I have investigated two different forward models for image formation in transmission electron microscopy of thick specimens, the 3DCtf model, which introduces a defocus gradient in the linear approximation, and the multislice model. An important result is that the 3DCtf model does not seem to be compatible with the multislice image formation model. A second very useful finding is that the exit wave in the multislice model has an imaginary part, which, in first-order approximation, is a pure projection of the specimen and is not affected by the defocus gradient. The defocus gradient only comes into play in real valued and higher-order imaginary terms. If the multislice model is closer to reality than the 3DCtf-model, then the best way to retrieve the specimen projection for thicker specimens should be a procedure for retrieving the exit wave's imaginary term, for example using images recorded at different defocus values.