4 Details of specific functions

The following section contains information specific to some functions. If any of your questions are not covered in these sections, please refer to the function help files in R, send me an email (guillert@tcd.ie), or raise an issue on GitHub. The several tutorials below describe specific functionalities of certain functions; please always refer to the function help files for the full function documentation!

Before each section, make sure you loaded the Beck and Lee (2014) data (see example data for more details).

4.1 Time slicing

The function chrono.subsets allows users to divide the matrix into different time subsets or slices given a dated phylogeny that contains all the elements (i.e. taxa) from the matrix. Each subset generated by this function will then contain all the elements present at a specific point in time or during a specific period in time.

Two types of time subsets can be performed by using the method option:

  • Discrete time subsets (or time-binning) using method = discrete
  • Continuous time subsets (or time-slicing) using method = continuous

For the time-slicing method details see Guillerme and Cooper (2018). For both methods, the function takes the time argument which can be a vector of numeric values for:

  • Defining the boundaries of the time bins (when method = discrete)
  • Defining the time slices (when method = continuous)

Otherwise, the time argument can be set as a single numeric value for automatically generating a given number of equidistant time-bins/slices. Additionally, it is also possible to input a dataframe containing the first and last occurrence data (FAD/LAD) for taxa that span over a longer time than the given tips/nodes age, so taxa can appear in more than one time bin/slice.

Here is an example for method = discrete:

##  ---- dispRity object ---- 
## 3 discrete time subsets for 50 elements:
##     120 - 80, 80 - 40, 40 - 0.

Note that we can also generate equivalent results by just telling the function that we want three time-bins as follow:

##  ---- dispRity object ---- 
## 3 discrete time subsets for 50 elements:
##     133.51104 - 89.00736, 89.00736 - 44.50368, 44.50368 - 0.

In this example, the taxa were split inside each time-bin according to their age. However, the taxa here are considered as single points in time. It is totally possible that some taxa could have had longer longevity and that they exist in multiple time bins. In this case, it is possible to include them in more than one bin by providing a table of first and last occurrence dates (FAD/LAD). This table should have the taxa names as row names and two columns for respectively the first and last occurrence age:

##             FAD  LAD
## Adapis     37.2 36.8
## Asioryctes 83.6 72.1
## Leptictis  33.9 33.3
## Miacis     49.0 46.7
## Mimotona   61.6 59.2
## Notharctus 50.2 47.0
##  ---- dispRity object ---- 
## 3 discrete time subsets for 50 elements:
##     120 - 80, 80 - 40, 40 - 0.

When using this method, the oldest boundary of the first bin (or the first slice, see below) is automatically generated as the root age plus 1% of the tree length, as long as at least three elements/taxa are present at that point in time. The algorithm adds an extra 1% tree length until reaching the required minimum of three elements. It is also possible to include nodes in each bin by using inc.nodes = TRUE and providing a matrix that contains the ordinated distance among tips and nodes.

For the time-slicing method (method = continuous), the idea is fairly similar. This option, however, requires a matrix that contains the ordinated distance among taxa and nodes and an extra argument describing the assumed evolutionary model (via the model argument). This model argument is used when the time slice occurs along a branch of the tree rather than on a tip or a node, meaning that a decision must be made about what the value for the branch should be. The model can be one of the following:

  • Punctuated models
    • acctran where the data chosen along the branch is always the one of the descendant
    • deltran where the data chosen along the branch is always the one of the ancestor
    • random where the data chosen along the branch is randomly chosen between the descendant or the ancestor
    • proximity where the data chosen along the branch is either the descendant or the ancestor depending on branch length
  • Gradual models
    • equal.split where the data chosen along the branch is both the descendant and the ancestor with an even probability
    • gradual.split where the data chosen along the branch is both the descendant and the ancestor with a probability depending on branch length

Note that the four first models are a proxy for punctuated evolution: the selected data is always either the one of the descendant or the ancestor. In other words, changes along the branches always occur at either ends of it. The two last models are a proxy for gradual evolution: the data from both the descendant and the ancestor is used with an associate probability. These later models perform better when bootstrapped, effectively approximating the “intermediate” state between and the ancestor and the descendants.

##  ---- dispRity object ---- 
## 4 continuous (proximity) time subsets for 99 elements:
##     120, 80, 40, 0.
##  ---- dispRity object ---- 
## 4 continuous (proximity) time subsets for 99 elements:
##     133.51104, 89.00736, 44.50368, 0.

