Cramér type large deviations for generalized rank statistics (Q762850): Difference between revisions

Consider a triangular array \(X_{N1},X_{N2},...,X_{NN}\) \((N=1,2,...)\) of independent (not necessarily identically distributed) random variables. The authors are concerned with the relative error of the normal approximation of generalized linear rank statistics \(T_ N\) based on \(X_{N1},...,X_{NN}\). Their main result is as follows: Under suitable regularity conditions on the regression constants and score generating function, uniformly in the region \(0<x\leq \rho_ NN^{1/6}\), \(\rho_ N=o(1)\), one has as \(N\to\infty:\) \[ P(T_ N-ET_ N>\tau_ Nx)/[1-\Phi (x)]=1+o(1),\text{ where }\tau^ 2_ N=Var(T_ N) \] and \(\Phi\) denotes the standard normal d.f. For the special case of signed linear rank statistics a similar result is stated under somewhat weaker assumptions on regression constants and score function, but assuming row-wise identical continuous d.f.s \(F_ N\).