Correlation

In statistics, correlation or dependence is any statistical relationship, whether causal or not, between two random variables or bivariate data. Although in the broadest sense, "correlation" may indicate any type of association, in statistics it usually refers to the degree to which a pair of variables are linearly related. Familiar examples of dependent phenomena include the correlation between the height of parents and their offspring, and the correlation between the price of a good and the quantity the consumers are willing to purchase, as it is depicted in the so-called demand curve.

Correlations are useful because they can indicate a predictive relationship that can be exploited in practice. For example, an electrical utility may produce less power on a mild day based on the correlation between electricity demand and weather. In this example, there is a causal relationship, because extreme weather causes people to use more electricity for heating or cooling. However, in general, the presence of a correlation is not sufficient to infer the presence of a causal relationship (i.e., correlation does not imply causation).

Formally, random variables are dependent if they do not satisfy a mathematical property of probabilistic independence. In informal parlance, correlation is synonymous with dependence. However, when used in a technical sense, correlation refers to any of several specific types of mathematical relationship between the conditional expectation of one variable given the other is not constant as the conditioning variable changes; broadly correlation in this specific sense is used when ${\displaystyle E(Y|X=x)}$ is related to ${\displaystyle x}$ in some manner (such as linearly, monotonically, or perhaps according to some particular functional form such as logarithmic). Essentially, correlation is the measure of how two or more variables are related to one another. There are several correlation coefficients, often denoted ${\displaystyle \rho }$ or ${\displaystyle r}$, measuring the degree of correlation. The most common of these is the Pearson correlation coefficient, which is sensitive only to a linear relationship between two variables (which may be present even when one variable is a nonlinear function of the other). Other correlation coefficients – such as Spearman's rank correlation – have been developed to be more robust than Pearson's, that is, more sensitive to nonlinear relationships.^[1][2][3] Mutual information can also be applied to measure dependence between two variables.

The most familiar measure of dependence between two quantities is the Pearson product-moment correlation coefficient (PPMCC), or "Pearson's correlation coefficient", commonly called simply "the correlation coefficient". It is obtained by taking the ratio of the covariance of the two variables in question of our numerical dataset, normalized to the square root of their variances. Mathematically, one simply divides the covariance of the two variables by the product of their standard deviations. Karl Pearson developed the coefficient from a similar but slightly different idea by Francis Galton.^[4]

A Pearson product-moment correlation coefficient attempts to establish a line of best fit through a dataset of two variables by essentially laying out the expected values and the resulting Pearson's correlation coefficient indicates how far away the actual dataset is from the expected values. Depending on the sign of our Pearson's correlation coefficient, we can end up with either a negative or positive correlation if there is any sort of relationship between the variables of our data set.^{[citation needed]}

The population correlation coefficient ${\displaystyle \rho _{X,Y}}$ between two random variables ${\displaystyle X}$ and ${\displaystyle Y}$ with expected values ${\displaystyle \mu _{X}}$ and ${\displaystyle \mu _{Y}}$ and standard deviations ${\displaystyle \sigma _{X}}$ and ${\displaystyle \sigma _{Y}}$ is defined as:

$${\displaystyle \rho _{X,Y}=\operatorname {corr} (X,Y)={\operatorname {cov} (X,Y) \over \sigma _{X}\sigma _{Y}}={\operatorname {E} [(X-\mu _{X})(Y-\mu _{Y})] \over \sigma _{X}\sigma _{Y}},\quad {\text{if}}\ \sigma _{X}\sigma _{Y}>0.}$$

where ${\displaystyle \operatorname {E} }$ is the expected value operator, ${\displaystyle \operatorname {cov} }$ means covariance, and ${\displaystyle \operatorname {corr} }$ is a widely used alternative notation for the correlation coefficient. The Pearson correlation is defined only if both standard deviations are finite and positive. An alternative formula purely in terms of moments is:

$${\displaystyle \rho _{X,Y}={\operatorname {E} (XY)-\operatorname {E} (X)\operatorname {E} (Y) \over {\sqrt {\operatorname {E} (X^{2})-\operatorname {E} (X)^{2}}}\cdot {\sqrt {\operatorname {E} (Y^{2})-\operatorname {E} (Y)^{2}}}}}$$

Consider the joint probability distribution of X and Y given in the table below.

More information ...

