For Pearson correlation coefficient one has a "population" formula in terms of the expectations / variances / covariances, and a "sample" estimate formula based on actual observations. However, for Spearman's correlation, I have only seen it given in terms of the "sample" formula, never the population one.

It seems that in order to have a population formula for Spearman's correlation, one has to at least fix the sample size. Assuming absolutely continuous densities $f_X$, $f_Y$, and $f_{XY}$ (both joint and marginal) for random variables $X$ and $Y$, and denoting by $\mathbf X_n = \{X_1,X_2,...X_n\}$ a sample of size $n$, the probability of any two $X$ in the sample being the same is 0 and hence the ranks are always given by the set of integers from $1$ to $n$, which has the same "variance" as the uniform discreet distribution given by $$ \mathrm {Var}(\mathrm{rank}(X)) = \frac{n^2-1}{12}$$ and similarly for $Y$. So, this is independent of the marginal distributions. We now need to find the expression for the covariance of ranked samples. If we denote by $\mathbf X_{(n)} = \{X_{(1)},X_{(2)},...X_{(n)}\}$ the corresponding order statistics in the increasing order (and similarly for $Y$), then the density function for the order statistics are given by

$$f_{\mathbf X_{(n)}} = n!f_X(x_1)f_X(x_2) \cdots f_X(x_n) \mathbb I_{x_1 \le x_2} \mathbb I_{x_2 \le x_3} \cdots \mathbb I_{x_{n-1} \le x_n}$$


$$f_{\mathbf Y_{(n)}} = n!f_Y(y_1)f_Y(y_2) \cdots f_Y(y_n) \mathbb I_{y_1 \le y_2} \mathbb I_{y_2 \le y_3} \cdots \mathbb I_{y_{n-1} \le y_n}$$

However, it is not clear to me how to express the joint distribution of ranked variables in order to calculate their covariance.


1 Answer 1


It depends on the ranking criteria of $(X_1,Y_1), (X_2, Y_2), \ldots, (X_n, Y_n)$.


  1. $(X_i, Y_i) \leq (X_j, Y_j) \iff X_i \leq X_j$ then,

$$f_{(X,Y)_{(n)}} = n! f_{X,Y}(x_1,y_1)f_{X,Y}(x_2,y_2) \ldots f_{X,Y}(x_n, y_n)\mathbb{I}_{x_1 \leq x_2 \leq x_3 \ldots \leq x_n}$$

  1. $(X_i, Y_i) \leq (X_j, Y_j) \iff Y_i \leq Y_j$ then,

$$f_{(X,Y)_{(n)}} = n! f_{X,Y}(x_1,y_1)f_{X,Y}(x_2,y_2) \ldots f_{X,Y}(x_n, y_n)\mathbb{I}_{y_1 \leq y_2 \leq y_3 \ldots \leq y_n}$$

  1. $(X_i, Y_i) \equiv (\text{ith smallest }_{j=1}^{n} \ X_j, \ \text{ith smallest }_{j=1}^{n} \ Y_j)$ then,

$$f_{(X,Y)_{(n)}} = n!n! f_{X,Y}(x_1,y_1)f_{X,Y}(x_2,y_2) \ldots f_{X,Y}(x_n, y_n){I}_{x_1 \leq x_2 \leq x_3 \ldots \leq x_n}\mathbb{I}_{y_1 \leq y_2 \leq y_3 \ldots \leq y_n}$$

In each case, you multiply $f_{X,Y}(x_1,y_1)f_{X,Y}(x_2,y_2) \ldots f_{X,Y}(x_n, y_n)$ with number of possible ways of assigning a realization of the sample to the random variables in the sample such that the ranking criteria is satisfied i.e. if the realization of the sample is $x_1, x_2, \ldots, x_n, y_1, y_2, \ldots y_n$, then multiply with the number of ways you can assign these $x_i$'s to $X_j$'s and $y_k$'s to $Y_l$'s such that the ranking criteria is satisfied. You also multiple with a proper indicator function which is one if the given realization of the sample satisfies the ranking criteria otherwise zero.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .