This post tries to unsettle some of the methods used in economics today ‘ Regression analysis and Granger causality. Apart from this objective, the post also tries to understand the meaning (rather, meanings) of causality. Do we economists mistake correlation for causality’ Can we have a single method for capturing causality’
Causation is defined in the following ways:
-the action of causing or producing.
-the relation of cause to effect; causality
-anything that produces an effect; cause [Dictionary.com]
And the definition of Correlation is:
-mutual relation of two or more things, parts, etc. [Dictionary.com]
I once asked a professor who had offered to give a lecture for our Advanced Economic Theory class this: ‘Sir, is it possible to establish causation conclusively” He replied ‘That is simple. There are these tests- Granger test, Sims test, Sargent test, McClave-Hsiao test, Haugh-Pierce test, etc.’ And he wrote the names of these so-called ‘scientific’ tests on our black board.
What we often forget is that, there is no single and simple understanding of causation. There are various kinds of causality like epistemic, conceptual, factual, counterfactual etc. Causality in economics also are of different natures- poverty is causing unemployment, increased demand for oil has resulted in oil price rise, supply constraints are hiking up the price, etc. For instance, conceptually we know that poverty causes unemployment (vicious cycle of poverty) and that increased demand causes a concomitant price rise. In economics, it is important to have an account of both conceptual as well as factual causality.
For causality to be present between two variables A and B, it is necessary for them to be related in some way. This relationship among them can be of a linear nature or a non-linear one. If it is linear in nature, then it is called correlation. [Note that regression analysis (OLS) is based on correlation and is linear in nature.] But, correlation alone does not imply causation. Hence, all those who think that causation and correlation are the same make an inductive leap ‘ from correlation to that of causation.
As R G D Allen writes in his Statistics for Economics: ‘This statistical concept of correlation is quite neutral as regards causation. One of the variables may be ’caused’ by the other, but this can only be known from other than statistical considerations.’ Usually, causal hypotheses are derived from economic theory, because ‘data does not speak on its own’. We need to pass data through theory in order to make sense of it.
However, economic theory (like any other theory) contains a lot of assumptions, mostly unrealistic. What happens to causality then’ Causality, then is dependent on these assumptions. Hence, drawing inferences from such theoretical models for practical purposes should be undertaken with caution.
It is this caution that seems to be missing amongst econometricians. This will be evident after looking into the workings of Granger causality. Granger causality analyses probabilistic causality. However, this per se is not a limitation of the test. For, in social sciences, it is exceedingly difficult to talk about deterministic causes in real-world scenario- especially, in a macroeconomic environment.
Granger causality
If X and Y are probabilistically dependent and X precedes Y, then X causes Y.
And, in actual testing, Y is regressed on X (t) and also on X (t-1). If the latter regression is found to be more significant, then X is said to Granger-cause Y. In actual practice, there are economists who forget the prefix. That is, again, some sort of correlation analysis is carried out between Y and X (t-1). In any case, the concept of causation is a problem-ridden one. And as economists who contribute to policy-making, one ought to be on their toes all the time.
I really think that all these causality tests are generally used as surgical operation tools on data by researchers to drive home some point. These test can rarely be used to know the causation (for example Granger’s test performed on some time lagged same sourced phenomenon will give strong causation.), it can only support a hypothesis.
On a lighter note, as they say, its just an “academic Mumbo-Jumbo”.
Agreed;that all this could be academic mumbo-jumbo.BUT that is the best we can do to understand stuff in complicated systems as in social sciences.
We cannot reduce the social and economic phenomena to neat equations like in Physics and Chemistry.Despite the higher precision of “hard sciences” remember Heisenberg’s UNCERTAINTY principle in physical chemistry!
It is true that proving causality is very difficult.Especially in complex systems where there are many variables at work and are difficult to control.
Being a medical man let me share few things that medical scientists use to deduce causality.These are useful in understanding economic problems too.
1.If factor A (say eating stale food)is supposed to cause effect X( loose stools)it must precede X.(temporal relationship)
2.Stronger the correlation greater is the chance for the two to be truly associated.Of course which is primary cannot be known in this and also if both are effects of some other cause cannot be known.
3.If greater dose of a proposed cause(no of cigarettes smoked) is associated with greater effect(incidence of lung cancer) and smaller dose is associated with smaller effect then causality can be suspected.(Dose-Response relationship)
4.If studies done in different centers by different investigators signals the same thing, it could be an indication of truth.(consistency)
5.If the finding is compatible with theories which are already established, it could indicate a signal.(plausibility)
6.If the effect cannot be explained by any other alternate explanation than the factor that is being evaluated then one could consider the factor to be causal.
7.If an ethical,feasible experiment can be designed to test the hypothesis with adequate control of known factors to see difference in effect in a experimental group and a control group, then significant difference could prompt us to consider causality.
8.If a factor is associated with one effect only, when it is studied could indicate some form of causality.
On the issue of Granger’s causality (though I do not understand high end mathematics)- It makes intuitive sense of outcomes/ effects to be correlated with factor being seen as a cause at that particular time but also at a time preceding it (depending on possible cause effect duration).
Correction :
It makes intuitive sense of outcomes/effects to be not only correlated with factor being seen as a cause at that particular time but also at a time preceding it (depending on possible cause effect duration).
It is quite clear that correlation does not imply causality. Correlation between two variables; whether positive or negative; just means an association. You cannot extrapolate that to mean that one variable CAUSES the other. For example, a study could show a positive correlation between the husband’s income and the wife’s height! This might be a perfectly valid correlation in a small study. Obviously one cannot imply causality with that. On the other hand regression analysis might point to a cause and effect.