4
$\begingroup$

I am having trouble following a short derivation that uses the Law of Iterated Expectations that is found in the answer to another question: How to derive a regression formula

I will repeat it here:

Let $E(y|z) = \mu_{y|z}.$ Then it is shown that $E(y \mu_{y|z}) = Var(\mu_{y|z})$ in the following steps:

(1) $E(y \mu_{y|z}) = E(E(y|z, \mu_{y|z}) *\mu_{y|z})$

(2) $~~~~~~~~~~~~~= E(E(y|z) *\mu_{y|z})$

(3) $~~~~~~~~~~~~~= E(\mu^2_{y|z})$

(4) $~~~~~~~~~~~~~= Var(\mu^2_{y|z})$

I don't know all the properties of the LIE, but I do know that in general it gives $E(W) = E_Z(E(W|Z))$. With that said:

Q1. In line 1, what is going on? How is the LIE being applied in this way?

Q2. In going from line 1 to line 2, why do we not condition on $\mu_{y|z}$ anymore?

Q3. In going from line 3 to line 4, why is $(E(\mu_{y|z}))^2 = 0$ so that we get the variance?

$\endgroup$

1 Answer 1

4
$\begingroup$

By the defining property of the conditional expectation,

$$E\big[E(W\mid Q)\big] = E(W)$$

Set $W \equiv y\cdot E(y\mid z)$ and $Q \equiv z$. Substitute to get

$$E\Big(E\big [y\cdot E(y\mid z) \Big|z\big]\Big)= E\big [y\cdot E(y\mid z)\big] $$

The right-hand-side is what you are starting with, so we need to manipulate the left-hand-side.

By the measurability of $E(y\mid z)$ with respect to $z$ ("take out what is known") we have

$$E\Big(E\big [y\cdot E(y\mid z) \Big|z\big]\Big) = E\Big(E(y\mid z)\cdot E\big [y \mid z\big]\Big)$$

$$=E\left(\left[E(y\mid z)\right]^2\right) = {\rm Var}[E(y\mid z)] +\big(E\left[E(y\mid z)\right]\big)^2$$

$$={\rm Var}[E(y\mid z)] +\big(E[y]\big)^2$$

So for the stated result to go through, it must be the case that the unconditional expected value of $y$ is zero. This is of course not a general result, but it is an assumption explicitly made in Angrist's paper (which is used in the CV thread the OP linked to), where just above his eq. $(2)$ the author writes:

"Let $\beta$ denote the population regression coefficient from a regression of (mean zero) $y$ on $μ_{y|z} = E[y|z]$, for any random variables, $y$ and $z$."

$\endgroup$
1
  • 1
    $\begingroup$ +1 That's a good point Alecos, I had not mentioned this assumtpion in my previous answer to the OP. I will edit the answer accordingly. $\endgroup$
    – Andy
    Commented Dec 1, 2014 at 8:40

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.