Bootstrap Tests of Nonnested Linear Regression Models

The J test for nonnested regression models often works badly as an asypmtotic test, but it generally works very well when bootstrapped. We provide a theroretical analysis of the J test which explains both of these phenomena. We also propose a modified version of the test which works even better than the ordinary J test when bootstrapped. Using our theoretical results to make simulation much faster, we obtain extremely accurate Monte Carlo results which demonstrate just how well the bootstraped tests perform.