${\displaystyle \mathrm {P} (X=x,Y=y)}$

y x	−1	0	1
0	0	⁠1/3⁠	0
1	⁠1/3⁠	0	⁠1/3⁠

Close

For this joint distribution, the marginal distributions are:

${\displaystyle \mathrm {P} (X=x)={\begin{cases}{\frac {1}{3}}&\quad {\text{for }}x=0\\{\frac {2}{3}}&\quad {\text{for }}x=1\end{cases}}}$

${\displaystyle \mathrm {P} (Y=y)={\begin{cases}{\frac {1}{3}}&\quad {\text{for }}y=-1\\{\frac {1}{3}}&\quad {\text{for }}y=0\\{\frac {1}{3}}&\quad {\text{for }}y=1\end{cases}}}$

This yields the following expectations and variances:

${\displaystyle \mu _{X}={\frac {2}{3}}}$

${\displaystyle \mu _{Y}=0}$

${\displaystyle \sigma _{X}^{2}={\frac {2}{9}}}$

${\displaystyle \sigma _{Y}^{2}={\frac {2}{3}}}$

Therefore:

${\displaystyle {\begin{aligned}\rho _{X,Y}&={\frac {1}{\sigma _{X}\sigma _{Y}}}\mathrm {E} [(X-\mu _{X})(Y-\mu _{Y})]\\[5pt]&={\frac {1}{\sigma _{X}\sigma _{Y}}}\sum _{x,y}{(x-\mu _{X})(y-\mu _{Y})\mathrm {P} (X=x,Y=y)}\\[5pt]&={\frac {3{\sqrt {3}}}{2}}\left(\left(1-{\frac {2}{3}}\right)(-1-0){\frac {1}{3}}+\left(0-{\frac {2}{3}}\right)(0-0){\frac {1}{3}}+\left(1-{\frac {2}{3}}\right)(1-0){\frac {1}{3}}\right)=0.\end{aligned}}}$

Rank correlation coefficients, such as Spearman's rank correlation coefficient and Kendall's rank correlation coefficient (τ) measure the extent to which, as one variable increases, the other variable tends to increase, without requiring that increase to be represented by a linear relationship. If, as the one variable increases, the other decreases, the rank correlation coefficients will be negative. It is common to regard these rank correlation coefficients as alternatives to Pearson's coefficient, used either to reduce the amount of calculation or to make the coefficient less sensitive to non-normality in distributions. However, this view has little mathematical basis, as rank correlation coefficients measure a different type of relationship than the Pearson product-moment correlation coefficient, and are best seen as measures of a different type of association, rather than as an alternative measure of the population correlation coefficient.^[7][8]

To illustrate the nature of rank correlation, and its difference from linear correlation, consider the following four pairs of numbers ${\displaystyle (x,y)}$:

(0, 1), (10, 100), (101, 500), (102, 2000).

As we go from each pair to the next pair ${\displaystyle x}$ increases, and so does ${\displaystyle y}$. This relationship is perfect, in the sense that an increase in ${\displaystyle x}$ is always accompanied by an increase in ${\displaystyle y}$. This means that we have a perfect rank correlation, and both Spearman's and Kendall's correlation coefficients are 1, whereas in this example Pearson product-moment correlation coefficient is 0.7544, indicating that the points are far from lying on a straight line. In the same way if ${\displaystyle y}$ always decreases when ${\displaystyle x}$ increases, the rank correlation coefficients will be −1, while the Pearson product-moment correlation coefficient may or may not be close to −1, depending on how close the points are to a straight line. Although in the extreme cases of perfect rank correlation the two coefficients are both equal (being both +1 or both −1), this is not generally the case, and so values of the two coefficients cannot meaningfully be compared.^[7] For example, for the three pairs (1, 1) (2, 3) (3, 2) Spearman's coefficient is 1/2, while Kendall's coefficient is 1/3.

The information given by a correlation coefficient is not enough to define the dependence structure between random variables. The correlation coefficient completely defines the dependence structure only in very particular cases, for example when the distribution is a multivariate normal distribution. (See diagram above.) In the case of elliptical distributions it characterizes the (hyper-)ellipses of equal density; however, it does not completely characterize the dependence structure (for example, a multivariate t-distribution's degrees of freedom determine the level of tail dependence).