If you want to generate time subsets based on stratigraphy, the package proposes a useful functions to do it for you: get.bin.ages (check out the function’s manual in R)!

4.2 Customised subsets

Another way of separating elements into different categories is to use customised subsets as briefly explained above. This function simply takes the list of elements to put in each group (whether they are the actual element names or their position in the matrix).

##  ---- dispRity object ---- 
## 2 customised subsets for 50 elements:
##     crown, stem.

Like in this example, you can use the utility function crown.stem that allows to automatically separate the crown and stems taxa given a phylogenetic tree. Also, elements can easily be assigned to different groups if necessary!

The custom.subsets function can also take a phylogeny (as a phylo object) as an argument to create groups as clades:

This automatically creates 49 (the number of nodes) groups containing between two and 50 (the number of tips) elements.

4.3 Bootstraps and rarefactions

One important step in analysing ordinated matrices is to pseudo-replicate the data to see how robust the results are, and how sensitive they are to outliers in the dataset. This can be achieved using the function boot.matrix to bootstrap and/or rarefy the data. The default options will bootstrap the matrix 100 times without rarefaction using the “full” bootstrap method (see below):

##  ---- dispRity object ---- 
## 50 elements with 48 dimensions.
## Data was bootstrapped 100 times (method:"full").

The number of bootstrap replicates can be defined using the bootstraps option. The method can be modified by controlling which bootstrap algorithm to use through the boot.type argument. Currently two algorithms are implemented:

  • full where the bootstrapping is entirely stochastic (n elements are replaced by any m elements drawn from the data)
  • single where only one random element is replaced by one other random element for each pseudo-replicate
##  ---- dispRity object ---- 
## 50 elements with 48 dimensions.
## Data was bootstrapped 100 times (method:"single").

This function also allows users to rarefy the data using the rarefaction argument. Rarefaction allows users to limit the number of elements to be drawn at each bootstrap replication. This is useful if, for example, one is interested in looking at the effect of reducing the number of elements on the results of an analysis.

This can be achieved by using the rarefaction option that draws only n-x at each bootstrap replicate (where x is the number of elements not sampled). The default argument is FALSE but it can be set to TRUE to fully rarefy the data (i.e. remove x elements for the number of pseudo-replicates, where x varies from the maximum number of elements present in each subset to a minimum of three elements). It can also be set to one or more numeric values to only rarefy to the corresponding number of elements.

##  ---- dispRity object ---- 
## 50 elements with 48 dimensions.
## Data was bootstrapped 20 times (method:"full") and fully rarefied.
##  ---- dispRity object ---- 
## 50 elements with 48 dimensions.
## Data was bootstrapped 20 times (method:"full") and rarefied to 6, 7, 8, 3 elements.

One other argument is dimensions that specifies how many dimensions from the matrix should be used for further analysis. When missing, all dimensions from the ordinated matrix are used.

##  ---- dispRity object ---- 
## 50 elements with 24 dimensions.
## Data was bootstrapped 100 times (method:"full").
##  ---- dispRity object ---- 
## 50 elements with 10 dimensions.
## Data was bootstrapped 100 times (method:"full").

It is also possible to specify the sampling probability in the bootstrap for each elements. This can be useful for weighting analysis for example (i.e. giving more importance to specific elements). These probabilities can be passed to the prob argument individually with a vector with the elements names or with a matrix with the rownames as elements names. The elements with no specified probability will be assigned a probability of 1 (or 1/maximum weight if the argument is weights rather than probabilities).

##  ---- dispRity object ---- 
## 50 elements with 48 dimensions.
## Data was bootstrapped 100 times (method:"full").

Of course, one could directly supply the subsets generated above (using chrono.subsets or custom.subsets) to this function.

##  ---- dispRity object ---- 
## 2 customised subsets for 50 elements with 48 dimensions:
##     crown, stem.
## Data was bootstrapped 200 times (method:"full") and fully rarefied.
##  ---- dispRity object ---- 
## 4 continuous (proximity) time subsets for 99 elements with 97 dimensions:
##     120, 80, 40, 0.
## Data was bootstrapped 100 times (method:"full").

4.4 Disparity metrics

There are many ways of measuring disparity! In brief, disparity is a summary metric that will represent an aspect of an ordinated space (e.g. a MDS, PCA, PCO, PCoA). For example, one can look at ellipsoid hyper-volume of the ordinated space (Donohue et al. 2013), the sum and the product of the ranges and variances (Wills et al. 1994) or the median position of the elements relative to their centroid (Wills et al. 1994). Of course, there are many more examples of metrics one can use for describing some aspect of the ordinated space, with some performing better than other ones at particular descriptive tasks, and some being more generalist.

