Comments on: Bootstrapping and Subsampling: Part II http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/ Thoughts on Statistics and Machine Learning Sat, 25 May 2013 23:18:25 +0000 hourly 1 http://wordpress.com/ By: Comunidades tribais são mais violentas? O quão próxima é a distribuição normal? O papel do BNDES. | Análise Real http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-7060 Tue, 05 Feb 2013 01:14:43 +0000 http://normaldeviate.wordpress.com/?p=358#comment-7060 [...] – Sobre bootstraping I e II; [...]

]]> By: normaldeviate http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6975 Fri, 01 Feb 2013 13:40:18 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6975 I agree. I think with some work, something could be said about the finite sample properties although
it would not be easy.

]]> By: Christian Hennig http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6968 Fri, 01 Feb 2013 12:46:08 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6968 I see a problem with this kind of asymptotic result that is not a problem with all asymptotic results in statistics (but with some others, too). This is that in reality we have a fixed n and a fixed b, and the asymptotic relies on *how b changes with n*. Now the interpretation of increasing n is clear: We observe more and get more information. No statistician would *choose* to have a small n if large n is available (OK, let’s say I don’t bother for the moment with situations where she would). However, what happens to b if n goes to infinity is actually *decided* by the statistician. And it is only *hypothetically* decided by the statistician if in reality there is only one fixed n and one fixed b. So if my n is 500 and I choose b to be 250, whether the asymptotic result holds or not depends on what b I’d choose for n=1,000, 10,000, 1 billion and so on, which I don’t need to bother about because that’s not the n that I have.
So how useful is an asymptotic result in a given real situation if its validity depends on how the reader believes the statistician would choose b with much larger n, which is a choice that in fact the statistician never has to make?

]]> By: Mark Schaffer http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6927 Thu, 31 Jan 2013 13:49:11 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6927 Great, thanks!

]]> By: normaldeviate http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6925 Thu, 31 Jan 2013 13:32:08 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6925 Yes. That was one of the motivations for subsampling, it works on dependent data.
You need to divide the data into blocks and subsample blocks.
Most of the Romano book is actually about the dependent case.
If you google “time series subsampling” you’ll
find lots of good stuff.

]]> By: Mark Schaffer http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6924 Thu, 31 Jan 2013 13:28:01 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6924 First of all, thanks for writing this – really interesting. And a question – what if your data aren’t i.i.d.? There are bootstrap flavours for that, and I wonder if there are subsampling flavours as well.

]]> By: Andrew Beam http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6911 Thu, 31 Jan 2013 02:03:08 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6911 I would love to see this lead into a post about bagging. Bagging is obviously a technique widely used in machine learning (which as you mentioned the bootstrap is not) and I would love to hear this type of discussion on it.

]]> By: László Sándor http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6837 Tue, 29 Jan 2013 15:18:01 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6837 This is really helpful, thanks. Back to an older question of mine, would you say that subsampling is more promising for “bunching” estimates that people try to do from big data, distributions these days? So far, bootstrapping was what all there was, see e.g. http://obs.rc.fas.harvard.edu/chetty/denmark_adjcost.pdf

I also wonder whether this is could be used from regression discontinuity estimates, the literature seems to be heading into a different direction, e.g. http://www-personal.umich.edu/~cattaneo/papers/RD-biascorrection.pdf

And by the way, if we are “sticking to the data” this much, are we getting standard errors (closer to those) conditional on covariates, or not?
I basically mean this paper: http://www.nber.org/papers/w17442.pdf

]]> By: normaldeviate http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6806 Mon, 28 Jan 2013 20:23:12 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6806 No. You can actually estimate that too.

]]> By: anon http://normaldeviate.wordpress.com/2013/01/27/bootstrapping-and-subsampling-part-ii/comment-page-1/#comment-6804 Mon, 28 Jan 2013 20:22:05 +0000 http://normaldeviate.wordpress.com/?p=358#comment-6804 Don’t you need to know \beta when constructing the subsampling distribution L?

]]>