Re: [R] Skewed t distribution
On 3/28/06, Prof Brian Ripley [EMAIL PROTECTED] wrote: You need to learn to supply adequate information. The current version of sn *does* have such an argument, and I was careful to check. So it seems that you are using an unstated obselete version of sn. Do ugrade as the posting guide asked you to. Ok, point taken, I just forgot about it and had no way to do it before computations started (I have to work on two machines, and the one with R on it has no access to the internet). My apologies for wasting Your everyone else's time. rg, konrad -- We are what we pretend to be, so we must be careful about what we pretend to be Kurt Vonnegut Jr. Mother Night [[alternative HTML version deleted]] __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
[R] Skewed t distribution
Dear All, I am working with skewed-t copula in my research recently, so I needed to write an mle procedure instead of using a standard fit one; I stick to the sn package. On subsamples of the entire population that I deal with, everything is fine. However, on the total sample (difference in cross-sectional dimension: 30 vs 240) things go wrong - the objective function diverges to infinity. I located the rotten line to be t1 - dmst(vector, mu, P, alpha, nu) where vector is the matrix row, on which I evaluate my likelihood and the rest in parametrized in a standard way, just as the help pages give it. In large dimensions, I get a zero value of the density (which is probably due to numerical issues). I tried the following dummy example t1 - rmst(1,mu,P,alpha, nu) t2 - dmst(t1, mu, alpha,nu) and t2 remains to be zero. Can anyone help me on this one? thanks in advance, Konrad -- We are what we pretend to be, so we must be careful about what we pretend to be Kurt Vonnegut Jr. Mother Night [[alternative HTML version deleted]] __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
Try maximizing the log-likelihood and using the log=TRUE argument to dmst. (You have told us so little about what you are doing that we can but guess at what you mean by `write an mle procedure': what is wrong with st.mle, for example?) On Tue, 28 Mar 2006, Konrad Banachewicz wrote: Dear All, I am working with skewed-t copula in my research recently, so I needed to write an mle procedure instead of using a standard fit one; I stick to the sn package. On subsamples of the entire population that I deal with, everything is fine. However, on the total sample (difference in cross-sectional dimension: 30 vs 240) things go wrong - the objective function diverges to infinity. I located the rotten line to be t1 - dmst(vector, mu, P, alpha, nu) where vector is the matrix row, on which I evaluate my likelihood and the rest in parametrized in a standard way, just as the help pages give it. In large dimensions, I get a zero value of the density (which is probably due to numerical issues). I tried the following dummy example t1 - rmst(1,mu,P,alpha, nu) t2 - dmst(t1, mu, alpha,nu) and t2 remains to be zero. Can anyone help me on this one? thanks in advance, Konrad -- We are what we pretend to be, so we must be careful about what we pretend to be Kurt Vonnegut Jr. Mother Night [[alternative HTML version deleted]] __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html -- Brian D. Ripley, [EMAIL PROTECTED] Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UKFax: +44 1865 272595 __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
On Tue, 28 Mar 2006 11:41:19 +0200, Konrad Banachewicz wrote: please supply the ingredients needed to reproduce the problem that you have faced (including the values of the parameters mu,P,alpha,nu, among the rest) best wishes, Adelchi Azzalini KB Dear All, KB I am working with skewed-t copula in my research recently, so I KB needed to write an mle KB procedure instead of using a standard fit one; I stick to the sn KB package. On subsamples of the entire population that I deal with, KB everything is fine. However, on the total sample (difference in KB cross-sectional dimension: 30 vs 240) things go wrong - the KB objective function diverges to infinity. I located the rotten KB line to be KB KB t1 - dmst(vector, mu, P, alpha, nu) KB KB where vector is the matrix row, on which I evaluate my KB likelihood and the rest in parametrized in a standard KB way, just as the help pages give it. In large dimensions, I get a KB zero value of the density (which is probably due to numerical KB issues). I tried the following dummy example KB KB t1 - rmst(1,mu,P,alpha, nu) KB t2 - dmst(t1, mu, alpha,nu) KB KB and t2 remains to be zero. Can anyone help me on this one? KB KB thanks in advance, KB Konrad KB KB -- KB We are what we pretend to be, so we must be careful about what we KB pretend to be KB KB Kurt Vonnegut Jr. Mother Night KB KB [[alternative HTML version deleted]] KB KB __ KB R-help@stat.math.ethz.ch mailing list KB https://stat.ethz.ch/mailman/listinfo/r-help KB PLEASE do read the posting guide! KB http://www.R-project.org/posting-guide.html KB __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
