Testing for trends in correlated data

The problem of testing for the significance of a linear trend in the presence of positively correlated errors is considered. Test criteria based on ordinary least squares, conditional maximum likelihood, estimated generalized least squares and maximum likelihood estimates tend to have higher significance levels than nominal levels for positively correlated series of moderate length. In this paper, we study three alternative methods: (a) pre-test, (b) bias-adjusted, and (c) bootstrap-based procedures. A simulation study is used to compare the empirical level and power of different procedures. An example is used to illustrate the procedures.