Is there a difference between omitted variable bias and confounding bias in linear models?

To my knowledge, when investigating the causal effect of $X$ on $Y$, a confounder is a variable $Z$ that is causally related to both $X$ and $Y$ with a corresponding dag: $Z\rightarrow X\rightarrow Y \leftarrow Z $

But why does the OMV, commonly derived as $\hat{\beta} = \beta + γ\cdot cov(X,Z)/var(X) = \beta + γ\kappa$, consist of the effect γ of the regression of $Y$ on $Z,X$ and the effect $\kappa$ of $Z$ on $X$ instead of $X$ on $Z$.

Edit, spelling out the classic OMV example and cleaning notation/correcting mistakes:

$Y =\beta X+γZ + ε \\ Y = \hat{\beta}X+ \hat{\epsilon}$

  • 1
    $\begingroup$ How about a little more context for that last sentence? What is the modelling framework? What are the b parameters? $\endgroup$ Nov 13, 2020 at 17:47
  • 1
    $\begingroup$ You might also be interested in the following paper by Cinelli/Hazlett (2020):doi.org/10.1111/rssb.12348. They extend the idea of OMV for sensitivity analyses; the paper is also helpful to grasp the general concept. $\endgroup$
    – persephone
    Nov 16, 2020 at 17:21

1 Answer 1


Omitted variable bias (OVB) is agnostic to the causal relationship between $X$ and $Z$. It concerns only the ability to estimate $\tau$ in the structural model for $Y$. The joint distribution of $Y$, $X$, and $Z$ is compatible both with a data-generating process in which $Z$ is a confounder of the $X \rightarrow Y$ relationship, so that $\tau$ represents the total effect of $X$ on $Y$, and with a data-generating process in which $Z$ is a mediator of the $X \rightarrow Y$ relationship, so that $\tau$ represents the direct effect of $X$ on $Y$.

In the confounding model, the data-generating process for $X$ and $Z$ is: $$ Z := \epsilon_Z \\ X := \gamma Z + \epsilon_X $$ In the mediation model, the data-genertaing process for $X$ and $Z$ is: $$ Z := \alpha X + \epsilon_Z \\ X := \epsilon_X $$

For the confounding process, omitting $Z$ from the model for $Y$ yields a biased estimate of $\tau$, the total effect of $X$ on $Y$. Thisis the classic bias due to an omitted confounder.

For the mediation process, the $X \rightarrow Y$ relationship is not confounded. The estimated coefficient $\hat \tau$ in the model omitting $Z$ is unbiased for the total causal effect of $X$ on $Y$. However, it is biased for $\tau$, the direct effect of $X$ on $Y$.

This is all to say that it's possible to have OVB without confounding if the coefficient you are trying to estimate is a direct effect, in which case omitting the mediator yields a biased estimate of this quantity. In the absence of confounding, the model omitting the mediator yields the total effect. The formula for the bias is the same regardless of the data-generating process of $X$ and $Z$, but the interpretation of the biased parameter depends on the causal relationship between $X$ and $Z$.

  • 1
    $\begingroup$ Excellent! I always appreciate your causality posts. $\endgroup$ Nov 16, 2020 at 15:17
  • 1
    $\begingroup$ Thanks, this is a really clear answer. Maybe you could add a little bit of clarification regarding $\kappa$ in the OVB formula and make it extra awesome. Say I know the data-generating process (e.g. im simulating data): Z(0.2)→X(0.4)→Y←(0.8)Z + some normal errors, whereby (x) are the values of the coefficients. If I now estimate the resticted model, why is the bias not 0.2*0.8? Or in the mediation case: X(0.2)→Z(0.8)→Y←(0.4)X, not 0.2*0.8? I thought the total effect would be 0.4 + (0.2*0.8) in this case? $\endgroup$
    – Rob G.
    Nov 16, 2020 at 17:16

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.