# Non-normal observations in regression modelling [duplicate]

I read an article that says the dependent variables in a regression model must be normally distributed. The way i understand it, is that the observations for the regression model must then be normally distributed. Or in other words if i choose sample data from a population then the sample must be normally distributed. But then i read answers to a similar question in another forum and apparently the the observations or dependent variables does not need to be normally distributed, just the error terms after the model is implemented must be normally distributed.

So now i am not sure, does the observations (dependent variables) has to be normally distributed as well as the error terms or is it just the error terms.

• Apr 19, 2014 at 21:39
• The question is clear and has a meaning from the viewpoint of users. If you have n > or = 30, the regression should yield correct results.
– user10619
Apr 20, 2014 at 8:13
• @subhash what evidence is there that this claim is true? Apr 20, 2014 at 9:24
• @Glen_b The large sample theory presumes that a set of 30 or more observation should generally result in a normal distribution. It is a hunch. The convergence ie mean and divergence 9S.D.) are based on central limit theorem and law of statistical regularity. These ideas are hunch-led, may be you are better equipped with statistical nuances.
– user10619
Apr 20, 2014 at 16:07
• @subhash Which large sample theory mentions 30 observations? The central limit theorem talks about $n\to \infty$. What "law of statistical regularity" do you mean? Where can I find it? On what are these hunches you refer to based? Where does a notion that $n$ of at least $30$ "should be sufficient" come from? Apr 20, 2014 at 22:05