UNSTABLE MODELS FROM INCORRECT FORMS

Parametric tests for structural change are conditional on the joint hypothesis of functional form and other aspects of the model specification. This problem is often disregarded. Monte Carlo evidence using three data sets indicates that apparently innocuous specification errors can lead to substantial increases in the probability of finding structural change when it is not present in the data-generating mechanism. Significant Chow tests and autocorrelation are much more likely when the wrong functional form is used. Maximum Chow tests falsely reject stable preferences much more often than their nominal size suggests, even when the correct model is estimated.