Residuals non-normal

[Vera Tar]

Hi everyone,

I am running into a problem when running an LSDV model. My residuals are non-normal (I have run the jb and the shapiro wilk tests). Probably this is because my dependent variable is highly skewed. The problem is that I cannot take the simple log transformation of the DV, because I have several values of zero, and if I took the log, those observations would go missing. Any suggestions what I could do to fix my model?

Thanks a lot in advance!

[Nick Cox]

A generalised linear model with log link doesn't hinge on all values being positive, just on the idea that the mean outcome is positive.

On the other hand, much hinges on what the zeros mean. For example, if the outcome was tobacco consumption, there could be a large spike of zeros because many people just don't smoke at all. If the data are essentially a mixture of those who do and those who don't, some other kind of model may make more sense.

[Carlo Lazzaro]

Vera:
if the most concerning options is about residual non-normality please note that OLS remains BLUE regardless of the shape the residual distribution takes (see Koop G. https://www.wiley.com/en-as/Introduc...-9780470032701, page 72).

[Vera Tar]

Thanks for both answers, it helped me a lot!