Local and global rank tests for multivariate varying-coefficient models

In a multivariate varying-coefficient model, the response vectors Y are regressed on known functions u(X) of some explanatory variables X and the coefficients in an unknown regression matrix q(Z) depend on another set of explanatory variables Z. We provide statistical tests, called local and global rank tests, which allow to estimate the rank of an unknown regression coefficient matrix q(Z) locally at a fixed level of the variable Z or globally as the maximum rank over all levels of Z, respectively. In the case of local rank tests, we do so by applying already available rank tests to a kernel-based estimator of the coefficient matrix q(z). Global rank tests are obtained by integrating test statistics used in estimation of local rank tests. We present a simulation study where, focusing on global ranks, we examine small sample properties of the considered statistical tests. We also apply our results to estimate the so-called local and global ranks in a demand system where budget shares are regressed on known functions of total expenditures and the coefficients in a regression matrix depend on prices faced by a consumer.