0
$\begingroup$

I know there is already at least one question answer with this, but I think the solution does not apply in my case.

I have a population for which I know the mean, variance and skewness. I saw how to estimate the parameters for the skew normal distribution given this data, but the problem is that when calculating $\alpha$, $\delta > 1$ (by wikipedia notation), and thus $\alpha$ is undetermined?

How can I solve this?

$\endgroup$
3
$\begingroup$

By looking at the formulae in the wikipedia page, one can see that if we know skewness, we know $\delta$ (note that skewness is monotonic in $\delta$, so the relation is unique). In turn this determines the sign of $\alpha$.

Also, $$|\delta| = \frac{|\alpha|}{\sqrt{1+\alpha^2}} \implies \delta^2(1+\alpha^2)=\alpha^2 \implies (\delta^2-1)\alpha^2 +\delta^2 = 0$$

So if $\delta \geq1$ this is impossible, not undetermined.

And indeed, this is one of the limitations of the skew normal distribution: $\delta$ should be strictly smaller than unity, which in turn imposes limits on the magnitude of the skewness coefficient

So if someone specifies a skew normal distribution with such a theoretical skew coefficient, that $\delta$ ends up equal or above unity, it is simply wrong.

When the sample skew coefficient is higher than theoretically possible, then one can apply maximum likelihood to estimate $\alpha$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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