I've been dealing with the problem of how to construct confidence intervals on penalized spline estimators in the presence of cluster-wise auto-correlation and heteroskedasticity. My previous thread on this is here: Penalized spline confidence intervals based on cluster-sandwich VCV
One of the commenters suggested using a wild bootstrap. This is a bit tricky in the presence of clustered data, but this paper by Cameron et al. at UC Davis proposes what seems to be a relatively simple approach that accommodates the clustered context: http://ideas.repec.org/a/tpr/restat/v90y2008i3p414-427.html
Basically, the residual vector for each cluster has an equal chance of being left alone, or being multiplied by -1. Thats it. No resampling. Over many iterations, you get many different combinations of cluster-wise residual vectors, but your combinations are restricted to the set of residual vectors that you already have, potentially multiplied by -1. (I'm aware that there are alternative things that you might multiply your residual vector by, summarized on wikipedia: http://en.wikipedia.org/wiki/Bootstrapping_%28statistics%29#Wild_bootstrap)
Basically this all seems too simple. Am I missing some complicating wrinkle?
Given that I'm programming my own estimators in R, I need to program my own bootstraps as well. I want to make sure I get the procedure right before I go and code.
Thanks in advance.