Why do we need causal inference when we have correlation?

In this post, I delineate a topic which is garnering attention, pretty quickly!, over the past years inspiring revolution about how we perceive AI. The idea of Causal Inference & its tools was propounded by Judea Pearl, who also espoused a probabilistic approach to AI. He describes causal inference as establishing a cause-effect relationship in any event by analysing the data collected instead of merely curve fitting.

Causal inference defines a direct relationship between the observed data(effect) and the actual source(cause) that produced the observed data. This relationship is poorly explained by classical correlation because it is not designed to capture such causal-effect-direct relationships, but only proportionality relationships.

There is a stark difference between proportionality & causal-effect relationships, for instance, a rooster crowing in the morning doesn’t establish the fact that the crow of rooster caused the Sun to rise! It is just that the two events, sunrise & crowing occurred at the same time which had their very own causes. However, the fact that it was night a while ago the Sun appeared is a very concrete cause for the morning(effect), introducing such information to our models while building them is deemed to have profound impact at realising artifitial intelligence.

The best examples can be found in health care or clinical sciences. For example, one may ask, “Does smoking cause cancer?”. If correlations are used, then only correspondence is established, which is due to the lurking factors called confounding variables in statistics; rather, prima facie questions like “Will a non-smoker smoking develop lung cancer?” & “Does holding a lighter cause cancer?” guide the journey in causal inference.

To conclude correlation does not imply causation!