Thanks Enrique. I found this very interesting indeed. With hindsight, my 2012 paper did not focus sufficiently on the number of regressors or on how they were generated (discrete vs. continuous), so it is great to see some results on these points. Let me note that there is more than one variant of the wild bootstrap, and their finite-sample properties can differ noticeably. My 2012 paper discusses several variants. A very recent paper with Morten Nielsen and Matt Webb, QED Working Paper 1485, has extended the jackknife-based variants to the wild cluster bootstrap. The most promising of these wild cluster bootstrap methods have already been incorporated into boottest by its author, David Roodman.
↧