6
$\begingroup$

At Mathematics Stack Exchange, user940 provided a general formula to calculate the variance of the sample variance based on the fourth central moment $\mu_4$ and the population variance $\sigma^2$ (1):

$$\text{Var}(S^2)=\frac{\mu_4}{n}-\frac{\sigma^4(n-3)}{n(n-1)}$$

I would be interested in an unbiased estimator for this, without knowing the population parameters $\mu_4$ and $\sigma^2$, but using the fourth and second sample central moment $m_4$ and $m_2$ (or the unbiased sample variance $S^2=\frac{n}{n-1}m_2$) instead.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
10
$\begingroup$

The question is to find an unbiased estimator of:

$$\text{Var}(S^2)=\frac{\mu_4}{n}-\frac{(n-3)}{n(n-1)} {\mu_2^2}$$

... where $\mu_r$ denotes the $r^\text{th}$ central moment of the population. This requires finding unbiased estimators of $\mu_4$ and of $\mu_2^2$.

An unbiased estimator of $\mu_4$

By defn, an unbiased estimator of the $r^\text{th}$ central moment is the $r^\text{th}$ h-statistic: $$\mathbb{E}[h_r] = \mu_r$$ The $4^\text{th}$ h-statistic is given by:

enter image description here where:

i) I am using the HStatistic function from the mathStatica package for Mathematica

ii) $s_r$ denotes the $r^\text{th}$ power sum $$s_r=\sum _{i=1}^n X_i^r$$

Alternative: The OP asked about finding an unbiased solution in terms of sample central moments $m_r=\frac{1}{n} \sum _{i=1}^n \left(X_i-\bar{X}\right)^r$. An unbiased estimator of $\mu_4$ in terms of $m_i$ is:

enter image description here

An unbiased estimator of $\mu_2^2$

An unbiased estimator of a product of central moments (here, $\mu_2 \times \mu_2$)is known as a polyache (play on poly-h). An unbiased estimator of $\mu_2^2$ is given by:

enter image description here

where:

i) I am using the PolyH function from the mathStatica package for Mathematica

ii) For more detail on polyaches, see section 7.2B of Chapter 7 of Rose and Smith, Mathematical Statistics with Mathematica (am one of the authors), a free download of which is available here.

While mathStatica does not have an automated converter to express PolyH in terms of sample central moments $m_i$ (nice idea), doing that conversion yields:

enter image description here

Putting it all together:

An unbiased estimator of $\frac{\mu_4}{n}-\frac{(n-3)}{n(n-1)} {\mu_2^2}$ is thus:

enter image description here

or, more compactly, in terms of sample central moments $m_i$:

........... enter image description here

And as a check, we can run the expectations operator over the above (the $1^\text{st}$ RawMoment of sol), expressing the solution in terms of Central moments of the population:

enter image description here

... and all is good.

Improve this answer
$\endgroup$
6
  • $\begingroup$ Great answer! Thanks! Two things: 1. It would be great if the estimator of $\mu_2^2$ could also be expressed in terms of $m_r$. 2. I'm often encountering terms such as $(n-1)(n-2)(n-3)\ldots$ in the denominator when unbiased quantities are involved. For instance, the unbiased estimator of $\mu_4$ that you have presented has $(n-1)(n-2)(n-3)$ in the denominator. Another example is Bessel's correction with $(n-1)$ in the denominator. Can I generally assume that any unbiased estimator of $\mu_r$ is only defined for $n \geq r$? Is there a theorem for that? $\endgroup$
    – Hiro
    Commented Oct 13, 2017 at 6:51
  • 1
    $\begingroup$ Yes - the solutions assume the sample size $n \ge w$, where $w$ denotes the weight of the expression. For instance, an expression containing $\mu_2^3$ has weight = 6. It also assumes, of course, that the moments of the parent random variable exist up to that weight. I have updated the answer above to provide the unbiased estimator of $\mu_2^2$ in terms of $m_i$ which is indeed neater. $\endgroup$
    – wolfies
    Commented Oct 13, 2017 at 15:53
  • $\begingroup$ Thanks for the explanation! However, I think that there is an issue with the unbiased estimator of $\mu_2^2$ in terms of $m_i$. I have checked it in Mathematica by calculating the average of multiple estimators and the result is quite off. It would be great if you could take a look at it again! $\endgroup$
    – Hiro
    Commented Oct 16, 2017 at 7:36
  • 1
    $\begingroup$ @Hiro - checked and seems fine to me. I have further checked by converting the expression in sample central moments back to power sums which can be done with expr /. Subscript[m, i_] :> SampleCentralToPowerSum[i][[2]] // Simplify which returns the expression in power sums above. If you can post your (simulation?) code, I will be happy to look at what you are doing. $\endgroup$
    – wolfies
    Commented Oct 17, 2017 at 15:26
  • $\begingroup$ I have checked my code and noticed that I've made a mistake that introduced the bias. After correcting it, everything works great! :-) Sorry for that and again, thank you very much! $\endgroup$
    – Hiro
    Commented Oct 18, 2017 at 13:41

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.