Wooldridge Introductory Econometrics: A Modern Approach (2018), pages 561 and 572, gives the following definitions:

Latent variable model (LVM): $$ y^*=\beta_0+\mathbf{x} \boldsymbol{\beta}+e, y=1\left[y^*>0\right] $$ where the indicator function takes the value 1 and if the event in the brackets is true and 0 otherwise. So $y$ is 1 if $y^* > 0$.

Tobit: $$ y^*=\beta_0+\mathbf{x} \boldsymbol{\beta}+u, u \mid \mathbf{x} \sim \operatorname{Normal}\left(0, \sigma^2\right) \\ y=\max \left(0, y^*\right) $$

where $y=\max \left(0, y^*\right)$ implies that the observed variable $y$ equals $y^*$ when $y^* \geq 0$, $0$ otherwise.


So it seems to me that one difference is that $y^*$ in the LVM takes the value 0 or, while in the Tobit model $y^*$ takes the values 0 or any positive values otherwise. Do I have that right?

Are there any other key takeaways w.r.t to the differences between the models?


1 Answer 1


Your understanding is correct.

The latent-variable representation is convenient because you can express various models in this form.

The LVM in your question corresponds to a probit model. If the errors come from a standard logistic (instead of standard normal) distribution, you get a logistic regression model.

If instead of a single threshold at 0 you have multiple cutpoints at $\zeta_1<\zeta_2<\zeta_3<\ldots$, you get an ordinal probit model (or ordinal logistic if you have logistic errors).


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.