The context of this question is ordinary least squares. $X$ denotes the design matrix.

I would like a proof of the claim – or a corrected version thereof – made in this other question that the exogeneity assumption $E[\epsilon|X] = 0$ implies that $E\!\left[h(X)^T\epsilon\right] = 0$ for any function $h$ on $X$. (By "any function" I guess what is meant is an arbitrary function $\mathbb{R}^{n\times p}\to \mathbb{R}^{n\times p}$, where $n\times p$ is the shape of the matrix $X$.)

The claim for the special case $h(X)=X$ says that the errors are uncorrelated with the regressors, which we can be prove using the expected value of the product of two random variables:

\begin{align} E\!\left[X^T\epsilon\right] &= E_X\!\left[\vphantom{\sum}X^T E[\epsilon|X]\right] \\ &= E_X\!\left[X^T 0\right] \\ &= 0 \end{align}

However, if we try the same proof for the general case, then we seem to run into trouble:

\begin{align} E\!\left[h(X)^T\epsilon\right] &= E_{h(X)}\!\left[\vphantom{\sum}h(X)^T E[\epsilon|h(X)]\right] \\ &\overset{?}{=} E_{h(X)}\!\left[\vphantom{\sum}h(X)^T E[\epsilon|X]\right] \\ &= E_{h(X)}[h(X)^T 0] \\ &= 0 \end{align}

Does the equality $E[\epsilon|h(X)] = E[\epsilon|X]$ hold if and only if $h$ is injective?


$E[h(X)^T\epsilon] = 0$ holds for any $h(X)$ because $h(X)$ can be treated as a constant when conditional on $X$. Using law of iterated expectations:

$E[h(X)^T\epsilon] = E[E[h(X)^T\epsilon|X]] = E[h(X)^TE[\epsilon|X]] = 0$


This result is a consequence of the law of iterated expectation. Applying that law for any function $h$ gives:

$$\begin{align} \mathbb{E}(h(X)^\text{T} \epsilon) &= \mathbb{E} \Big( \mathbb{E}(h(X)^\text{T} \epsilon | X ) \Big) \\[6pt] &= \mathbb{E} \Big( h(X)^\text{T} \mathbb{E}(\epsilon | X) \Big) \\[6pt] &= \mathbb{E} \Big( h(X)^\text{T} \mathbf{0} \Big) \\[6pt] &= \mathbb{E} (\mathbf{0}) = \mathbf{0}. \\[6pt] \\[6pt] \end{align}$$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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