MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I compute well-known "sample Pearson correlation coefficient" of two vectors:

$r(X,Y) = \frac{\sum ^n _{i=1}(X_i - \bar{X})(Y_i - \bar{Y})}{\sqrt{\sum ^n _{i=1}(X_i - \bar{X})^2} \sqrt{\sum ^n _{i=1}(Y_i - \bar{Y})^2}}$ . So far so good.

I need to generalize it as follows. I need to add "weights", $w_i\in[0..1]$, to the formula.
The idea is that smaller values of $w_i$ make affect of $(X_i,Y_i)$ on $r$ smaller.
(Proportionally to value of $w_i$.)
Without $w_i$, every $(X_i,Y_i)$ and $(Y_j,Y_j)$ affec $r$ equaly.
$w_i$ makes then affect $r$ differently, with different "weight".

Would it be right to calculate $r(X,Y,w)$ as
$r(X',Y')$ where $X'_i=w_i(X_i-\bar{X})$, and $Y'_i=w_i(Y_i-\bar{Y})$ ? Or else, what would be the
right way to insert $w_i$ into the formula ?

share|cite|improve this question
First of all - what do you need the generalization for? What kind of quantity you would like to measure, what properties you would like to have, etc... – Piotr Migdal Mar 11 '11 at 18:14
I need it to bevhave just like Person. Except that "older" values affect r less than newer values. I give older values smaller weight = smaller "importance". Does it make sense. For example if all $w_i$ are equal then it would become regular Pearson. – Andrei Mar 11 '11 at 18:57
up vote 1 down vote accepted

I think that the correct thing to do is treat the values as if they were repeated $w_i$ times. That is, assume first that the weights $w_i$ are non-negative integers, the data actually comes in the form such that the pair $(x_i, y_i)$ appears $w_i$ times, and compute the standard Pearson correlation for such (repeated) data. Doing so gives: $\bar{X}' = \frac{\sum_{i=1}^n w_i x_i}{\sum_{i=1}^n w_i}$ (and similarly for $Y$). And in terms of the original $n$ data points, the formula is: $$ r(X,Y,w) = \frac{\sum_{i=1}^n w_i (X_i-\bar{X}')(Y_i-\bar{Y}')}{\sqrt{\sum_{i=1}^n w_i (X_i - \bar{X}')^2} \sqrt{\sum_{i=1}^n w_i (Y_i - \bar{Y}')^2}} $$

and then you can use it of course for general real weights $w_i$. I don't think this is equivalent to your suggestion.

share|cite|improve this answer
Makes sense to me. Thanks. – Andrei Mar 11 '11 at 19:27
@Or zuk: that's exactly how -for instance- SPSS handles that weighting; it's implemented as a standard option in each statistical procedure. Actually I know its use of some studies, where the parameters of a distorted sample were compared with that of a population, where the underrepresentated subsamples are weighted higher and the overrepresented samples were weighted lower such that each subsample/group has the same relative size as in the population. The algorithms of SPSS are openly availabe as far as I recall ( but may be you have to go over the new owner instead (IBM) – Gottfried Helms Mar 11 '11 at 19:31
Yes, should have guessed this is pretty standard - good to know, thanx. I've actually used it lately for a case control genetic study (where the proportion of cases in the sample is different than in the population) – Or Zuk Mar 11 '11 at 19:38

Could this item be what you're looking for?

share|cite|improve this answer
Thanks, this is definitely interesting. – Andrei Mar 15 '11 at 20:28

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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