Despite years of stat courses I'm afraid I may still not completely understand bootstrapping.

My question here relates to nonparametric boostrapping of regression models. As i understand it you draw N bootstrap resamples of your original data to approximate the true n^n possible bootstrap samples, which in turn approximate the true underlying distribution.

However, if you were to draw all n^n, that would include rank deficient samples of the form (X1, X1, X1,... X1) for which there isn't a unique regression estimate. Even if you only do N resamples, there is still some non-zero probability mass on the rank-deficient resamples.

So how can the nonparametric bootstrap actually give you any meaningful result?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.