There seems to be a common use of the uncorrected sample standard deviation in calculating the correlation coefficient:

I'm wondering if I'm missing something about the apparent peferrability of using the corrected sample standard deviation:

I've posted this to:


It makes no difference whatsoever, provided you compute consistently.

Suppose you define your version of the variance of a batch of data $X = x_1, x_2, \ldots, x_n$, with mean $\bar x = (x_1 + x_2 + \cdots + x_n)/n$, to be

$$\text{var}_f(X) = f(n)\left((x_1-\bar x)^2 + (x_2 - \bar x)^2 + \cdots + (x_n - \bar x)^2\right)$$

where $f:\mathbb{N}\to (0,\infty)$ is a personalized function giving some positive value $f(n)$ for each positive integer $n$. Some people might use $f(n) = 1/n$, others might use $f(n) = 1/(n-1)$, and others might use something else altogether such as $f(n)=\Gamma((n-1)/2)^2/(2\Gamma(n/2)^2)$. It doesn't matter, because as we have seen you must use the same function $f$ to compute the covariance of any batch of paired data $(X,Y) = (x_1,y_1), (x_2,y_2), \ldots, (x_n,y_n)$:

$$\text{cov}_f(X,Y) = f(n)\left((x_1-\bar x)(y_1-\bar y) + (x_2-\bar x)(y_2-\bar y) + \cdots + (x_n-\bar x)(y_n-\bar y)\right).$$

Consequently your personal correlation coefficient will be

$$\eqalign{ \rho_f(X,Y) &= \frac{\text{cov}_f(X,Y)}{\sqrt{\text{var}_f(X)}\sqrt{\text{var}_f(Y)}} \\ &= \frac{f(n)\sum_i(x_i-\bar x)(y_i-\bar y)}{\sqrt{f(n)\sum_i(x_i-\bar x)^2}\sqrt{f(n)\sum_i(y_i-\bar y)^2}}\\ &= \frac{\sum_i(x_i-\bar x)(y_i-\bar y)}{\sqrt{\sum_i(x_i-\bar x)^2}\sqrt{\sum_i(y_i-\bar y)^2}}. }$$

Since the factor $f(n)/(\sqrt{f(n)}\sqrt{f(n)}) = 1$ disappears, your correlation coefficient will be the same as anyone else's. As a bonus, now you know you don't have to compute $f$.


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.