Study of partial and average conditional Kendall's tau (Q2236383)

Consider the dependence structure between two random variables \(Y_1\) and \(Y_2\) conditioned by another variable \(X\). According to the simplifying assumption, only the marginal distributions of \(Y_1\) and \(Y_2\) are influenced by \(X\). The authors compare in detail the average and the partial Kendall's tau. If \(X=x\), let \(H_x(y_1,y_2)=P(Y_1\le y_1,Y_2\le y_2\mid X=x)\) be the conditional joint distribution function and \(F_{1x}(y_1)=H_x(y_1,+\infty)\) and \(F_{2x}(y_2)=H_x(+\infty,y_2)\) its marginals. Then a copula \(C_x\) exists such that \(H_x(y_1,y_2)=C_x(F_{1x}(y_1),F_{2x}(y_2))\); \(C_x\) is the conditional copula. Kendall's conditional copula and its average version are respectively defined by \[ \tau(x)=4\,\int C_x(u_1,u_2)\,dC_x(u_1,u_2)\quad\text{and}\quad E_x(\tau(x)). \] These are studied in the case of the mixture of two copulas \[ C_x(u_1,u_2)=w(x)\,C_1(u_1,u_2)+(1-w(x))\,C_2(u_1,u_2) \] and when \((Y_1,Y_2,X)\) has an Archimedean copula. In both cases theoretical results and meaningful examples are presented.\par In the Archimedean case the absence of an explicit analytical expression, the doubt remains that \(\tau^A=\bar{\tau}\) might still be possible.