This is pretty hard for me to describe, but I'll try to make my problem understandable. So first you have to know that I've done a very simple linear regression so far. Before I estimated the coefficient, I watched the distribution of my $y$. It is heavy left skewed. After I estimated the model, I was quite sure to observe a left-skewed residual in a QQ-Plot as wel, but I absolutely did not. What might be the reason for this solution? Where is the mistake? Or has the distribution $y$ nothing to do with the distribution of the error term?

  • $\begingroup$ @Aniko gives a nice answer in response to your previous question. $\endgroup$
    – whuber
    Commented May 29, 2011 at 21:17

2 Answers 2


To answer your question, let's take a very simple example. The simple regression model is given by $y_i = \beta_0 + \beta_1 x_i + \epsilon_i$, where $\epsilon_i \sim N(0,\sigma^2)$. Now suppose that $x_i$ is dichotomous. If $\beta_1$ is not equal to zero, then the distribution of $y_i$ will not be normal, but actually a mixture of two normal distributions, one with mean $\beta_0$ and one with mean $\beta_0 + \beta_1$.

If $\beta_1$ is large enough and $\sigma^2$ is small enough, then a histogram of $y_i$ will look bimodal. However, one can also get a histogram of $y_i$ that looks like a "single" skewed distribution. Here is one example (using R):

xi <- rbinom(10000, 1, .2)
yi <- 0 + 3 * xi + rnorm(10000, .7)
hist(yi, breaks=20)
qqnorm(yi); qqline(yi)

It's not the distribution of $y_i$ that matters -- but the distribution of the error terms.

res <- lm(yi ~ xi)
hist(resid(res), breaks=20)
qqnorm(resid(res)); qqline(resid(res))

And that looks perfectly normal -- not only figuratively speaking =)


With reference to the excellent answer from @Wolfgang, here are the plots from his R code:

enter image description here


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.