tobin model is normally applied when we have censored data, which means that the upper or lower part of the data for y is censored though we may still have data on x. An example can be the income levels, where an upper bound is imposed and income (y) exceeding the upper bound is censored.
Tobit (by the way, Tobit is the name of the economists who pioneered the study of this type of problems) model can also be expressed in terms of a latent variable y*:
yi=yi*, iff yi*>0; yi=0 otherwise.
where yi*=xi(beta)+ei, ei is the independently and identically distributed normal error term.
Then the probability of yi=0 can be expressed as:
P(yi=0)=P(yi*<0)=P(ei<-xi(beta))=1-F(xi(beta)). F(.) is the cummulative density function (cdf) of the normal distribution.
Then the sample log likelihood function can be derived (sorry, the formular is really messy and I will not write it).
Truncated data is present, where x is unobserved and y is censored. Both censored and truncated data regressions belong to the limited dependent variable regression.
Another point is that Tobit model can be expressed as the sum of a binary choice model and a trucated regression.