Let's say we want to determine the impact of $y_2$ on $y_1$, which are related as follows:

$$y_1 = f(y_2, x_1,e)$$


$$e = g(y_3,x_2, u_1), \space y_2 = h(y_4,x_3, u_2), \space y_3 = m(y_4, x_4,u_3)$$

and the $u$'s are the exogenous errors. Substituting, we have

$$y_1 = f(h(y_4,x_3, u_2),x_1, g(m(y_4, x_4,u_3),x_2, u_1))$$

We can see here that there is a correlation between $h$ and $g$ in $f$. Therefore, if we do not control for $y_3$ or $y_4$ in $f$, we would have endogeneity because then $h$ would be correlated with the error. Is this correct?

In other words, we are looking for $\frac{\partial f}{\partial h}$, but a change in $h$ without holding $y_3$ or $y_4$ constant in $g$ implies that our estimate of $\frac{\partial f}{\partial h}$ is not ceteris paribus; when we adjust $h$, we also adjust $g$.

My main concern here is whether it matters that the correlation between $y_2$ and the error in $f$ occurs through two functions: $g$ and $m$.


1 Answer 1


I think expressing your Structural Causal Model as Causal Graph gives simple answer. The model looks like:

enter image description here

Much clearer view we can obtain removing non-confounding exogenous variables (however this step is not necessary):

enter image description here

We can clearly see here, that $y_4$ is the confounder which creates backdoor path from $y_2$ to $y_1$ through $y_3$ and $e$. In order to block that backdoor path we need to condition for any nonempty set of the variables: $y_4$, $y_3$, $e$ (which I guess is unobservable error).

Therefore: yes, there exist endogeneity problem with identification of the causal effect because of existing confounder. The estimator in the model $y_1 = f(y_2, x_1, e)$ will most likely be biased even if $f$ is assumed correctly.

  • $\begingroup$ Excellent. You confirmed my intuition, as I drew a very similar graph for myself. Now, the question I have is, what would the next step be to resolve this issue? If $y_3$ and $y_4$ are not directly observable and there are no good proxies for them, I would say, at least in theory, a good next step would be finding an instrument for $y_2$. $\endgroup$ Mar 3, 2021 at 1:05
  • 1
    $\begingroup$ Yes, Instrument, (natural) experiment, sole mediator (front-door criterion) or measuring variables that let you close backdoor paths. I do not see any other options, however there are tons of natural-experimental designs. $\endgroup$
    – cure
    Mar 3, 2021 at 9:32
  • $\begingroup$ Ok, but what if we could control for $y_3$ but not $y_4$? Wouldn't that solve the endogeneity issue? $\endgroup$ Mar 4, 2021 at 5:58
  • 1
    $\begingroup$ Yes, controlling for a mediator blocks backdoor path. As I stated in the answer, any nonempty set of those three variables leads to the correct identification of effect of $y_2$ on $y_1$. It could be any of those variables, or more of them. $\endgroup$
    – cure
    Mar 4, 2021 at 17:12
  • $\begingroup$ Thanks for the help. If you have time, I have a related question here that I would love to hear your feedback on: stats.stackexchange.com/questions/512340/… $\endgroup$ Mar 4, 2021 at 23:40

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.