On 3/28/06, Prof Brian Ripley [EMAIL PROTECTED] wrote: Try maximizing the log-likelihood and using the log=TRUE argument to dmst. seems like dmst does not support this argument (the way e.g. dt does) (You have told us so little about what you are doing that we can but guess at what you mean by `write an mle procedure': what is wrong with st.mle, for example?) st.mle assumes skewed-t marginals (for a whole distribution), whereas I am working with a copula so my margins are uniform. The whole point is separating the joint and marginal dynamics. rg, konrad [[alternative HTML version deleted]] __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
P is an identity matrix 240X240, mu and alpha are vectors of zeros 240X1, nu equals 10, so alltogether You need: P - matrix(0,244,244) diag(P) - 1 nu - 10 alpha - rep(0,244) mu - rep(0,244) require(sn) t1 - rmst(1,mu,P, alpha, nu) t2 - dmst(t1,mu,P,alpha,nu) please supply the ingredients needed to reproduce the problem that you have faced (including the values of the parameters mu,P,alpha,nu, among the rest) best wishes, Adelchi Azzalini [[alternative HTML version deleted]] __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
You need to learn to supply adequate information. The current version of sn *does* have such an argument, and I was careful to check. So it seems that you are using an unstated obselete version of sn. Do ugrade as the posting guide asked you to. On Tue, 28 Mar 2006, Konrad Banachewicz wrote: On 3/28/06, Prof Brian Ripley [EMAIL PROTECTED] wrote: Try maximizing the log-likelihood and using the log=TRUE argument to dmst. seems like dmst does not support this argument (the way e.g. dt does) (You have told us so little about what you are doing that we can but guess at what you mean by `write an mle procedure': what is wrong with st.mle, for example?) st.mle assumes skewed-t marginals (for a whole distribution), whereas I am working with a copula so my margins are uniform. The whole point is separating the joint and marginal dynamics. rg, konrad -- Brian D. Ripley, [EMAIL PROTECTED] Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UKFax: +44 1865 272595 __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
On Tue, 28 Mar 2006 13:11:02 +0200, Konrad Banachewicz wrote: KB P is an identity matrix 240X240, mu and alpha are vectors of zeros KB 240X1, nu equals 10, so alltogether You need: KB P - matrix(0,244,244) KB diag(P) - 1 KB nu - 10 KB alpha - rep(0,244) KB mu - rep(0,244) KB require(sn) KB t1 - rmst(1,mu,P, alpha, nu) KB t2 - dmst(t1,mu,P,alpha,nu) KB KB With such a large dimension, numerical problems are obiquitous. At the very least, I suggest that you work on the log scale: R t2 - dmst(t1,mu,P,alpha,nu, log=TRUE) R t2 [1] -250.3 R exp(t2) [1] 2.002e-109 The next release of the 'sn' package will handle this sort of things in a more consisent way (the R code is largely updated, but the documentation is far behind..) -- Adelchi Azzalini [EMAIL PROTECTED] Dipart.Scienze Statistiche, Università di Padova, Italia tel. +39 049 8274147, http://azzalini.stat.unipd.it/ KB KB KB please supply the ingredients needed to reproduce the problem KB that you have faced (including the values of the parameters KB mu,P,alpha,nu, among the rest) KB KB best wishes, KB KB Adelchi Azzalini KB KB KB __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Skewed t distribution
On Tue, 28 Mar 2006 12:59:34 +0200, Konrad Banachewicz wrote: KB On 3/28/06, Prof Brian Ripley [EMAIL PROTECTED] wrote: KB KB Try maximizing the log-likelihood and using the log=TRUE KB argument to dmst. KB KB KB seems like dmst does not support this argument (the way e.g. dt KB does) KB here I get the following R library(sn) Loading required package: mvtnorm Library 'sn', version 0.3-5 (2005-12-30) , © 1998-2005 A.Azzalini type 'help(SN)' for summary information R args(dmst) function (x, xi = rep(0, d), Omega, alpha, df = Inf, log = FALSE) NULL R notice that 0.3-5 is the current version on CRAN KB KB KB (You have told us so little about what you are doing that we can KB but guess at what you mean by `write an mle procedure': what is KB wrong with st.mle, for example?) KB KB KB st.mle assumes skewed-t marginals (for a whole distribution), KB whereas I am working with a copula so my margins are uniform. The KB whole point is separating the joint and marginal dynamics. KB I am not a copula expert, but what Brian Ripley suggests makes sense to me; and I know of people that have used st.mle to obtain the marginal components (which is what is needed for the copula mechanism) -- Adelchi Azzalini [EMAIL PROTECTED] Dipart.Scienze Statistiche, Università di Padova, Italia tel. +39 049 8274147, http://azzalini.stat.unipd.it/ __ R-help@stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html