This topic was covered well in 2010 by Tom Eagle at the 2010 Sawtooth Software Conference. You can read his paper within the following volume:
The simple one-step approach as done by the “tools/scripts” tool you can find on our website appears to work reasonably well for standard types of CBC datasets. But, Tom was able to create some more complex data sets where his two-stage joint discrete/continuous models performed better.
The key issue in all of this is whether respondents can give you good volumetric CBC data. Many researchers with deep experience with this say that respondents often cannot provide very good volumetric data, so there is a healthy deal of skepticism surrounding the volumetric CBC approach.