5
$\begingroup$

I'm trying to find the variance of $L$, $Var(L)$, using the delta method (I want to find a closed form). $L$ is defined as: $$L = \frac{A}{B} + \frac{C}{D}$$ All $A$, $B$, $C$, and $D$ are dependent.

But I'm not sure how to proceed.

$\endgroup$
  • 2
    $\begingroup$ The answer to stats.stackexchange.com/questions/21875/… will get you started; your problem has four variables and will therefore be more complicated, but the core approach is the same. $\endgroup$ – jbowman Apr 22 '12 at 17:03
  • 1
    $\begingroup$ Note, too, that the delta method is an approximation, so please don't overinterpret any "closed form" result! $\endgroup$ – whuber Apr 22 '12 at 21:06
6
$\begingroup$

For notational simplicity define $X_1=A$, $X_2=B$, $X_3=C$, $X_4=D$ and $L=f(X_1,X_2,X_3,X_4)$ with $f(x_1,x_2,x_3,x_4) = \frac{x_1}{x_2}+\frac{x_3}{x_4}$. A first order expansion around $\mu = (\mu_1, \mu_2,\mu_3, \mu_4)$ where $\mu_i=\mathbb{E}[X_i]$ shows that the following approximation holds $$L \approx f(\mu) + \sum_{i=1}^4 \partial_i f(\mu) \, (X_i-\mu_i)$$ as soon as the quantities $X_i-\mu_i$ are small enough. Since $\mathbb{E}\big[\sum_{i=1}^4 \partial_i f(\mu) \, (X_i-\mu_i) \big]=0$ it thus follows that the variance of $L$ can also be approximated by $$\textrm{var}(L) \approx \mathbb{E} \Big[ \Big\{\sum_{i=1}^4 \partial_i f(\mu) \, (X_i-\mu_i) \Big\}^2 \Big] \\ = \sum_{1 \leq i,j \leq 4} \textrm{Cov}(X_i,X_j) \partial_i f(\mu) \partial_j f(\mu).$$

$\endgroup$

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.