Negative Binomial Regression 0 Problem I am dealing with a count data set.  I am trying to predict the amount of purchasers per day by the amount of previous purchasers for different products, counting the amount of purchases per day.  Therefore, I am using the negative binomial panel regression (panel= different products).  Did I understand  correctly that the negative binomial regression transforms the DV into log form for the analysis?  Then I would need to transform my IV manually into log form as well? 
 A: Answered in comments copied here below:  
The idea is not that negative binomial regression requires transformations of either the response or outcome variable (DV in your abbreviation) or indeed any others. If it did, taking logarithm of zero values would indeed be problematic, as you seem to be hinting in your title. Leave the variables as they arrive unless you have specific grounds for using a predictor (so-called IV) on logged scale; the technique itself does not require that you do any such transformation.  
I have to say that I don't like that translation on the UCLA website of what negative binomial regression does. I'd say that the outcome is predicted by exponentiating a linear combination, although it's a one-step process. That may sound like a distinction without a difference, but it doesn't imply (or is less likely to be read as implying) that there is log transformation at any stage. As said, you have no reason to transform anything just because you are using negative binomial regression. I think you need to read around a bit. I wouldn't rely on websites, but on textbooks.
I agree with Nick's comments. Just two hints which may help while looking for more precise recommendations in textbooks etc.: To check what transformations of your regressors are useful you could try continuity-corrected logs, i.e. log(x + 0.5) instead of just x. Or you could employ a hurdle idea and use I(x > 0) as a regressor (whether or not a product was purchased previously) and additionally log(x) for just the observations with x > 0 (and zero otherwise). This might help you get a better understanding of which regressors are relevant and how they could be incorporated into the model.