Because of this great diversity of metrics, the package dispRity does not have one way to measure disparity but rather proposes to facilitate users in defining their own disparity metric that will best suit their particular analysis. In fact, the core function of the package, dispRity, allows the user to define any metric with the metric argument. However the metric argument has to follow certain rules:

  1. It must be composed from one to three function objects;
  2. The function(s) must take as a first argument a matrix or a vector;
  3. The function(s) must be of one of the three dimension-levels described below;
  4. At least one of the functions must be of dimension-level 1 or 2 (see below).

4.4.1 The function dimension-levels

The metric function dimension-levels determine the “dimensionality of decomposition” of the input matrix. In other words, each dimension-level designates the dimensions of the output, i.e. either three (a matrix); two (a vector); or one (a single numeric value) dimension.

Illustration of the different dimension-levels of functions with an input matrix

Illustration of the different dimension-levels of functions with an input matrix

4.4.1.1 Dimension-level 1 functions

A dimension-level 1 function will decompose a matrix or a vector into a single value:

## [1] 0.2986208
## [1] 0.3686598

Any summary metric such as mean or median are good examples of dimension-level 1 functions as they reduce the matrix to a single dimension (i.e. one value).

4.4.1.2 Dimension-level 2 functions

A dimension-level 2 function will decompose a matrix into a vector.

## [1]  0.06266985 -0.18910202  0.73229527  1.36514343

Several dimension-level 2 functions are implemented in dispRity (see ?dispRity.metric) such as the variances or ranges functions that calculate the variance or the range of each dimension of the ordinated matrix respectively.

4.4.1.3 Dimension-level 3 functions

Finally a dimension-level 3 function will transform the matrix into another matrix. Note that the dimension of the output matrix doesn’t need to match the the input matrix:

##            [,1]       [,2]       [,3]
## [1,]  0.6893456 -0.1707521 -0.1734104
## [2,] -0.1707521  1.1508417  1.0649556
## [3,] -0.1734104  1.0649556  0.9860865
##          1        2        3        4
## 1 0.000000 1.970182 2.590456 1.162856
## 2 1.970182 0.000000 2.451574 1.927741
## 3 2.590456 2.451574 0.000000 3.499225
## 4 1.162856 1.927741 3.499225 0.000000

4.4.2 make.metric

Of course, functions can be more complex and involve multiple operations such as the centroids function (see ?dispRity.metric) that calculates the Euclidean distance between each element and the centroid of the ordinated space. The make.metric function implemented in dispRity is designed to help test and find the dimension-level of the functions. This function tests:

  1. If your function can deal with a matrix or a vector as an input;
  2. Your function’s dimension-level according to its output (dimension-level 1, 2 or 3, see above);
  3. Whether the function can be implemented in the dispRity function (the function is fed into a lapply loop).

For example, let’s see if the functions described above are the right dimension-levels:

## mean outputs a single value.
## mean is detected as being a dimension-level 1 function.
## prod.rows outputs a matrix object.
## prod.rows is detected as being a dimension-level 2 function.
## var outputs a matrix object.
## var is detected as being a dimension-level 3 function.
## Additional dimension-level 2 and/or 1 function(s) will be needed.

A non verbose version of the function is also available. This can be done using the option silent = TRUE and will simply output the dimension-level of the metric.

## The metric is not dimension-level 1.

4.4.3 Metrics in the dispRity function

Using this metric structure, we can easily use any disparity metric in the dispRity function as follows:

##   subsets  n   obs
## 1       1 50 0.201
##   subsets  n   obs
## 1       1 50 0.028
##   subsets  n obs
## 1       1 50   0

Note that the order of each function in the metric argument does not matter, the dispRity function will automatically detect the function dimension-levels (using make.metric) and apply them to the data in decreasing order (dimension-level 3 > 2 > 1).

##      subsets    n  obs
## [1,]    TRUE TRUE TRUE

In these examples, we considered disparity to be a single value. For example, in the previous example, we defined disparity as the standard deviation of the variances of each column of the variance/covariance matrix (metric = c(variances, sd, var)). It is, however, possible to calculate disparity as a distribution.

4.4.4 Metrics implemented in dispRity

Several disparity metrics are implemented in the dispRity package. The detailed list can be found in ?dispRity.metric along with some description of each metric.

