Limit theorems for change in linear regression (Q1323141)

Asymptotic tests are proposed to detect the possible changepoint in a multiple linear regression model based on independent, nonsystematic errors with constant variance and \(\nu>2\) moments. The design points may be nonrandom or random. Normality of the error terms is not required. The tests are developed in the ``at most one changepoint'' setting with \(H_ A\) stating that the \(d\)-dimensional parameter vector \(\beta\) changes to \(\gamma\), \(\beta\neq \gamma\), at some unknown time, say at the \(k^*\)th observation. References for reviews of this problem and historical accounts as well as a clear motivation of the test statistic are given at the beginning of the paper. The test statistic which is shown to be consistent is \[ \max_{1\leq k\leq n} (k/(1-k/n))^{1/2} \Biggl| {{\overline{y}_ k- \overline{y}_ n- \widehat{\beta}_ n (\overline{\mathbf x}_ k- \overline{\mathbf x}_ n)^ T} \over {(1-k( \overline{\mathbf x}_ k- \overline{\mathbf x}_ n) (\overline{\mathbf x}_ k- \overline{\mathbf x}_ n )^ T/ (Q_ n(1- k/n)))^{1/2}}} \Biggr|; \] \[ \text{where} \qquad \qquad \overline{y}_ k= k^{-1} \sum_{1\leq i\leq k} y_ i, \quad \overline{\mathbf x}_ k= k^{-1} \sum_{1\leq i\leq k} {\mathbf x}_ i, \quad Q_ n= \sum_{1\leq i\leq n} ({\mathbf x}_ i- \overline{\mathbf x}_ n) ({\mathbf x}_ i- \overline{\mathbf x}_ n)^ T. \] Several limit theorems regarding this statistic (and a related one) are proven in the last section of the paper. Following the presentation of each result, a concise interpretation is given. After addressing the problem of variance estimation under \(H_ A\), three examples and the results of a small scale Monte Carlo simulation study are discussed.