knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) options(digits = 3)
AnthropMMD was primarily designed as an R-Shiny application [@Santos18], that can be launched with the following instruction:
Starting with the version 3.0.0, it can also be used from the command line, to make it convenient in a context of reproducible research.
AnthropMMDfrom the command line
The MMD is a dissimilarity measure among groups of individuals described by binary (presence-absence) traits [@Sjovold73; @Sjovold77; @Harris04]. The MMD formula only requires to know the sample sizes and relative trait frequencies within each group: providing the raw individual data is not mandatory.
Usually, the user will have recorded the data in one of the the two following formats.
Most of the time, the data will be formatted as a classical dataframe with $n$ rows (one per individual) and $p+1$ columns (the first column must be a group indicator; the $p$ other columns correspond to the $p$ binary traits observed). Each trait has two possible values: $1$ if present, $0$ if absent (missing values are allowed). Rownames are optional but colnames (i.e.,
headers) are mandatory.
An artificial dataset,
toyMMD, is available in the package to give an example of such a format:
It is not necessary for the
Traits to be manually converted as factors prior to the analysis.
toyMMD includes nine traits with many missing values, and the individuals belong to five groups:
If the data were recorded as a raw binary dataset, they must be converted into a table of relative frequencies prior to the MMD analysis. The R function
binary_to_table with the argument
relative = TRUE can perform this operation:
tab <- binary_to_table(toyMMD, relative = TRUE) tab
Sometimes, in particular if the data were extracted from research articles rather than true exprimentation, the user will only know the sample sizes and absolute trait frequencies in each group. As previously stated, those data are perfectly sufficient to compute MMD values. Here is an example of such a table:
If there are $k$ groups in the data, the first $k$ rows (whose names must start with an
N_ prefix, followed by the group labels) indicates the sample sizes of each trait in each group, and the last $k$ rows indicates the absolute trait frequencies. Such a table can be directly imported by the user, but must be converted to relative trait frequencies prior to the analysis, for instance by using the function
tab <- table_relfreq(absolute_freqs) print(tab)
Due to rounding issues, it is clearly not advised for the user to submit directly a table of relative frequencies. If the absolute frequencies do not perfectly sum up to 1 for each trait, an error will be triggered.
A table of relative frequencies is the starting point of all subsequent analyses, but it should be computed from the data loaded in R; it should not be loaded itself.
A quick summary:
binary_to_table(relative = TRUE)as shown above;
AnthropMMD proposes a built-in feature for trait selection, in order to discard the traits that could be observed on too few individuals, or that do not show enough variability among groups. The function
select_traits allows such a selection according to several strategies [@Harris04; @Santos18].
For instance, with the following instruction, we can discard the traits that could be observed on at least 10 individuals per group, and that exhibit no significant variability in frequencies among groups according to Fisher's exact tests:
tab_selected <- select_traits(tab, k = 10, strategy = "keepFisher") tab_selected$filtered
Trait1 (which could be observed on only 8 individuals in Group B),
Trait8 (whose variation in frequencies was not significant among groups) were removed from the data.
Once the trait selection has been performed, the MMD can be computed with the
mmd function. With the following instruction, the MMD is computed using Anscombe's angular transformation of trait frequencies:
mmd.result <- mmd(tab_selected$filtered, angular = "Anscombe") mmd.result
$MMDMatrix, follows the presentation adopted in most research articles [@Donlon00]: the true MMD values are indicated above the diagonal, and their standard deviations are indicated below the diagonal.
*in the component
$MMDSymis a symmetrical matrix of MMD values, with a null diagonal. This matrix of dissimilarities can be used to perform a multidimensional scaling or a hierarchical clustering to visualize the distances among the groups.
$MMDpval. Theoretical details for this test of significance can be found in @Souza77.
mmd.result$MMDSym can perfectly be passed to your favourite function to produce a MDS plot,
AnthropMMD also proposes a built-in generic function for such a graphical representation:
The MDS can be computed using the classical
stats::cmdscale function (and the produces a metric MDS), or several variants of MDS algorithms implemented in the R package
For instance, we plot here the MDS coordinates computed with one variant of SMACOF algorithms:
par(cex = 0.8) plot(x = mmd.result, method = "interval", gof = TRUE, axes = TRUE, xlim = c(-1.2, 0.75))
gof = TRUE displays goodness of fits statistics for the MDS configurations directly on the plot.
AnthropMMD does not propose a built-in function for hierarchical clustering, but such a plot can easily be obtained with the usual R functions.
library(cluster) par(cex = 0.8) plot(agnes(mmd.result$MMDSym), which.plots = 2, main = "Dendrogram of MMD dissimilarities")
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.