Level Name Description Source
2 ancestral.dist The distance between an element and its ancestor dispRity
2 centroids1 The distance between each element and the centroid of the ordinated space dispRity
1 convhull.surface The surface of the convex hull formed by all the elements geometry::convhulln$area
1 convhull.volume The volume of the convex hull formed by all the elements geometry::convhulln$vol
1 diagonal The longest distance in the ordinated space (like the diagonal in two dimensions) dispRity
2 displacements The ratio between the distance from a reference and the distance from the centroid dispRity
1 ellipse.volume1 The volume of the ellipsoid of the space Donohue et al. (2013)
1 func.div The functional divergence (the ratio of deviation from the centroid) dispRity (similar to FD::dbFD$FDiv but without abundance)
1 func.eve The functional evenness (the minimal spanning tree distances evenness) dispRity (similar to FD::dbFD$FEve but without abundance)
1 mode.val The modal value dispRity
1 n.ball.volume The hyper-spherical (n-ball) volume dispRity
2 neighbours The distance to specific neighbours (e.g. the nearest neighbours - by default) dispRity
2 pairwise.dist The pairwise distances between elements vegan::vegist
2 quantiles The nth quantile range per axis dispRity
2 radius The radius of each dimensions dispRity
2 ranges The range of each dimension dispRity
2 span.tree.length The minimal spanning tree length ape (but see also vegan::spantree)
2 variances The variance of each dimension dispRity

1: Note that by default, the centroid is the centroid of the elements. It can, however, be fixed to a different value by using the centroid argument centroids(space, centroid = rep(0, ncol(space))), for example the origin of the ordinated space.

2: This function uses an estimation of the eigenvalue that only works for MDS or PCoA ordinations (not PCA).

4.4.5 Equations and implementations

Some of the functions described below are implemented in the dispRity package and do not require any other packages to calculate (see implementation here).

\[\begin{equation} ancestral.dist = \sqrt{\sum_{i=1}^{n}{({k}_{n}-Ancestor_{n})^2}} \end{equation}\]

\[\begin{equation} centroids = \sqrt{\sum_{i=1}^{n}{({k}_{n}-Centroid_{k})^2}} \end{equation}\]

\[\begin{equation} diagonal = \sqrt{\sum_{i=1}^{k}|max(k_i) - min(k_i)|} \end{equation}\]

\[\begin{equation} displacements = \frac{\sqrt{\sum_{i=1}^{n}{({k}_{n}-Reference_{k})^2}}}{\sqrt{\sum_{i=1}^{n}{({k}_{n}-Centroid_{k})^2}}} \end{equation}\]

\[\begin{equation} ellipse.volume = \frac{\pi^{k/2}}{\Gamma(\frac{k}{2}+1)}\displaystyle\prod_{i=1}^{k} (\lambda_{i}^{0.5}) \end{equation}\]

\[\begin{equation} n.ball.volume = \frac{\pi^{k/2}}{\Gamma(\frac{k}{2}+1)}\displaystyle\prod_{i=1}^{k} R \end{equation}\]

\[\begin{equation} radius = |\frac{\sum_{i=1}^{n}k_i}{n} - f(\mathbf{v}k)| \end{equation}\]

\[\begin{equation} ranges = |max(k_i) - min(k_i)| \end{equation}\]

\[\begin{equation} variances = \sigma^{2}{k_i} \end{equation}\]

\[\begin{equation} span.tree.length = \mathrm{branch\ length} \end{equation}\]

Where k is the number of dimensions, n the number of elements, \(\Gamma\) is the Gamma distribution, \(\lambda_i\) is the eigenvalue of each dimensions, \(\sigma^{2}\) is their variance and \(Centroid_{k}\) is their mean, \(Ancestor_{n}\) is the coordinates of the ancestor of element \(n\), \(f(\mathbf{v}k)\) is function to select one value from the vector \(\mathbf{v}\) of the dimension \(k\) (e.g. it’s maximum, minimum, mean, etc.), R is the radius of the sphere or the product of the radii of each dimensions (\(\displaystyle\prod_{i=1}^{k}R_{i}\) - for a hyper-ellipsoid), and \(Reference_{k}\) is an arbitrary point’s coordinates (usually 0).

4.4.6 Using the different disparity metrics

Here is a brief demonstration of the main metrics implemented in dispRity. First, we will create a dummy/simulated ordinated space using the space.maker utility function (more about that here:

We will use this simulated space to demonstrate the different metrics.

4.4.6.1 Volumes and surface metrics

The functions ellipse.volume, convhull.surface, convhull.volume and n.ball.volume all measure the surface or the volume of the ordinated space occupied:

Because there is only one subset (i.e. one matrix) in the dispRity object, the operations below are the equivalent of metric(dummy_space) (with rounding).