For continuous variables, multiple alternative measures of dependence were introduced to address the deficiency of Pearson's correlation that it can be zero for dependent random variables (see ^[9] and reference references therein for an overview). They all share the important property that a value of zero implies independence. This led some authors ^[9][10] to recommend their routine usage, particularly of Distance correlation.^[11][12] Another alternative measure is the Randomized Dependence Coefficient.^[13] The RDC is a computationally efficient, copula-based measure of dependence between multivariate random variables and is invariant with respect to non-linear scalings of random variables.

One important disadvantage of the alternative, more general measures is that, when used to test whether two variables are associated, they tend to have lower power compared to Pearson's correlation when the data follow a multivariate normal distribution.^[9] This is an implication of the No free lunch theorem. To detect all kinds of relationships, these measures have to sacrifice power on other relationships, particularly for the important special case of a linear relationship with Gaussian marginals, for which Pearson's correlation is optimal. Another problem concerns interpretation. While Person's correlation can be interpreted for all values, the alternative measures can generally only be interpreted meaningfully at the extremes.^[14]

For two binary variables, the odds ratio measures their dependence, and takes range non-negative numbers, possibly infinity: ⁠${\displaystyle [0,+\infty ]}$⁠. Related statistics such as Yule's Y and Yule's Q normalize this to the correlation-like range ⁠${\displaystyle [-1,1]}$⁠. The odds ratio is generalized by the logistic model to model cases where the dependent variables are discrete and there may be one or more independent variables.

The correlation ratio, entropy-based mutual information, total correlation, dual total correlation and polychoric correlation are all also capable of detecting more general dependencies, as is consideration of the copula between them, while the coefficient of determination generalizes the correlation coefficient to multiple regression.

The degree of dependence between variables X and Y does not depend on the scale on which the variables are expressed. That is, if we are analyzing the relationship between X and Y, most correlation measures are unaffected by transforming X to a + bX and Y to c + dY, where a, b, c, and d are constants (b and d being positive). This is true of some correlation statistics as well as their population analogues. Some correlation statistics, such as the rank correlation coefficient, are also invariant to monotone transformations of the marginal distributions of X and/or Y.

Most correlation measures are sensitive to the manner in which X and Y are sampled. Dependencies tend to be stronger if viewed over a wider range of values. Thus, if we consider the correlation coefficient between the heights of fathers and their sons over all adult males, and compare it to the same correlation coefficient calculated when the fathers are selected to be between 165 cm and 170 cm in height, the correlation will be weaker in the latter case. Several techniques have been developed that attempt to correct for range restriction in one or both variables, and are commonly used in meta-analysis; the most common are Thorndike's case II and case III equations.^[15]

Various correlation measures in use may be undefined for certain joint distributions of X and Y. For example, the Pearson correlation coefficient is defined in terms of moments, and hence will be undefined if the moments are undefined. Measures of dependence based on quantiles are always defined. Sample-based statistics intended to estimate population measures of dependence may or may not have desirable statistical properties such as being unbiased, or asymptotically consistent, based on the spatial structure of the population from which the data were sampled.

Sensitivity to the data distribution can be used to an advantage. For example, scaled correlation is designed to use the sensitivity to the range in order to pick out correlations between fast components of time series.^[16] By reducing the range of values in a controlled manner, the correlations on long time scale are filtered out and only the correlations on short time scales are revealed.

The correlation matrix of ${\displaystyle n}$ random variables ${\displaystyle X_{1},\ldots ,X_{n}}$ is the ${\displaystyle n\times n}$ matrix ${\displaystyle C}$ whose ${\displaystyle (i,j)}$ entry is

${\displaystyle c_{ij}:=\operatorname {corr} (X_{i},X_{j})={\frac {\operatorname {cov} (X_{i},X_{j})}{\sigma _{X_{i}}\sigma _{X_{j}}}},\quad {\text{if}}\ \sigma _{X_{i}}\sigma _{X_{j}}>0.}$

Thus the diagonal entries are all identically one. If the measures of correlation used are product-moment coefficients, the correlation matrix is the same as the covariance matrix of the standardized random variables ${\displaystyle X_{i}/\sigma (X_{i})}$ for ${\displaystyle i=1,\dots ,n}$. This applies both to the matrix of population correlations (in which case ${\displaystyle \sigma }$ is the population standard deviation), and to the matrix of sample correlations (in which case ${\displaystyle \sigma }$ denotes the sample standard deviation). Consequently, each is necessarily a positive-semidefinite matrix. Moreover, the correlation matrix is strictly positive definite if no variable can have all its values exactly generated as a linear function of the values of the others.

