Clustered data with Design package--bootcov() vs. robcov()
jjh21 wrote:
Another question related to bootcov(): A reviewer is concerned with the fact that bootstrapping the standard errors does not give the same answers each time. What is a good way to address this concern? Could I bootstrap, say, 100 times and report the mean standard error of those 100 estimates? I am already doing 1,000 replications in the bootstrap, but of course the answer is still slightly different each time.
First, you can argue that everything we estimate has a margin of error and that the variation across different runs of the bootstrap is within the statistical precision of what can be estimated. Second, run the bootstrap with 10,000 replications and be done with it. Frank
Frank E Harrell Jr wrote:
robcov does not use bootstrapping. It uses the cluster sandwich (Huber-White) variance-covariance estimator for which there are references in the help file (see especially Lin). Both robcov and bootcov work best when there is a large number of small clusters. If the clusters are somewhat large and greatly vary in size, expect to be in trouble and consider a full modeling approach (generalized least squares, mixed models, etc.). One advantage of robcov is that you get the same result every time, unlike bootstrapping. But even in the case of cluster sizes of one, the sandwich estimator can be inefficient (see the Gould paper) or can result in the "right" estimates of the "wrong" quantity (see a paper by Friedman in American Statistician). Frank
Thank you.
--
Frank E Harrell Jr Professor and Chair School of Medicine
Department of Biostatistics Vanderbilt University
______________________________________________ R-help at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
Frank E Harrell Jr Professor and Chair School of Medicine
Department of Biostatistics Vanderbilt University