# Should I use log transformation for my linear regression? [duplicate]

I am modeling commissions earned based on non sales data. My equation looks like this:

commissions = Number of times logged into the system + ...

(many binary variables for HR related info)

Realistically, there would be a peak login, or plateau where any more logins is not correlated to commissions. So should I take the natural log of this variable?

• Without some further diagnostics (eg. $\beta$'s conf. intervals, plots of observed vs fitted, residual vs fitted, $R^2$, etc.) it is a bit hard to say. Please provide some additional diagnostics first. In general, have you thought of a GLM with Gamma or Inverse Gaussian? – usεr11852 Aug 24 '14 at 3:06
• I have no idea what those models are. (I have only taken stats 101). Are you saying that I cannot rely on intuition to make this decision? It feels reasonable to me that the shape of a log function would fit my variable. Therefore if I take the log of it, it will linearize my variable and be a better representation of reality. I am ok with discarding intuition in favor of more systematic diagnostics, but it is an awkward pill to swallow. – Jay Aug 24 '14 at 3:22
• Your intuition could be correct, I do not know. Nevertheless, given that you admit not having substantial experience on this field I urge you to follow standard procedures and not rely on your intuition immediately. Even in a Stats101 course the importance of plotting your data should have been stressed anyway. The models I proposed are specifically designed to model dependent variables that are (any) positive numbers. – usεr11852 Aug 24 '14 at 3:37