On Sun, Aug 12, 2012 at 1:46 PM, Boel Brynedal <bryne...@gmail.com> wrote: > A clarification - yes, calculating the pearson covariance does give > the expected results. I dont fully understand why yet, but many thanks > for this help!
I'm not sure that the spearman correlation is an appropriate estimator for the covariance matrix of a multivariate normal, which is defined in terms of the pearson correlation matrix. (More bluntly, pearson and spearman are different measures and one won't converge to the other) A quick unscientific test: ################# set.seed(1) covMat <- matrix(c(1, 0.4875, 0.4875, 1), 2, 2) # Arbitrary library(MASS) library(TTR) # For fast pearson cor n <- 5000 rands <- mvrnorm(n, c(0,0), covMat, empirical = TRUE) # I'm pretty sure we want empirical = TRUE runPearson <- runCor(rands[,1], rands[,2], cumulative = TRUE) # This takes a little while but I'm doing my best to make it fast ;-) runSpearman <- vapply(seq(10, n), function(n) cor(rands[seq_len(n), ], method = "spearman")[2], numeric(1)) plot(runPearson, type = "l") lines(runSpearman, col = 2) # Show that we get good/decent convergence abline(h = covMat[2], col = 3) ############## That long stable difference suggests to me that you don't want to use cor(,,"spearman") to estimate a quantity defined in terms of cor(,, "pearson"). I am not sure if this is a general/fixed bias in the spearman estimator or if it's just a function of the covMat I randomly chose. Prof. Dalgaard and many others on this list must know. Cheers, Michael ______________________________________________ R-help@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.