##   subsets  n   obs
## 1       1 10 1.061

WARNING: in such dummy space, this gives the estimation of the ellipsoid volume, not the real ellipsoid volume! See the cautionary note in ?ellipse.volume.

##   subsets  n   obs
## 1       1 10 11.91
##   subsets  n   obs
## 1       1 10 1.031
##   subsets  n  obs
## 1       1 10 4.43

The convex hull based functions are a call to the geometry::convhulln function with the "FA" option (computes total area and volume). Also note that they are really sensitive to the size of the dataset.

Cautionary note: measuring volumes in a high number of dimensions can be strongly affected by the curse of dimensionality that often results in near 0 disparity values. I strongly recommend reading this really intuitive explanation from Toph Tucker.

4.4.6.2 Ranges, variances, quantiles, radius, pairwise distance, neighbours, modal value and diagonal

The functions ranges, variances radius, pairwise.dist, mode.val and diagonal all measure properties of the ordinated space based on its dimensional properties (they are also less affected by the “curse of dimensionality”):

ranges, variances quantiles and radius work on the same principle and measure the range/variance/radius of each dimension:

## [1] 2.430909 3.726481 2.908329 2.735739 1.588603
##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      2.736 1.673 2.431 2.908 3.645
##   subsets  n   obs
## 1       1 10 13.39
##   subsets  n   obs
## 1       1 10 114.5
## [1] 0.6093144 1.1438620 0.9131859 0.6537768 0.3549372
##   subsets  n obs.median 2.5%   25%   75% 97.5%
## 1       1 10      0.654 0.38 0.609 0.913 1.121
##   subsets  n   obs
## 1       1 10 3.675
##   subsets  n   obs
## 1       1 10 0.148
## [1] 2.234683 3.280911 2.760855 2.461077 1.559057
##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      2.461 1.627 2.235 2.761 3.229
##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      0.967 0.899 0.951 0.991 1.089
## [1] 1.4630780 2.4635449 1.8556785 1.4977898 0.8416318
## [1] 0.05144054 0.14099827 0.02212226 0.17453525 0.23044528
## [1] 0.6233501 0.7784888 0.7118713 0.6253263 0.5194332
##   subsets  n   obs
## 1       1 10 0.652

The pairwise distances and the neighbours distances uses the function vegan::vegdist and can take the normal vegdist options:

##   subsets  n   obs
## 1       1 10 2.539
##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      4.427 2.566 3.335 5.672  9.63
##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      1.517 1.266 1.432 1.646 2.787
##   subsets  n obs.median 2.5%   25%   75% 97.5%
## 1       1 10      7.895 6.15 6.852 9.402 10.99

Note that this function is a direct call to vegan::vegdist(matrix, method = method, diag = FALSE, upper = FALSE, ...).

The diagonal function measures the multidimensional diagonal of the whole space (i.e. in our case the longest Euclidean distance in our five dimensional space). The mode.val function measures the modal value of the matrix:

##   subsets  n   obs
## 1       1 10 3.659
##   subsets  n   obs
## 1       1 10 -2.21

This metric is only a Euclidean diagonal (mathematically valid) if the dimensions within the space are all orthogonal!

4.4.6.3 Centroids, displacements, and ancestral distance metrics

The centroids metric allows users to measure the position of the different elements compared to a fixed point in the ordinated space. By default, this function measures the distance between each element and their centroid (centre point):

##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      1.435 0.788 1.267 1.993 3.167
##   subsets  n   obs
## 1       1 10 1.435

It is however possible to fix the coordinates of the centroid to a specific point in the ordinated space, as long as it has the correct number of dimensions:

##   subsets  n obs.median  2.5% 25%   75% 97.5%
## 1       1 10      1.487 0.785 1.2 2.044 3.176
##   subsets  n obs.median  2.5%   25%   75% 97.5%
## 1       1 10      5.489 4.293 5.032 6.155 6.957

The displacements distance is the ratio between the centroids distance and the centroids distance with centroid = 0. Note that it is possible to measure a ratio from another point than 0 using the reference argument. It gives indication of the relative displacement of elements in the multidimensional space: a score >1 signifies a displacement away from the reference. A score of >1 signifies a displacement towards the reference.

