7
$\begingroup$

I am working on a bivariate probit model and I want to calculate the gradient. In order to do that I have to take the devirative w.r.t. $\beta$ in the following expression: $$\int_{-\infty}^{x_{g1}\beta} \int_{-\infty}^{x_{g2}\beta}\phi_2(z_1, z_2, \rho)dz_2dz_1$$ Could somebody give me a hint how to do it? I tried to use differentiation under the integral sign, however, I could not get the required result.

$\endgroup$
  • $\begingroup$ I know little about this other than that probit models are almost always simulated for the precise reason that these integrals are very difficult to calculate. I believe Kenneth Train's book would be a good place to start getting a perspective on this. $\endgroup$ – gregmacfarlane Jan 24 '13 at 21:29
  • $\begingroup$ Do you want to calculate the derivatives on the log likelihood? $\endgroup$ – Dimitriy V. Masterov Jan 25 '13 at 2:06
  • $\begingroup$ Yes, I want to calculate the derivative of the log-likelihood function. I didn't write the entire expression, since I have a problem only with this part. $\endgroup$ – Kolibris Jan 25 '13 at 10:15
  • 1
    $\begingroup$ That expression is a function $f(\beta)$. The numerator of the difference quotient, $df(h)=f(\beta+h)-f(\beta)$, is an integral over two semi-infinite strips whose widths are proportional to $h$. Assuming suitable continuity properties of $\phi_2$ in neighborhoods of these strips, the limiting value of $df(h)/h$ as $h\to 0$ will therefore be a linear combination of integrals over the two corresponding rays: one extending to the left from $(x_{g1}\beta, x_{g2}\beta)$ and the other down from that point. With this geometric image in mind you can obtain the derivative with almost no computation. $\endgroup$ – whuber Apr 28 '14 at 20:18
1
$\begingroup$

Let's take a step back and solve a simpler problem - how do derivatives with respect to a variable in the limits of the integral work, at least when everything is sufficiently nice? Let's take a very basic approach:

$\frac{d}{dz} \int_{a}^{h(z)} g(x) \,dx$

Let $\frac{dG}{dz}=g(z)$.

$=\frac{d}{dz} [G(h(z))-G(a)]=\frac{d}{dz} G(h(z))=h'(z)\,g(h(z))$

The same approach should be sufficient for your problem.

(whuber's approach gets you there faster though)

$\endgroup$

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.