The medcouple is a robust measure of skewness yielding values between \(-1\) and \(1\). For left- and right-skewed data the medcouple is negative and positive respectively.
The medcouple is defined as the median of the kernel function
\(h(x_i,x_j) = \frac{(x_j - med(x)) - (med(x)-x_i)}{x_j-x_i}\)
evaluated over all couples \((x_i,x_j)\) where
\(x_i\) is smaller than the median of x and \(x_j\) larger than the median of x. When there are multiple observations tied to the median, the kernel is defined separately as the denominator is not defined for these observations. Let \(m_1 < ... < m_k\) denote the indices of the observations which are tied to the median. Then \(h(x_{m_i},x_{m_j})\) is defined to equal \(-1\) if \(i + j - 1 < k\), \(0\) when \(i + j - 1 = k\) and \(+1\) if \(i + j - 1 > k\). To compute the medcouple an algorithm with time complexity \(O(n log(n))\) is applied. For details, see https://en.wikipedia.org/wiki/Medcouple.
For numerical accuracy it is advised, for small data sets, to compute the medcouple on both x and -x. The final value of the medcouple may then be obtained as a linear combination of both calculations. This procedure is warranted by the properties of the medcouple. Indeed the medcouple of the distribution \(X\) equals minus the medcouple of the reflected distribution \(-X\). Moreover the medcouple is location and scale invariant.
Note that missing values are not allowed.