##   subsets  n obs.median  2.5%   25% 75% 97.5%
## 1       1 10      1.014 0.841 0.925 1.1 1.205
##   subsets  n obs.median  2.5%  25%   75% 97.5%
## 1       1 10      3.368 2.066 3.19 4.358 7.166

The ancestral.dist metric works on a similar principle as the centroids function but changes the centroid to be the coordinates of each element’s ancestor. Therefore this functions needs a tree and node coordinates as additional arguments:

##   subsets n obs.median  2.5%   25%   75% 97.5%
## 1       1 9      3.862 0.346 1.809 6.239 7.445
##   subsets n   obs
## 1       1 9 35.91

4.4.6.4 Minimal spanning tree length

The span.tree.length uses the vegan::spantree function to heuristically calculate the minimum spanning tree (the shortest multidimensional tree connecting each elements) and calculates its length as the sum of every branch lengths.

##   subsets  n  obs
## 1       1 10 15.4

Note that because the solution is heuristic, this metric can take a long time to compute for big matrices.

4.4.6.5 Functional divergence and evenness

The func.div and func.eve functions are based on the FD::dpFD package. They are the equivalent to FD::dpFD(matrix)$FDiv and FD::dpFD(matrix)$FEve but a bit faster (since they don’t deal with abundance data). They are pretty straightforward to use:

##   subsets  n   obs
## 1       1 10 0.747
##   subsets  n   obs
## 1       1 10 0.898
##   subsets  n   obs
## 1       1 10 0.913

4.5 Summarising dispRity data (plots)

Because of its architecture, printing dispRity objects only summarises their content but does not print the disparity value measured or associated analysis (more about this here). To actually see what is in a dispRity object, one can either use the summary function for visualising the data in a table or plot to have a graphical representation of the results.

4.5.1 Summarising dispRity data

This function is an S3 function (summary.dispRity) allowing users to summarise the content of dispRity objects that contain disparity calculations.

These objects are easy to summarise as follows:

##   subsets  n   obs bs.median  2.5%   25%   75% 97.5%
## 1     120  5 2.823     2.403 1.481 2.144 2.594 2.823
## 2      80 19 3.233     3.064 2.863 2.983 3.160 3.234
## 3      40 15 3.359     3.163 2.719 3.032 3.259 3.402
## 4       0 10 4.055     3.684 3.029 3.581 3.783 3.946

Information about the number of elements in each subset and the observed (i.e. non-bootstrapped) disparity are also calculated. This is specifically handy when rarefying the data for example:

##   subsets  n   obs bs.median  2.5%   25%   75% 97.5%
## 1   crown 30 1.995     1.929 1.874 1.912 1.942 1.970
## 2   crown 29    NA     1.930 1.881 1.915 1.947 1.975
## 3   crown 28    NA     1.936 1.871 1.914 1.951 1.978
## 4   crown 27    NA     1.937 1.879 1.919 1.947 1.970
## 5   crown 26    NA     1.928 1.869 1.913 1.947 1.972
## 6   crown 25    NA     1.923 1.873 1.908 1.945 1.970

The summary functions can also take various options such as:

  • quantiles values for the confidence interval levels (by default, the 50 and 95 quantiles are calculated)
  • cent.tend for the central tendency to use for summarising the results (default is median)
  • digitsoption corresponding to the number of decimal places to print (default is2`)
  • recall option for printing the call of the dispRity object as well (default is FALSE)

These options can easily be changed from the defaults as follows:

##   subsets  n   obs bs.sd    6%   94%
## 1     120  5 2.823 0.338 1.767 2.823
## 2      80 19 3.233 0.119 2.908 3.216
## 3      40 15 3.359 0.178 2.841 3.369
## 4       0 10 4.055 0.226 3.265 3.900
##  ---- dispRity object ---- 
## 4 continuous (proximity) time subsets for 99 elements with 97 dimensions:
##     120, 80, 40, 0.
## Data was bootstrapped 100 times (method:"full").
## Disparity was calculated as: c(sum, variances).
##   subsets  n     obs bs.median    2.5%     25%     75%   97.5%
## 1     120  5 2.82292   2.40331 1.48096 2.14432 2.59439 2.82292
## 2      80 19 3.23312   3.06403 2.86262 2.98261 3.16024 3.23409
## 3      40 15 3.35947   3.16330 2.71937 3.03205 3.25855 3.40218
## 4       0 10 4.05457   3.68403 3.02885 3.58090 3.78324 3.94556

Note that the summary table is a data.frame, hence it is as easy to modify as any dataframe using dplyr. You can also export it in csv format using write.csv or write_csv or even directly export into LaTeX format using the following;

4.5.2 Plotting dispRity data

An alternative (and more fun!) way to display the calculated disparity is to plot the results using the S3 method plot.dispRity. This function takes the same options as summary.dispRity along with various graphical options described in the function help files (see ?plot.dispRity).

