Dear Khaled,

In addition what Alex has recommended which is mostly recommended, if theory supports you, transform the variables with serious outlier problem into natural logarithm.

OUTLIERS may be verified whether the total counts of outliers are statistically significant? If yes, then you might considered analyze the data under extreme Value Theory (EVT). Outliers are not problem; they are values in a set of observation. Data of any kind should be treated "as they are." let the nature of the data lead to your model selection. Do not pre-select a model before the data. Data should take precedence, not model. A model is a tool for the job. A particular type of tool may not be applicable for all types of jobs. Each data set represents a a type of job waiting to be analyzed. if the outliers counts are significant, work with it. It requires a particular tool of analysis.

You may also winsorize your data at one percent or five percent level to minimize outliers in STATA.

I know of sevreral researchs where outliers were not taken into account. However, when many outliers are te case it may be risky to ignore or even overlook them all.

KInd regards,

Orlando

Thank you paul.

It is wise to try to understand whether those outliers result from data generation process or from a measurement error. You would probably be interested in understanding those outliers at first place. It could be the case that they are unique observations that motivate a search for deeper explanation. (such outliers you would not want to drop out).

Outliers that may bother you are those which may affect the model fit or the predictive power of your estimation or both.

Aguinis, Gottfreson and Joo (2013) wrote a survey paper on outliers with very rich details. You may be interested in reading that paper.