Currenty, I am working with panel data to estimate spatial spillover effects. 3 of my variables, percentage of forest area, share of diesel cars and share of people working in research and development, show skewness more than 1. I am not sure whether I should transform them using log or other transformation or it is better to keep them as they are?
Also, one of variable, which is cement production have a lot of zero. But this is meaningful because I look at how cement production in one region impacts pollution in another region and it is clear that some region does not produce cement at all. Should I transform it as it also shows high skewness?
Thank you in advance for your recommendations!
skewness means the process is persistent at some value, usually its a risk to transform data, for skewness use regression modelling, because for random data skewness is close to zero, for others, its just a formula in the field
Ivan Avila Caceres Thank you very much for your answer. I am estimating spatial spillover effects of cement production and I have some variables which are highly skewed. I have read similar papers and they wrote that in order to eliminate estimation errors caused by hetero scedasticity and collinearity, we performed logarithmic processing on non-ratio variables. That is why I was a bit confused which criteria should be used to decide whether to transform variables or not.
50 years ago it was usual to use log function, but the error is affected there, if you want to use log, instead use regression of Y~e^x. its different from log(y)~X in the thing we don't apply log on data, just write on software lm(Y~exp(x)) generating X. the difference is on the scientific side: using log(Y) or just transforming data can create misconceptions at the deduction, meanwhile creating or "simulating" data has not so much problem in respect to the variable of study. a book like statistiscal simulations allows to generate random data for exp(X). probably is more work. but this is comparing with technics from 50 years ago. good day
When percentage or ratio variables are highly skewed, best practices often involve applying transformations such as log, square root, or Box-Cox to normalize distribution, especially if the data violates assumptions of parametric models.
There is no requirement that the dependent or explanatory variables in a regression should not be skewed. Econometrics makes no such demands.
You need to transform your variables if your underlying theory specifies that they should be changed.
When you complete model specification tests, there may be indications of misspecification that might require the transformation of some variables. Such transformation must be consistent with your economic theory.
I stress that you must start with a sensible economic model. You can then provide appropriate explanations of your estimates