The plots can be of four different types:

  • continuous for displaying continuous disparity curves
  • box, lines, and polygons to display discrete disparity results in respectively a boxplot, confidence interval lines, and confidence interval polygons.

This argument can be left empty. In this case, the algorithm will automatically detect the type of subsets from the dispRity object and plot accordingly.

It is also possible to display the number of elements in each subset (as a horizontal dotted line) using the option elements = TRUE. Additionally, when the data is rarefied, one can indicate which level of rarefaction to display (i.e. only display the results for a certain number of elements) by using the rarefaction argument.

Since plot.dispRity uses the arguments from the generic plot method, it is of course possible to change pretty much everything using the regular plot arguments:

In addition to the classic plot arguments, the function can also take arguments that are specific to plot.dispRity like adding the number of elements or rarefaction level (as described above), and also changing the values of the quantiles to plot as well as the central tendency.

Note that the argument observed = TRUE allows to plot the disparity values calculated from the non-bootstrapped data as crosses on the plot.

For comparing results, it is also possible to add a plot to the existent plot by using add = TRUE:

Finally, if your data has been fully rarefied, it is also possible to easily look at rarefaction curves by using the rarefaction = TRUE argument:

Note that all the options above are plotting disparity objects for which a disparity metric has been calculated. This makes totally sense for dispRity objects but sometimes it might be interesting to look at what the trait-space looks like before measuring the disparity. This can be done by plotting dispRity objects with no calculated disparity!

For example, we might be interested in looking at how the distribution of elements change as a function of the distributions of different sub-settings. For example custom subsets vs. time subsets:

## [1] TRUE
## [1] TRUE

DISCLAIMER: This functionality can be handy for exploring the data (e.g. to visually check whether the subset attribution worked) but it might be misleading on how the data is actually distributed in the multidimensional space! Groups that don’t overlap on two set dimensions can totally overlap in all other dimensions!

4.6 Testing disparity hypotheses

The dispRity package allows users to apply statistical tests to the calculated disparity to test various hypotheses. The function test.dispRity works in a similar way to the dispRity function: it takes a dispRity object, a test and a comparisons argument.

The comparisons argument indicates the way the test should be applied to the data:

  • pairwise (default): to compare each subset in a pairwise manner
  • referential: to compare each subset to the first subset
  • sequential: to compare each subset to the following subset
  • all: to compare all the subsets together (like in analysis of variance)

It is also possible to input a list of pairs of numeric values or characters matching the subset names to create personalised tests. Some other tests implemented in dispRity such as the dispRity::null.test have a specific way they are applied to the data and therefore ignore the comparisons argument.

The test argument can be any statistical or non-statistical test to apply to the disparity object. It can be a common statistical test function (e.g. stats::t.test), a function implemented in dispRity (e.g. see ?null.test) or any function defined by the user.

This function also allows users to correct for Type I error inflation (false positives) when using multiple comparisons via the correction argument. This argument can be empty (no correction applied) or can contain one of the corrections from the stats::p.adjust function (see ?p.adjust).

Note that the test.dispRity algorithm deals with some classical test outputs (h.test, lm and numeric vector) and summarises the test output. It is, however, possible to get the full detailed output by using the options details = TRUE.

Here we are using the variables generated in the section above:

## [[1]]
##              statistic: t
## crown : stem     71.35019
## 
## [[2]]
##              parameter: df
## crown : stem      176.0866
## 
## [[3]]
##                    p.value
## crown : stem 7.057325e-132
## 
## [[4]]
##                   stderr
## crown : stem 0.004116325
## $`crown : stem`
## $`crown : stem`[[1]]
## 
##  Welch Two Sample t-test
## 
## data:  dots[[1L]][[1L]] and dots[[2L]][[1L]]
## t = 71.35, df = 176.09, p-value < 2.2e-16
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
##  0.2855769 0.3018243
## sample estimates:
## mean of x mean of y 
##  1.928095  1.634394
## [[1]]
##          statistic: W
## 120 : 80           44
## 80 : 40          3219
## 40 : 0            489
## 
## [[2]]
##               p.value
## 120 : 80 2.843186e-33
## 80 : 40  4.075680e-05
## 40 : 0   9.089860e-28
## Warning in test.dispRity(disparity_time_bins, test = bhatt.coeff): Multiple p-values will be calculated without adjustment!
## This can inflate Type I error!
##                    bhatt.coeff
## 120 - 80 : 80 - 40   0.0000000
## 120 - 80 : 40 - 0    0.0000000
## 80 - 40 : 40 - 0     0.5032302

Because of the modular design of the package, tests can always be made by the user (the same way disparity metrics can be user made). The only condition is that the test can be applied to at least two distributions. In practice, the test.dispRity function will pass the calculated disparity data (distributions) to the provided function in either pairs of distributions (if the comparisons argument is set to pairwise, referential or sequential) or a table containing all the distributions (comparisons = all; this should be in the same format as data passed to lm-type functions for example).

