[Vegan-commits] r1869 - in pkg/vegan: R inst

[noreply at r-forge.r-project.org]

Author: jarioksa
Date: 2011-09-22 12:26:09 +0200 (Thu, 22 Sep 2011)
New Revision: 1869

Modified:
   pkg/vegan/R/meandist.R
   pkg/vegan/inst/ChangeLog
Log:
meandist could internally reorder the 'grouping' levels and return a randomly permuted matrix

Modified: pkg/vegan/R/meandist.R
===================================================================
--- pkg/vegan/R/meandist.R	2011-09-21 19:27:18 UTC (rev 1868)
+++ pkg/vegan/R/meandist.R	2011-09-22 10:26:09 UTC (rev 1869)
@@ -1,26 +1,27 @@
 `meandist` <-
     function(dist, grouping, ...)
 {
+    if (!inherits(dist, "dist"))
+        stop("'dist' must be dissimilarity object inheriting from", dQuote(dist))
     ## check that 'dist' are dissimilarities (non-negative)
     if (any(dist < -sqrt(.Machine$double.eps)))
         warning("some dissimilarities are negative -- is this intentional?")
-    ## merge levels so that lower is always first (filling lower triangle)
-    mergenames <- function(X, Y, ...) {
-        xy <- cbind(X, Y)
-        xy <- apply(xy, 1, sort)
-        apply(xy, 2, paste, collapse = " ")
-    }
     grouping <- factor(grouping, exclude = NULL)
-    cl <- outer(grouping, grouping, mergenames)
-    cl <- cl[lower.tri(cl)]
+    ## grouping for rows and columns
+    grow <- grouping[as.dist(row(as.matrix(dist)))]
+    gcol <- grouping[as.dist(col(as.matrix(dist)))]
+    ## The row index must be "smaller" of the factor levels so that
+    ## all means are in the lower triangle, and upper is NA
+    first <- as.numeric(grow) >= as.numeric(gcol)
+    cl1 <- ifelse(first, grow, gcol)
+    cl2 <- ifelse(!first, grow, gcol)
     ## Cannot have within-group dissimilarity for group size 1
     n <- table(grouping)
     take <- matrix(TRUE, nlevels(grouping), nlevels(grouping))
     diag(take) <- n > 1
     take[upper.tri(take)] <- FALSE
     ## Get output matrix
-    out <- matrix(NA, nlevels(grouping), nlevels(grouping))
-    out[take] <- tapply(dist, cl, mean)
+    out <- tapply(dist, list(cl1, cl2), mean)
     out[upper.tri(out)] <- t(out)[upper.tri(out)]
     rownames(out) <- colnames(out) <- levels(grouping)
     class(out) <- c("meandist", "matrix")

Modified: pkg/vegan/inst/ChangeLog
===================================================================
--- pkg/vegan/inst/ChangeLog	2011-09-21 19:27:18 UTC (rev 1868)
+++ pkg/vegan/inst/ChangeLog	2011-09-22 10:26:09 UTC (rev 1869)
@@ -4,31 +4,37 @@
 
 Version 2.1-1 (opened September 20, 2011)
 
+	* meandist bug fix: tapply() function used to find mean group x
+	group dissimilarities could reorder the class levels and return a
+	confused matrix. This could happen in particular when the
+	'grouping' was a vector of integers which then were ordered
+	alphabetically so that "1" < "10" < "2". Now uses internally more
+	stable way of applying tapply() which should have the danger of
+	reordering the levels.  Incidentally, this also seems to be
+	faster. The problem was found by Dr Miguel Alvarez (Univ Bonn).
+
 	* nestedness.c: changed interface in "swapcount" and "rswapcount"
 	which now require integer data matrix. The first argument in .C
 	call should now be defined as as.integer() instead of old
 	as.double().
 
 	* New functions: commsim is used to define Null Model Algorithms
-	via a function that returns n x m x nsim array of permuted
-	matrices based on structural constraints. make.commsim
-	contains Null Model Algorithms already defined in vegan 
-	from commsimulator and permat* functions (and some more).
-	The nullmodel function creates an environment, where statistics
-	of the input matrix are stored. The environment also stores
-	updated status of sequential algorithms and current number of 
-	iterations. The update and simulate methods are used to
-	update the nullmodel (for sequential algorithms) or
-	simulate random matrices, respectively. The simulate
-	method returns the n x m nsim array (simmat class).
-	Efficiency gains are sometimes high (because marginal
-	statistics are calculate only once by nullmodel), 
-	but not significant in most cases. 
-	Most advanageously, this implementation
-	can unite the commsimulator and permat* branches
-	and can serve as basis for further extensions.
-	Current intent is to investigate how this low level
-	infrastructure can be used within oecosimu and permat*
+	via a function that returns n x m x nsim array of simulated
+	matrices based on structural constraints. make.commsim contains
+	Null Model Algorithms already defined in vegan from commsimulator
+	and permat* functions (and some more).  The nullmodel function
+	creates an environment, where statistics of the input matrix are
+	stored. The environment also stores updated status of sequential
+	algorithms and current number of iterations. The update and
+	simulate methods are used to update the nullmodel (for sequential
+	algorithms) or simulate random matrices, respectively. The
+	simulate method returns the n x m x nsim array (simmat class).
+	Efficiency gains are sometimes high (because marginal statistics
+	are calculate only once by nullmodel), but not significant in most
+	cases.  Most advantageously, this implementation can unite the
+	commsimulator and permat* branches and can serve as basis for
+	further extensions.  Current intent is to investigate how this low
+	level infrastructure can be used within oecosimu and permat*
 	functions without breaking current vegan functionality.
 
 Version 2.1-0 (closed September 20, 2011)