Robust heteroskedasticity-robust tests

- Robustness of Hausman and Palmer's heteroscedasticity-robust test is investigated.
- It is found to have large size distortion in several data generating processes.
- A wild bootstrap test has much better size in finite samples.
- Hausman and Palmer provides no robust power improvement over the wild bootstrap.

Hausman and Palmer (2012) suggest using the Edgeworth corrected critical values of Rothenberg (1988) along with a pairs bootstrap covariance matrix estimator in order to obtain second order correct heteroskedasticity-robust inferences. According to their simulations, this test has size comparable to and power greater than a wild bootstrap test. In this note, I show that this does not hold in general. Using a more extensive set of simulations reveals that the wild bootstrap test is much more robust to the underlying data generating process.