21.2 Developing the package

used to print and plot the results. Here, oneway indicates that there is a single grouping factor.

It’s a good idea to place each function in a separate text file with a .R extension. This isn’t strictly necessary, but it makes organizing the work easier. Additionally, it isn’t necessary for the names of the functions and the names of the files to match, but again, it’s good coding practice. The files are provided in listings 21.2 through 21.5.

Each file has a header consisting of a set of comments that start with the characters #'. The R interpreter ignores these lines, but you’ll use the roxygen2 package to turn the comments into your package documentation. These header comments will be discussed in section 21.3.

The oneway() function computes the statistics, and the print(), summary(), and plot() functions display the results. In the next section, you’ll develop the oneway() function.

21.2.1Computing the statistics

The oneway() function in the oneway.R text file performs all the statistical computations required.

Listing 21.2 Contents of the oneway.R file

#' @title Nonparametric group comparisons #'

#' @description

#' \code{oneway} computes nonparametric group comparisons, including an #' omnibus test and post-hoc pairwise group comparisons.

#'

#' @details

#' This function computes an omnibus Kruskal-Wallis test that the #' groups are equal, followed by all pairwise comparisons using

#' Wilcoxon Rank Sum tests. Exact Wilcoxon tests can be requested if #' there are no ties on the dependent variable. The p-values are

#' adjusted for multiple comparisons using the \code{\link{p.adjust}} #' function.

#'

#' @param formula an object of class formula, relating the dependent #' variable to the grouping variable.

#' @param data a data frame containing the variables in the model.

#' @param exact logical. If \code{TRUE}, calculate exact Wilcoxon tests. #' @param sort logical. If \code{TRUE}, sort groups by median dependent #' variable values.

#' @param method method for correcting p-values for multiple comparisons. #' @export

#' @return a list with 7 elements: #' \item{CALL}{function call}

#' \item{data}{data frame containing the depending and grouping variable} #' \item{sumstats}{data frame with descriptive statistics by group}

#' \item{kw}{results of the Kruskal-Wallis test} #' \item{method}{method used to adjust p-values}

#' \item{wmc}{data frame containing the multiple comparisons} #' \item{vnames}{variable names}

#' @author Rob Kabacoff <rkabacoff@@statmethods.net>

function argument list b. The user provides a formula of the form dependent variable~grouping variable and a data frame containing the data. By default, approximate p-values are computed, and the groups are ordered by their median dependent variable values. The user can choose from among eight adjustment methods, with the holm method (the first option in the list) chosen by default.

Once the user enters the arguments, they’re scanned for errors c. The if() function tests that the formula isn’t missing, that it’s a formula (variables ~ variables), and that there is only one variable on each side of the tilde (~). If any of these three conditions isn’t true, the stop() function halts execution, prints an error message, and returns the user to the R prompt. For debugging purposes, you can alter the error action with the options(error=) function. See section 20.5.3 for details.

The match.arg(arg, choices) function ensures that the user has entered an argument that matches one of the strings in the choices character vector. If a match isn’t found, an error is thrown, and, again, oneway() exits.

Next, the model.frame() function is used to create a data frame containing the dependent variable as the first column and the grouping variable as the second column d. In general, model.frame() returns a data frame containing all the variables in a formula. From this data frame, you create a numeric vector (y) containing the dependent variable and a factor vector (g) containing the grouping variable. The character vector vnames contains the variable names.

If sort=TRUE, you use the reorder() function to reorder the levels of the grouping variable g by the median dependent variable values y e. This is the default. The character vector groups contains the names of the groups, and the value k contains the number of groups.

Next, a numeric matrix (sumstats) is created, containing the sample size, median, and median absolute deviation for each group f. The aggregate() function uses the getstats() function to calculate the summary statistics, and the remaining code formats the table so that groups are columns and statistics are rows (I thought this was more attractive).

The statistical tests are then computed g. The results of the Kruskal–Wallis test are saved to a list called kw. The for() functions calculate every pairwise Wilcoxon test. The results of these pairwise tests are saved in the wmc data frame:

Here, Group.1 and Group.2 indicate the groups being compared to each other, W is the Wilcoxon statistic, and p is the (adjusted) p-value for each comparison.

Finally, the results are bundled up and returned as a list h. The list contains seven components, which are summarized in table 21.1. Additionally, you set the class of the