the causal net literature; linearity in
Econometrics). Such problems have been particularly acute in Econometrics
where datasets have traditionally been rather small (this is starting to
change).

To summarize, there is no significant distinction between the two
approaches. The primary practical problems present in applications of
structural equation models - lack of data, large numbers of possible
variables and possible models, are also present for causal nets. It is
true that users of structural equation models have often been unclear or
even unaware of the causal interpretation of their models and this has led
to much confusion and obscurity. In this regard, the community of causal
net modellers may have some advantage (though this issue is also often not
addressed in many Bayes Net applications).

The issues of stationarity raised by Peter McBurney are of course also of
concern in many Econometric applications. Feedback is another possibility
in many econometric settings For more discussion of this see Lauritzen and
Richardson "Chain graph models and their causal interpretations (with
discussion)", (2001) Journal of the Royal Statistical Society Series B.

Thomas Richardson

Department of Statistics
University of Washington

Reply via email to