The correlation matrix is symmetric because the correlation between ${\displaystyle X_{i}}$ and ${\displaystyle X_{j}}$ is the same as the correlation between ${\displaystyle X_{j}}$ and ${\displaystyle X_{i}}$.

A correlation matrix appears, for example, in one formula for the coefficient of multiple determination, a measure of goodness of fit in multiple regression.

In statistical modelling, correlation matrices representing the relationships between variables are categorized into different correlation structures, which are distinguished by factors such as the number of parameters required to estimate them. For example, in an exchangeable correlation matrix, all pairs of variables are modeled as having the same correlation, so all non-diagonal elements of the matrix are equal to each other. On the other hand, an autoregressive matrix is often used when variables represent a time series, since correlations are likely to be greater when measurements are closer in time. Other examples include independent, unstructured, M-dependent, and Toeplitz.

In exploratory data analysis, the iconography of correlations consists in replacing a correlation matrix by a diagram where the "remarkable" correlations are represented by a solid line (positive correlation), or a dotted line (negative correlation).

Similarly for two stochastic processes ${\displaystyle \left\{X_{t}\right\}_{t\in {\mathcal {T}}}}$ and ${\displaystyle \left\{Y_{t}\right\}_{t\in {\mathcal {T}}}}$: If they are independent, then they are uncorrelated.^{[21]: p. 151} The opposite of this statement might not be true. Even if two variables are uncorrelated, they might not be independent to each other.

If a pair ${\displaystyle \ (X,Y)\ }$ of random variables follows a bivariate normal distribution, the conditional mean ${\displaystyle \operatorname {\boldsymbol {\mathcal {E}}} (X\mid Y)}$ is a linear function of ${\displaystyle Y}$, and the conditional mean ${\displaystyle \operatorname {\boldsymbol {\mathcal {E}}} (Y\mid X)}$ is a linear function of ${\displaystyle \ X~.}$ The correlation coefficient ${\displaystyle \ \rho _{X,Y}\ }$ between ${\displaystyle \ X\ }$ and ${\displaystyle \ Y\ ,}$ and the marginal means and variances of ${\displaystyle \ X\ }$ and ${\displaystyle \ Y\ }$ determine this linear relationship:

${\displaystyle \operatorname {\boldsymbol {\mathcal {E}}} (Y\mid X)=\operatorname {\boldsymbol {\mathcal {E}}} (Y)+\rho _{X,Y}\cdot \sigma _{Y}\cdot {\frac {\ X-\operatorname {\boldsymbol {\mathcal {E}}} (X)\ }{\sigma _{X}}}\ ,}$

where ${\displaystyle \operatorname {\boldsymbol {\mathcal {E}}} (X)}$ and ${\displaystyle \operatorname {\boldsymbol {\mathcal {E}}} (Y)}$ are the expected values of ${\displaystyle \ X\ }$ and ${\displaystyle \ Y\ ,}$ respectively, and ${\displaystyle \ \sigma _{X}\ }$ and ${\displaystyle \ \sigma _{Y}\ }$ are the standard deviations of ${\displaystyle \ X\ }$ and ${\displaystyle \ Y\ ,}$ respectively.

The empirical correlation ${\displaystyle r}$ is an estimate of the correlation coefficient ${\displaystyle \ \rho ~.}$ A distribution estimate for ${\displaystyle \ \rho \ }$ is given by

$${\displaystyle \pi (\rho \mid r)={\frac {\ \Gamma (N)\ }{\ {\sqrt {2\pi \ }}\cdot \Gamma (N-{\tfrac {\ 1\ }{2}})\ }}\cdot {\bigl (}1-r^{2}{\bigr )}^{\frac {\ N\ -2\ }{2}}\cdot {\bigl (}1-\rho ^{2}{\bigr )}^{\frac {\ N-3\ }{2}}\cdot {\bigl (}1-r\rho {\bigr )}^{-N+{\frac {\ 3\ }{2}}}\cdot F_{\mathsf {Hyp}}\left(\ {\tfrac {\ 3\ }{2}},-{\tfrac {\ 1\ }{2}};N-{\tfrac {\ 1\ }{2}};{\frac {\ 1+r\rho \ }{2}}\ \right)\ }$$

where ${\displaystyle \ F_{\mathsf {Hyp}}\ }$ is the Gaussian hypergeometric function.

This density is both a Bayesian posterior density and an exact optimal confidence distribution density.^[24][25]