4
$\begingroup$

I have just started this introductory course to causal inference. The DAG approach is completely new to me even though I come from an econometric background (though that dates back to 15 years ago).

The discussion around confounders reminds me of the endogeneity problem in econometrics, especially when confounders are unmeasured (e.g. omitted variable bias). In this case, I remember instrumental variables to be a workaround strategy, and explaining IV with DAGs makes a lot more sense!

However, the prefered strategy with DAGs seems to be blocking backdoor paths by conditioning on confounders (which will allow the ignorability assumption $(Y^{0}, Y^{1} \perp A|X)$ to be valid). I am wondering how blocking a path from a confounder is actually different from including those confounder in the regression?

For simplicity, suppose we have this DAG :

enter image description here

I would believe that if $X$ (the confounder) is measured in addition to $A$ (the treatment), then regressing $Y$ (the outcome) on both variables would lead to correct estimates (supposing we are using the correct regression function).

Am I wrong?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
7
$\begingroup$

You're not wrong. Including $X$ in the regression is typically how you condition on a variable in the regression setting. There are other ways to condition (stratify, backdoor adjustment, frontdoor adjustment, and instrumental variables).

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ Just want to add that propensity score methods do this too! $\endgroup$
    – Noah
    Commented Jul 30, 2021 at 1:38
  • $\begingroup$ Is there a rule of thumb indicating which method is superior to another (or maybe given enough context)? I have recently encountered this statement suggesting to ban matching: "Much has been written about [propensity matching and] why it’s terrible and how it will lead to wrong estimates. If you can compute propensity scores, you can use inverse propensity weighting instead." (see towardsdatascience.com/…). $\endgroup$
    – Tanguy
    Commented Jul 30, 2021 at 8:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.