Generic consistency of the break-point estimator under specification errors

This paper considers the asymptotic behavior of the break-point estimator when some or all of the variables in a structural-break model are misspecified. An obvious example is misspecifying a linear model as a log--log model. The results given here cover a large number of data transformations, including transformation of the independent variables, dependent variable and transformation of both variables at the same time. The true data generating process can be stationary or non-stationary. I establish a useful result that, under some relatively weak assumptions, the break point can always be consistently estimated regardless of how a structural-break model is misspecified. The asymptotic behavior of the SupWald test under model misspecification is studied. Simulations and empirical evidence are also provided. Copyright Royal Economic Society, 2003