5

Here is an argument I came across: by limiting your sample by some range of one of the variables the (Pearson) correlation coefficient between the two variables is likely to be reduced.

I can't see the logic. I would think it could both increase or reduce, depending on whether the linear relationship is stronger or not for that interval at hand.

Any clues?

snoram
  • 437
  • 3
  • 11
  • 1
    However, *"is likely to"* does not imply "*will always*." – Penguin_Knight Mar 19 '15 at 16:19
  • Of course! But it implies more likely than not, which I have hard time accepting. – snoram Mar 19 '15 at 16:20
  • 2
    "Limiting your sample" sounds like *truncation.* Censoring means that values in the sample are *replaced* by censoring limits. The effect on the nominal correlation coefficient (as incorrectly calculated by replacing each censored value by its censoring limit) is profoundly different from removing such values altogether (truncation). Which one do you really mean? – whuber Mar 19 '15 at 16:34
  • 1
    I mean "removing such values altogether (truncation)". Thanks, I will change the title. – snoram Mar 19 '15 at 16:36
  • 2
    This phenomenon is illustrated in my answer at http://stats.stackexchange.com/a/13317/919, which slices one variable into quantiles and quotes the $R^2$ (squared correlation coefficient) for the data within each slice. – whuber Mar 19 '15 at 21:41

2 Answers2

7

There's a number of ways to look at it, but this is a pretty straightforward one:

Imagine for a moment we're looking at a regression problem. The squared correlation between the two variables ($r^2$) is $R^2$, the coefficient of determination, which is $1-\frac{s^2_\epsilon}{\text{Var}(y)}$. When you restrict the range of $x$, you also reduce the range of $y$, so $\text{Var}(y)$ goes down with it, while $s^2_\epsilon$ (the noise about the line) should hardly change, since it still has an expected value of $\sigma^2_\epsilon$. Here's an example of that:

full x-y plot and truncated x-range

Since the denominator of the fraction decreases while the numerator hardly changes, the fraction gets larger, so $R^2$ gets smaller, so $r^2(x,y)$ and hence $|r|$ will be smaller. So we really should expect that the size of the correlation decreases.

Glen_b
  • 257,508
  • 32
  • 553
  • 939
4

Thinking of a 2D plot of one variable plotted against the other, limiting the range for one variable means only looking at a vertical or horizontal "slice". So my intuition is that the overall shape of the "cloud" of points will be more vertical or more horizontal, instead of "diagonal". A vertical or horizontal-looking cloud of points has zero correlation. So to me there is indeed an intuition that correlation is likely to decrease.

As a toy example, if your data points are (1,1), (1,20), and (20,20), you have 0.5 correlation, but if you limit the range of the first variable to [0,10] you are left with two points (1,1) and (1,20), and correlation =0. If you limit the second variable to [10,30] then you get two points aligned vertically, and again correlation =0.

user3780968
  • 361
  • 1
  • 4