The R Project SVN R

Rev

Rev 89312 | Show entire file | Ignore whitespace | Details | Blame | Last modification | View Log | RSS feed

Rev 89312 Rev 89313
Line 53... Line 53...
53 
  \I{Hochberg}'s, but the difference is usually small and the \I{Hochberg}
 53 
  \I{Hochberg}'s, but the difference is usually small and the \I{Hochberg}
54 
  p-values are faster to compute.
 54 
  p-values are faster to compute.
55 
 
 55 
 
56 
  The \code{"BH"} (aka \code{"fdr"}) and \code{"BY"} methods of
 56 
  The \code{"BH"} (aka \code{"fdr"}) and \code{"BY"} methods of
57 
  \I{Benjamini}, \I{Hochberg}, and \I{Yekutieli} control the false discovery rate,
 57 
  \I{Benjamini}, \I{Hochberg}, and \I{Yekutieli} control the false discovery rate,
58 
  \I{FDR}, the expected proportion of false discoveries amongst the
 58 
  \acronym{FDR}, the expected proportion of false discoveries amongst the
59 
  rejected hypotheses.
 59 
  rejected hypotheses.
60 
 
 60 
 
61 
  The \code{"BY"} correction modifies the \I{BH} procedure by replacing the
 61 
  The \code{"BY"} correction modifies the \acronym{BH} procedure by replacing the
62 
  target level \eqn{q} with \eqn{q / \sum_{i=1}^{m} 1/i}, where \eqn{m} is
 62 
  target level \eqn{q} with \eqn{q / \sum_{i=1}^{m} 1/i}, where \eqn{m} is
63 
  the number of tests (Theorem 1.3 in the reference), which controls the
 63 
  the number of tests (Theorem 1.3 in the reference), which controls the
64 
  \I{FDR} under the most general form of dependence structure.  This will
 64 
  \acronym{FDR} under the most general form of dependence structure.  This will
65 
  be more conservative than \code{"BH"}, for small \code{p} even more than
 65 
  be more conservative than \code{"BH"}, for small \code{p} even more than
66 
  \I{Bonferroni}, see the example.
 66 
  \I{Bonferroni}, see the example.
67 
  The \I{FDR} as implemented by the \code{"BH"} method is a less stringent
 67 
  The \acronym{FDR} as implemented by the \code{"BH"} method is a less stringent
68 
  condition than the family-wise error rate, so it is typically more
 68 
  condition than the family-wise error rate, so it is typically more
69 
  powerful than the others.
 69 
  powerful than the others.
70 
 
 70 
 
71 
  Note that you can set \code{n} larger than \code{length(p)} which
 71 
  Note that you can set \code{n} larger than \code{length(p)} which
72 
  means the unobserved p-values are assumed to be greater than all the
 72 
  means the unobserved p-values are assumed to be greater than all the