wilc.stat: SAM Analysis Using Wilcoxon Rank Statistics

Description

Generates the required statistics for a Significance Analysis of Microarrays analysis using standardized Wilcoxon rank statistics. Should not be called directly, but via sam(..., method = wilc.stat).

Usage

wilc.stat(data, cl, gene.names = NULL, R.fold = 1, use.dm = FALSE, R.unlog = TRUE, na.replace = TRUE, na.method = "mean",  approx50 = TRUE, ties.method=c("min","random","max"),  use.row = FALSE, rand = NA)

Arguments

data

a matrix or a data frame. Each row of data must correspond to a variable (e.g., a gene), and each column to a sample (i.e.\ an observation).

a numeric vector of length ncol(data) containing the class labels of the samples. In the two class paired case, cl can also be a matrix with ncol(data) rows and 2 columns. For details on how cl should be specified, see ?sam.

gene.names

a character vector of length nrow(data) containing the names of the genes.

R.fold

a numeric value. If the fold change of a gene is smaller than or equal to R.fold, or larger than or equal to 1/R.fold,respectively, then this gene will be excluded from the SAM analysis. The expression score $d$ of excluded genes is set to NA. By default, R.fold is set to 1 such that all genes are included in the SAM analysis. Setting R.fold to 0 or a negative value will avoid the computation of the fold change. The fold change is only computed in the two-class unpaired case.

use.dm

if TRUE, the fold change is computed by 2 to the power of the difference between the mean log2 intensities of the two groups, i.e.\ 2 to the power of the numerator of the test statistic. If FALSE, the fold change is determined by computing 2 to the power of data (if R.unlog = TRUE) and then calculating the ratio of the mean intensity in the group coded by 1 to the mean intensity in the group coded by 0. The latter is the default, as this definition of the fold change is used in Tusher et al.\ (2001).

R.unlog

if TRUE, the anti-log of data will be used in the computation of the fold change. Otherwise, data is used. This transformation should be done if data is log2-tranformed. (In a SAM analysis, it is highly recommended to use log2-transformed expression data.) Ignored if use.dm = TRUE.

na.replace

if TRUE, missing values will be removed by the genewise/rowwise statistic specified by na.method. If a gene has less than 2 non-missing values, this gene will be excluded from further analysis. If na.replace = FALSE, all genes with one or more missing values will be excluded from further analysis. The expression score $d$ of excluded genes is set to NA.

na.method

a character string naming the statistic with which missing values will be replaced if na.replace=TRUE. Must be either "mean" (default) or median.

approx50

if TRUE, the null distribution will be approximated by the standard normal distribution. Otherwise, the exact null distribution is computed. This argument will automatically be set to FALSE if there are less than 50 samples in each of the groups.

ties.method

either "min" (default), "random", or "max". If "random", the ranks of ties are randomly assigned. If "min" or "max", the ranks of ties are set to the minimum or maximum rank, respectively. For details, see the help of rank. If use.row = TRUE, ties.method = "max" will be used. For the handling of Zeros, see Details.

use.row

if TRUE, rowWilcoxon is used to compute the Wilcoxon rank statistics.

rand

numeric value. If specified, i.e. not NA, the random number generator will be set into a reproducible state.

Value

A list containing statistics required by sam.

Details

Standardized versions of the Wilcoxon rank statistics are computed. This means that $W* = (W - mean(W)) / sd(W)$ is used as expression score $d$, where $W$ is the usual Wilcoxon rank sum statistic or Wilcoxon signed rank statistic, respectively. In the computation of these statistics, the ranks of ties are by default set to the minimum rank. In the computation of the Wilcoxon signed rank statistic, zeros are randomly set either to a very small positive or negative value. If there are less than 50 observations in each of the groups, the exact null distribution will be used. If there are more than 50 observations in at least one group, the null distribution will by default be approximated by the standard normal distribution. It is, however, still possible to compute the exact null distribution by setting approx50 to FALSE.

References

Schwender, H., Krause, A. and Ickstadt, K. (2003). Comparison of the Empirical Bayes and the Significance Analysis of Microarrays. Technical Report, SFB 475, University of Dortmund, Germany.

Tusher, V.G., Tibshirani, R., and Chu, G. (2001). Significance analysis of microarrays applied to the ionizing radiation response. PNAS, 98, 5116-5121.