Lise,
I had the same question, so I fielded an R&D study to find out. I reported this on our LinkedIn discussion group a couple of years ago, but the bottom line is this: I didn't find significant differences in how well the model fit holdout data (as measured by mean absolute errors of prediction).
When I presented this work at one of our conferences, I found the audience about split on how they handled estimation in cases like this, and that they seemed to split more on which method they found more convenient than for any theoretical reasons.
So I think you're safe to proceed as you like.