I am creating a multiple Linear Regression model using a Dependent variable which is extremely positively-skewed: My skewness statistic is 27.610, std. deviation is 2832.139, range of values from 0 - 100,000 in a population of 234,270 (over 95% of my cases have values at or below 500.00).

Is splitting the file by the DV in order to reduce skewness and then modeling separately a reasonable course of action? Are there rules or guidelines for splitting the DV I should follow? For example, taking a certain number or percentage by Std. Deviation to inform where I should make the split?

Note: I'm new at this - be gentle!

  • 2
    $\begingroup$ Could you provide a bit more information about your problem: What is the dependent variable, how is it generated? What are the independent variables? Why do you think that a skewed DV would be a problem? Note that multiple linear regression makes no assumption about the distribution of the dependent variable. $\endgroup$ Oct 2, 2013 at 20:05
  • 1
    $\begingroup$ (... at least about its unconditional distribution). In some circumstances you might make an assumption about its conditional distribution (/the distribution of the errors). Depending on what your DV is and what you're trying to achieve, there are a number of possibilities (e.g. perhaps GLMs would suit). I would NOT advise splitting the DV. $\endgroup$
    – Glen_b
    Oct 2, 2013 at 23:22
  • $\begingroup$ @COOLSerdash: My data contains details of financial transactions and some demographic data about customers. It is all generated "in-house," ie, usually supplied by the customers themselves. All independent variables are either continuous or binary numerics. The DV represents customer's largest transaction with us. We are trying to predict their next transaction amount. I assumed the skewness of the DV would be problematic because customers at the bottom are very different from those at the top. $\endgroup$
    – user30257
    Oct 3, 2013 at 18:33

1 Answer 1


First, as @COOSerdash notes, multiple linear regression makes no assumption about the dependent variable. It assumes that the errors (as estimated by the residuals) are normal. So, first, test the model and look at the residuals.

Second, splitting the DV is unlikely to be the best idea. If the residuals turn out to be non-normal (which seems likely) then there are other possibilities: 1) Transform the DV (e.g. by taking square root or tenth roots or some such). 2) Use a nonlinear model 3) Use a robust regression method (to name 3).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.