SIMULATION EXPLORATION OF THE POWER OF MARGINAL REGRESSION ANALYSIS*

In either simple or multivariate regression models, measured coefficients represent a kind of average of co-variant relationships implicit in the data. For time-series data, if co-variant relationships follow no particular pattern over the sample period, use of this average in forecasting is rational. However, if relationships among independent and dependent variables is undergoing change over time, then measured regression coefficients are likely to be poor tools for prediction. This paper is an exploratory study of this general problem area. It develops implications as to the predictive impact of the problem and offers preliminary suggestions as to a method of developing predictions under such circumstances.