An alternative derivation of aligned rank tests for regression (Q1174647): Difference between revisions

A new method for deriving aligned rank tests and their efficacy is presented. This theory has been developed by \textit{C. Chiang} and \textit{M. L. Puri} [Ann. Inst. Stat. Math. 36, 35--50 (1984; Zbl 0572.62038)] and alternative proofs are given here. Let us consider the linear model \[ Y_ i=\beta_ 0+\beta'x_ i+\varepsilon_ i=\beta_ 0+\beta'_ 1 x_{1i}+\beta'_ 2 x_{2i}+ \varepsilon_ i \] where \(\beta_ 0\) is a scalar, \(\beta\) and \(x_ i\) are \(p\)-dimensional vectors, \(\beta_ 1\) and \(x_{1i}\) are \(q\)-dimensional vectors, \(\beta_ 0\) and \(\beta=(\beta_ 1,\beta_ 2)\) are unknown parameters and \(x_ i=(x_{1i},x_{2i})\), \(i=1,\dots,n\), are known. Two tests are usually associated to these models: a full hypothesis test with \(H_ 0: \beta=0\) and a subhypothesis \(H_ 0: \beta_ 2=0\). We suppose that \(\varepsilon_ i\), \(i=1,\dots,n\), are i.i.d. with distribution function \(F\) with zero mean, finite variance and density \(f\). Let \(G_ n\) denote the empirical distribution function of the \(Y\)'s; then \(n G_ n(Y_ i)\) is the rank of \(Y_ i\) and \(G_ n\) is the rank transformation. Let \(G\) be the limit of \(G_ n\) under the null hypothesis. Let us introduce the following notations \[ Y^*=(G(Y_ 1),\dots,G(Y_ n))'\text{ and } Y^{**}=(G_ n(Y_ 1),\dots,G_ n(Y_ n))'. \] With some hypothesis on the matrix \(X=[x_{ij}]\), and under the null hypothesis \(\beta=0\), it is then proved that \begin{align*} &G_n(x)\to G(x)=F(x-\beta_0);\\ &G(Y_i)=0,5 +e_i \text{ where } e_i= F(\varepsilon_i)- 0,5\text{ and } \text{Var}(e_i)=1/12;\\ &(Y^*)'X(X'X)^{-1} X'Y^* \to (1/12)\chi_p^2;\\ &(Y^*)'X(X'X)^{-1}X'Y^*- (Y^{**})' X(X'X)^{-1} X'Y^{**} \to 0\text{ in probability }. \end{align*} Consider now the hypothesis \(H_{2n}: \beta_{2n}=h/\sqrt n\). Let \(\hat \beta_ 1\) denote a consistent estimate of \(\beta_ 1\) such that \(\sqrt n(\hat \beta_ 1-\beta_ 1)=O_ p(1)\). Let us introduce the following notations: \[ Z_ i=Y_ i-\hat \beta_ 1x_{1i}=\beta_ 0+A_ i=\beta_ 2'x_{2i}+\varepsilon_ i \] and the aligned ranks for the \(Z_ i\)'s \[ G_ n(x;\Delta)=n^{-1}\sum_{i=1}^ n I(Y_ i- \beta'_ 1 x_{1i}-\Delta'x_{1i}\leq x). \] Then, we can see that \(G_ n(x;\hat \beta_ 1-\beta_ 1)\) is the empirical distribution function of the \(Z_ i\)'s. Moreover, with \(G(x;\Delta)=n^{- 1}\sum_{i=1}^ n G(x+\Delta'x_{1i})\) which implies \(G(x;0)=G(x)\), and, under \(H_ 0: \beta_ 2=0\), we obtain \[ G_ n(x;\Delta)=G_ n(x)\to G(x)=F(x-\beta_ 0). \] Then, with the following notations, \begin{align*} V & =(0,5+e_1,\dots,0,5+e_n),\quad \gamma^2=\lim_{n\to \infty} n^{-1} \sum (h'x_{2i})^2,\\ Z^* & =(G(Z_1), \dots, G(Z_n))' \text{ and }Z^{**} =(G_n(Z_1;\hat \beta_1-\beta_1), \dots, G_n(Z_n;\hat \beta_1- \beta_1))' \end{align*} it is proved that, under \(H_ 0: \beta_ 2=0\), \[ V'X_ 2(X_ 2'X_ 2)^{-1}X_ 2'V\to (1/12)\chi_{p-q}^ 2,\quad (Z^{**})'X_ 2(X_ 2'X_ 2)^{-1}X_ 2'Z^{**}\to (1/12)\chi_{p-q}^ 2 \] and that, under \(H_{2n}\), \[ (Z^{**})'X_ 2(X'_ 2 X_ 2)^{-1} X'_ 2 Z^{**} \to (1/12) \chi_{p-q}^ 2(12 \gamma^ 2\mathbb{E}(f(\varepsilon))^ 2). \] These results allow the study of the Pitman efficiency for this aligned rank test.