# Statistical Learning and Econometrics

The principal goal of Econometrics is to find causal effects between economic variables. But the growing technology level seems to show us that Big Data and Statistical Learning result in a tradeoff with econometrics, because of according to some professors (and professionals), the Statistics field, once is applied in economics, must be guide by economic theory because its alternatives find correlations (Not inference).

Given the above, is it possible to find causal effects between economic variables with Statistical Learning methods as alternatives to Econometrics?

Thanks!

(Pardon forerrors, I'm not a native English speaker)

• Take a look at work by Judea Pearl. I haven't tried to understand even the basics of i it but he develops a whole new framework and notation that supposedly can infer cause and effect. atleast that's as much as I understand about it. Commented Dec 4, 2018 at 8:22
• Thanks @markleeds, I'll take a look at work that you've mentioned.
– SMD
Commented Dec 6, 2018 at 1:21
• it's difficult but probably worth it once you understand it. some say he's turned statistics on its head. others claim he's done a good job of marketing. I truly don't know but I know he's quite talented and been recognized for his talent so that factors in. Commented Dec 6, 2018 at 22:24
• I could imagine how smart he is. This topic is, indeed, interesting. Thank you @markleeds
– SMD
Commented Dec 10, 2018 at 21:01
• I wish I had the time to try to understand it. Enjoy. But read other things to get alternative viewpoints. I've realized how important that can be over the years. Commented Dec 11, 2018 at 22:34

As @markleeds suggested, Judea Pearl made a significant contribution to the field of Causality.

His work (though not innovative) explains that causality cannot be represented by joint probabilities. In fact, even econometric models can't model causality.

So, we have to use Rubin's Causal model, Granger Causality, Difference-in-Differences or whatever, to explore causality.

So, the answer to your question is yes and no. We can use statistical learning techniques in combination with causal frameworks, but they alone can't provide causality rigorously.

For example, I came up with a very simple solution to use special impulse-responses (in my case, one-hot vectors like $$(1,0,0), (0,1,0), (0,0,1)$$) to clear out the partial effects of input variables to the outcome in Neural Nets. Link here: http://ravshansk.com/articles/irf-ann

But all of these are just hacks and don't quite have a rigor. Judea Pearl argues that we must have a new notation for causality or use graphical models whatsoever.