Bootstrap log-likelihood ratio test for linear hypothesis testing in problem with instrumental variables

In this talk bootstrap log-likelihhod ratio (BLR) test will be justified to recover real world log-likelihood ratio (LR) test statistic. It is demonstrated that BLR test is valid under $(\frac{p^3}{n})^{\frac{1}{6}}$ small condition. This procedure will be further used to test linear hypothesis on the target in parametric regression model with instrumental variables, where we admit that do not have any prior knowledge on instruments identification. It was shown that testing hypothesis using such a data driven approach provides numerical performance comparable to the other tests presented in literature.

References

1. Spokoiny V., Zhilova M., “Bootstrap confidence sets under a model misspecification”, 2014, arXiv: 1410.0347
2. D. Andrews D., M. J. Moreira, J. H.Stock, “Optimal Two-Sided Invariant Similar Tests for Instrumental Variables Regression”, Econometrica, 74:3 (2006), 715–752      
3. M. J. Moreira, “A conditional likelihood ratio test for structural models”, Econometrica, 71:4 (2003), 1027–1048