Going from ADMB to ADMB-RE

As a user of AD Model Builder you are used to being able to estimate several thousand parameters. When converting your model into a random effects model you may be disappointed about the performance. When turning an init_vector into a random_effects_vector you will most likely experience that the program runs much slower. The reason is that integration of the likelihood (the way ADMB-RE deals with random effects) is more computationally intensive than optimization. In some cases you can exploit special structure in the model (called separability) to reduce the complexity. You can read more about separability in the user manual.

The other warning you should be given is that programs that worded with init_vector's may run into trouble when you change to random_effects_vector's. The problems are related to whether the Laplace approximation is appropriate for the model, and is not related to ADMB-RE per se. The examples in the example collection do not have this kind of problems as these models are only "mildly" nonlinear. We do not yet have much experience with use of ADMB-RE in strongly nonlinear model.

Having said this, it should be stressed that with ADMB-RE you are solving a different (and more difficult) problem than you did with your original program. You have the possibility to estimate variance parameters that you before had to fix.

Updated November 2004