Re: st: About taking log on zero values

[Daniel Feenberg <feenberg@nber.org>]

On Thu, 20 Feb 2014, Maarten Buis wrote:

> On Thu, Feb 20, 2014 at 4:16 PM, Alfonso Sánchez-Peñalver wrote:
> > In any case, if possible the best possible solution would be to estimate the values that ln(sales) would take for those zeros using either a Tobit or a Heckman sample selection model.
>
> This might work if we were talking about an dependent variable (though
> in that case I would just use a glm with a log link). However, we are
> talking about an independent variable. What should happen with the 0s
> depends on the functional form of the relationship between sales (the
> independent variable, which Sebastian wants to log) and y (the
> dependent variable). A Tobit or a Heckman are not tools for
> determining such a functional form, and can thus not be a solution to
> this problem.

If the to be logged variable is on the RHS, then can't you put in an 
arbitrary value instead of log(0), and add a dummy variable which is one 
for that case, otherwise zero? Then you have a functional form that 
encompasses missing/non-missing and has a log relation for all non-zero 
values. The size of the arbitrary value won't actually affect the 
predicted values, T-statistics or regression summary statistics such as 
R**2, so it shouldn't be a source of controversy.

Daniel Feenberg

> -- Maarten
>
> ---------------------------------
> Maarten L. Buis
> WZB
> Reichpietschufer 50
> 10785 Berlin
> Germany
>
> http://www.maartenbuis.nl
> ---------------------------------
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/