Misleading Regressions with Constructed Variables.

It is common practice to examine empirical models in which one of the regressors is constructed as the weighted average or sum of a set of series that includes the dependent variable. Examples include models relating money and wealth, consumption and income and regional and national unemployment. In this paper we show that biased results are likely to be generated by such models and that the identified bias is distinct from the more familiar simultaneous equation bias. The theoretical arguments are illustrated with simulation experiments and as a practical example we consider the relationship between regional and national unemployment in Australia.