Robust Tests in Regression Models With Omnibus Alternatives and Bounded Influence

A robust approach for testing the parametric form of a regression function versus an omnibus alternative is introduced. This generalizes existing robust methods for testing subhypotheses in a regression model. The new test is motivated by developments in modern smoothingbased testing procedures and can be viewed as a robustification of a smoothing-based conditional moment test. It is asymptotically normal under both the null hypothesis and local alternatives. The robustified test retains the “omnibus” property of the corresponding smoothing test; that is, it is consistent for any fixed smooth alternative in an infinite-dimensional space. It is shown that the bias of the asymptotic level under shrinking local contamination is bounded only if the second-order Hampel’s influence function is bounded. The test’s performance is demonstrated through both Monte Carlo simulations and application to an agricultural dataset.