On the Stata blog, Bill Gould has a nice little post about using poisson regression to estimate models where the dependent variable- when the model is linearized- is a log, for example a Mincer earnings model. Remember with poisson, as with the negative binomial, E(y/x) = exp(x’b).
You might think, why bother? But he gives several credible reasons one being that it allows zeros to be included which “logging” will not. A second, more subtle reason is that it is a handier way of predicting the y variable (as opposed to log(y)).
An entirely separate argument for using poisson to estimate models of this form, rather than linear regression, is given by Santos Silva & Tenreyro in their RE Stats paper “The log of gravity” which is also well worth a look.
P.S. If you are trying this at home don’t forget to use White/Huber robust standard errors since the usual ML poisson standard errors will probably be wrong.