Regarding estimators of variance from a iid sample of size $n$, Karl Ove Hufthammer says Estimates of variance from an iid sample:

if they do have a normal distribution, dividing by n+1 (sic!) will give the lowest mean square error.

Why does it give the lowest mean square error?

What does he mean by "(sic!)"?


  • $\begingroup$ sic is Latin for 'thus', usually in text meaning "exactly as in the original" (even if that looks wrong, or unlikely, or questionable) $\endgroup$ – Glen_b -Reinstate Monica Feb 2 '14 at 2:29
  • $\begingroup$ @Glen_b: I know that meaning of sic, but I don't understand what Karl means in his comment. $\endgroup$ – StackExchange for All Feb 2 '14 at 2:33
  • $\begingroup$ That it's reasonable to ask in comments at the original; it's a clarification of his meaning, not a whole new question. However, I may offer the suggestion that since $n+1$ looks rather like the more usual $n-1$ (or $n$ for that matter), one might be tempted to think he simply made a sign error or something, and sic is perhaps intended to emphasize that, no, even if it might appear possible that he meant $n-1$ or $n$, he really does mean $n+1$, just as originally typed. $\endgroup$ – Glen_b -Reinstate Monica Feb 2 '14 at 2:38
  • $\begingroup$ You may be interested in this paper: tandfonline.com/doi/abs/10.1080/00031305.2012.735209?af=R $\endgroup$ – ekvall Feb 2 '14 at 7:49
  • $\begingroup$ @KarlOskar interesting; if I did it right, the estimator there corresponds to using $d = (n-1)\,+\,2/n$ in the notation of my answer. $\endgroup$ – Glen_b -Reinstate Monica Feb 2 '14 at 22:45


$$s^2_d = \frac{\sum_{i=1}^n\left(x_i-\bar{x}\,\right)^2}{d}$$

The statistic $(n - 1)s_{n-1}^2 / \sigma^2$ follows a $\chi^2_{n-1}$ distribution. A $\chi^2_{n-1}$ has mean $n-1$ and variance $2(n-1)$. Hence $\text{E}(s_{n-1}^2) = \sigma^2$, and $\text{Var}(s_{n-1}^2) = 2\frac{\sigma^4 }{ n - 1}$.

Now $s_d^2 = \frac{n-1}{d} s_{n-1}^2$

$$\text{Bias}(s_d^2) = E(s_d^2)-\sigma^2 = \frac{n-1}{d}\sigma^2 -\sigma^2=\frac{n-1-d}{d}\sigma^2$$

$$\text {Var}(s_d^2) = 2\sigma^4(n - 1) / d^2$$

$$\text{MSE} = \text{Bias}^2 + \text{Var}$$


\begin{eqnarray} \text{MSE}(s_d^2) &=& \left(\frac{n-1-d}{d}\right)^2\sigma^4+ 2\sigma^4(n - 1) / d^2\\ &=&\sigma^4\frac{(n-1-d)^2+2(n-1)}{d^2}\\ &=&\sigma^4\,f(d)\,,\end{eqnarray}

where $f(d)=1+\frac{(n-1)^2 + 2(n-1)-2(n-1)d}{d^2}$

$f(d)$ is at a turning point when $f'(d)=0$

i.e. when $d(-2(n-1))-2((n-1)^2 + 2(n-1)-2(n-1)d) = 0$

which occurs when $d-2=n-1$ ... i.e. when $d=n+1$

Showing that's a minimum rather than a maximum (or indeed a horizontal point of inflexion) is straightforward, but I'll leave it at that for now.

The relevant Wikipedia page does it more generally, getting a formula in terms of the excess kurtosis (which gives the same result in this case). I may incorporate an outline of that derivation at some point.


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.