I have been taught the following derivation of the tobit model from a latent underlying variable model:


  • $y*=$ latent variable (e.g. utility)
  • $y=max[0,y*]$

And $$y* = \mathbf X\beta + u,\qquad u|\mathbf X \sim Normal(0,\sigma^2) $$

Where the intercept has been absorbed into $\mathbf X_i$ for notational simplicity.

According to lectures and the textbook (Wooldridge, 2012, p. 597), the density of $y$ given $\mathbf X$ is:

$$ P(y=0| \mathbf X )\quad = \quad P(y*<0 \ | \ \mathbf X) $$ $$ \qquad \qquad \qquad \qquad \quad = \quad P( \mathbf X \beta \ + \ \epsilon<0 \ | \ \mathbf X) $$ $$\qquad \qquad \qquad \quad \ = \quad P( \epsilon<- \mathbf X \beta \ | \ \mathbf X)$$

$$ \qquad \qquad \qquad \quad \ = \quad P \left( \frac{\epsilon}{\sigma} < - \frac{\mathbf X \beta}{\sigma} \ | \ \mathbf X \right) $$

$$\qquad \qquad \qquad \quad \quad = \quad \phi \left(- \frac{\mathbf X \beta}{\sigma} \right), \ \frac{\epsilon}{\sigma} \sim N(0,1) $$

I understand the derivation up to here. According to the lecture, it then follows that the density of y given x is:

$$ \frac{1}{\sqrt{2\pi \sigma^2}} e^{-(y-\mathbf X \beta)^2/2\sigma^2) }$$

How is this implied by the preceding line? In my mind, if the standard normal function is given by:

$$ \phi(v) = \frac{1}{\sqrt{2\pi}} e^\frac{-v^2}{2} $$

where $v = \frac{\epsilon}{\sigma} = \frac{y- \mathbf X \beta}{\sigma} $

wouldn't this make the density: $$ \phi \left(\frac{y- \mathbf X\beta}{\sigma} \right) = \frac{1}{\sqrt{2\pi}} e^\frac{-(y- \mathbf X \beta)^2}{2 \sigma^2} $$

I apologize for the shoddy formatting, first time trying MathJax.

Thank you!


I think your notations are a little bit confusing. I think $\phi \left(- \frac{\mathbf x \beta}{\sigma} \right)$, should be$ ...P(y*<0 \ | \ \mathbf x)... =\Phi \left(- \frac{\mathbf x \beta}{\sigma} \right)$ which is CDF.

Above CDF is for $y=0$, note since $y=max[0,y*]$, $y$ only can be $\ge 0$

For $y>0$ then $y=y*$. The pdf of $y$ can be get from $y*$

Since $y* = \mathbf x\beta + u,\qquad u|\mathbf x \sim Normal(0,\sigma^2)$. (Note the book does not use captial $X$)

Therefore $y=y*\sim N(x\beta,\sigma^2),y>0$ i.e $(x\beta+u)\sim N(x\beta,\sigma^2)$

The pdf of $y*$ is

$$\frac{1}{\sqrt{2\pi \sigma^2}} e^{-(y-\mathbf X \beta)^2/2\sigma^2) }$$

which is also pdf of $y$ for $y>0$

Also your last formula for change of variable is not correct, you missed Jacobian $\frac{1}{\sigma}$, I think.


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.