1
$\begingroup$

Problem Statement: Suppose we have the following Structural Causal Model (SCM). Assume all exogenous variables ($U$) are independent identically distributed standard normals. \begin{align*} V&=\{X,Y,Z\},\qquad U=\{U_X, U_Y, U_Z\},\qquad F=\{f_X, f_Y, f_Z\}\\ f_X: X&=U_X\\ f_Y: Y&=\frac{X}{3}+U_Y\\ f_Z: Z&=\frac{Y}{16}+U_Z. \end{align*} Determine the best guess of $Y,$ given that we observed $X=1$ and $Z=3.$ [Hint: you may wish to use the technique of multiple regression, together with the fact that, for every three normally distributed variables, say $X, Y,$ and $Z,$ we have $E[Y|X=x,Z=z]=R_{Y\!X\cdot Z}x+R_{Y\!Z\cdot X}z.$]

My Answer: We assume the model $$Y=\alpha+\beta_XX+\beta_ZZ+\varepsilon,$$ with $\beta_X$ and $\beta_Z$ given by \begin{align*} \beta_X=R_{Y\!X\cdot Z} &=\frac{\sigma_Z^2\sigma_{Y\!X}-\sigma_{Y\!Z}\sigma_{Z\!X}}{\sigma_X^2\sigma_Z^2-\sigma_{X\!Z}^2}\\ \beta_Z=R_{Y\!Z\cdot X} &=\frac{\sigma_X^2\sigma_{Y\!Z}-\sigma_{Y\!X}\sigma_{X\!Z}}{\sigma_Z^2\sigma_X^2-\sigma_{Z\!X}^2}. \end{align*} Now here's where things get a little hazy. Going off the model, and the idea that if $C=aA+bB,$ then $\sigma_C^2=a^2\sigma_A^2+b^2\sigma_B^2,$ we have the following equations: \begin{align*} \sigma_X^2&=\sigma_{U_X}^2=1\\ \sigma_Y^2&=\frac19 \sigma_X^2+\sigma_{U_Y}^2=\frac{10}{9}\\ \sigma_Z^2&=\frac{1}{256}\sigma_Y^2+\sigma_{U_Z}^2=\frac{1157}{1152}. \end{align*} Here's where I have trouble: computing the covariances. I know that, for example, $$\sigma_{XY}=E(XY)-\underbrace{E(X)}_{=0}E(Y)=E(XY)$$ in this case, since $E(X)=0.$ Then I substitute in to obtain \begin{align*} E(XY) &=E(X(X/3+U_Y))\\ &=\frac13 E\big(X^2\big)+\underbrace{E(XU_Y)}_{=0}\\ &=\frac13\left(\sigma_X^2-(E(X))^2\right)\\ &=\frac13. \end{align*} A similar calculation reveals that \begin{align*} \sigma_{XZ}&=\frac{1}{48}\\ \sigma_{YZ}&=\frac{5}{72}. \end{align*}

My question is: are these calculations correct so far? If so, I think I can make my way to the end by plugging into the expressions above for the regression coefficients, and then the model.

Thank you for your time!

$\endgroup$
5
  • $\begingroup$ Just for my edification, what does the notation $R_{YX \cdot Z}x$ means? $\endgroup$ Jan 8 at 15:57
  • 1
    $\begingroup$ Sure! $R_{YX\cdot Z}$ means the slope of $Y$ on $X$ when we hold $Z$ constant; it's also known as a partial regression coefficient. See Causal Inference in Statistics: A Primer, p. 23. $\endgroup$ Jan 8 at 16:19
  • $\begingroup$ Got it, thanks! So, question then. Shouldn't the formula be $E[Y|X,Z]=y_0+R_{YX \cdot Z}x+R_{YZ \cdot X}z$ where $y_0=E[Y|X=0,Z=0]$? $\endgroup$ Jan 8 at 16:29
  • 1
    $\begingroup$ Well, the formula I typed up in the hint is exactly what Pearl has. If you notice in my answer, I allow an offset such as you suggest. In the context of this problem, I doubt it matters, as I think the $y_0$ you've got there would be zero, given that the exogenous variables are all standard normals. $\endgroup$ Jan 8 at 20:24
  • $\begingroup$ Ah, yes, good point. I did not consider that - I suppose, in general, its just an offset. Thanks! $\endgroup$ Jan 8 at 20:50
1
$\begingroup$

I believe that given that $X=1$, then $E(X)$ should be 1, not 0. I understand the confusion once $X=U_x$ and $E(U_x)=0$. But the previous only state that the value of $X$ is caused by the value of $U_x$, in this case the value of X is already observed to be 1, hence a constant $c$, and because $E(c)=c$ then $E(X)=1$. $E(Y|X=1)=E(1/3 +U_y)=1/3$.

Note that $\sigma_{xy}=0$ does not imply that X and Y are independent.

$\endgroup$
0
$\begingroup$

I was, indeed, on the right track. To finish the calculations: \begin{align*} \sigma_X^2&=1\\ \sigma_Y^2&=10/9\\ \sigma_Z^2&=1157/1152\\ \sigma_{XY}&=1/3\\ \sigma_{YZ}&=5/72\\ \sigma_{XZ}&=1/48\\ \\ E[Y|X=1,Z=3]&=400/771. \end{align*}

$\endgroup$

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.