chisq.test: Avoid overcorrection

git-svn-id: https://svn.r-project.org/R/branches/R-2-15-branch@59556 00db46b3-68df-0310-9c12-caf00c1e9a41

doc/NEWS.Rd

@@ -6,7 +6,7 @@
 \title{ R News}
 \encoding{UTF-8}
 
-\section{\Rlogo CHANGES IN R VERSION 2.15.0 patched}{
+\section{\Rlogo CHANGES IN R VERSION 2.15.1}{
   \subsection{NEW FEATURES}{
     \itemize{
       \item \code{source()} now uses \code{withVisible()} rather than
@@ -306,6 +306,10 @@
 
       \item \code{by()} failed for a zero-row data frame.  (Reported by
       Weiqiang Qian)
+
+      \item Yates correction in \code{chisq.test()} could be bigger than 
+      the terms it corrected, leading to an infinite test statistic in 
+      some corner cases. 
     }
   }
 }

src/library/stats/R/chisq.test.R

@@ -98,8 +98,9 @@ chisq.test <- function(x, y = NULL, correct = TRUE,
 	    if (simulate.p.value)
 		warning("cannot compute simulated p-value with zero marginals")
 	    if (correct && nrow(x) == 2 && ncol(x) == 2) {
-		YATES <- 0.5
-		METHOD <- paste(METHOD, "with Yates' continuity correction")
+		YATES <- min(0.5, abs(x-E))
+                if (YATES > 0)
+		    METHOD <- paste(METHOD, "with Yates' continuity correction")
 	    }
 	    else
 		YATES <- 0

src/library/stats/man/chisq.test.Rd

@@ -23,7 +23,8 @@ chisq.test(x, y = NULL, correct = TRUE,
     \code{x} is a factor, \code{y} should be a factor of the same length.}
   \item{correct}{a logical indicating whether to apply continuity
     correction when computing the test statistic for 2 by 2 tables: one
-    half is subtracted from all \eqn{|O - E|} differences.  No correction
+    half is subtracted from all \eqn{|O - E|} differences; however, the 
+    correction will not be bigger than the differences themselves.  No correction
     is done if \code{simulate.p.value = TRUE}.}
   \item{p}{a vector of probabilities of the same length of \code{x}.
     An error is given if any entry of \code{p} is negative.}