Testing Endogeneity with Possibly Invalid Instruments and High Dimensional Covariates

The Durbin-Wu-Hausman (DWH) test is a commonly used test for endogeneity in instrumental variables (IV) regression. Unfortunately, the DWH test depends, among other things, on assuming all the instruments are valid, a rarity in practice. In this paper, we show that the DWH test often has distorted size even if one IV is invalid. Also, the DWH test may have low power when many, possibly high dimensional, covariates are used to make the instruments more plausibly valid. To remedy these shortcomings, we propose a new endogeneity test which has proper size and better power when invalid instruments and high dimemsional covariates are present; in low dimensions, the new test is optimal in that its power is equivalent to the "oracle" DWH test's power that knows which instruments are valid. The paper concludes with a simulation study of the new test with invalid instruments and high dimensional covariates.