We highlight a subtle identification problem that afflicts models for non-negative data in which the conditional expectation is specified as the product of a logit and an exponential function. The results are illustrated with an empirical model for medical expenditures.