4.6.1 NPMANOVA in dispRity

One often useful test to apply to multidimensional data is the permutational multivariate analysis of variance based on distance matrices vegan::adonis. This can be done on dispRity objects using the adonis.dispRity wrapper function. Basically, this function takes the exact same arguments as adonis and a dispRity object for data and performs a PERMANOVA based on the distance matrix of the multidimensional space (unless the multidimensional space was already defined as a distance matrix). The adonis.dispRity function uses the information from the dispRity object to generate default formulas:

  • If the object contains customised subsets, it applies the default formula matrix ~ group testing the effect of group as a predictor on matrix (called from the dispRity object as data$matrix see dispRitu object details)
  • If the object contains time subsets, it applies the default formula matrix ~ time testing the effect of time as a predictor (were the different levels of time are the different time slices/bins)
## Warning: custom.subsets is applied on what seems to be a distance matrix.
## The resulting matrices won't be distance matrices anymore!
## 
## Call:
## vegan::adonis(formula = matrix ~ group, data = random_disparity,      method = "euclidean") 
## 
## Permutation: free
## Number of permutations: 999
## 
## Terms added sequentially (first to last)
## 
##           Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)
## group      1      14.2  14.200  1.2396 0.06443  0.166
## Residuals 18     206.2  11.456         0.93557       
## Total     19     220.4                 1.00000

Of course, it is possible to pass customised formulas if the disparity object contains more more groups. In that case the predictors must correspond to the names of the groups explained data must be set as matrix:

## Warning: custom.subsets is applied on what seems to be a distance matrix.
## The resulting matrices won't be distance matrices anymore!
## 
## Call:
## vegan::adonis(formula = matrix ~ g1 + g2, data = multi_groups,      method = "euclidean") 
## 
## Permutation: free
## Number of permutations: 999
## 
## Terms added sequentially (first to last)
## 
##           Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)
## g1         1      14.2  14.200 1.22042 0.06443  0.174
## g2         1       8.4   8.400 0.72194 0.03811  0.884
## Residuals 17     197.8  11.635         0.89746       
## Total     19     220.4                 1.00000

Finally, it is possible to use objects generated by chrono.subsets. In this case, adonis.dispRity will applied the matrix ~ time formula by default:

## Warning in adonis.dispRity(time_subsets): The input data for adonis.dispRity was not a distance matrix.
## The results are thus based on the distance matrix for the input data (i.e. dist(data$matrix)).
## Make sure that this is the desired methodological approach!
## 
## Call:
## vegan::adonis(formula = dist(matrix) ~ time, data = time_subsets,      method = "euclidean") 
## 
## Permutation: free
## Number of permutations: 999
## 
## Terms added sequentially (first to last)
## 
##           Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)    
## time       2     8.049  4.0245  2.1303 0.08311  0.001 ***
## Residuals 47    88.792  1.8892         0.91689           
## Total     49    96.841                 1.00000           
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Note that the function warns you that the input data was transformed into a distance matrix. This is reflected in the Call part of the output (formula = dist(matrix) ~ time).

To use each time subset as a separate predictor, you can use the matrix ~ chrono.subsets formula; this is equivalent to matrix ~ first_time_subset + second_time_subset + ...:

## Warning in adonis.dispRity(time_subsets, matrix ~ chrono.subsets): The input data for adonis.dispRity was not a distance matrix.
## The results are thus based on the distance matrix for the input data (i.e. dist(data$matrix)).
## Make sure that this is the desired methodological approach!
## 
## Call:
## vegan::adonis(formula = dist(matrix) ~ chrono.subsets, data = time_subsets,      method = "euclidean") 
## 
## Permutation: free
## Number of permutations: 999
## 
## Terms added sequentially (first to last)
## 
##           Df SumsOfSqs MeanSqs F.Model      R2 Pr(>F)    
## t100to85   1     3.090  3.0897  1.6354 0.03190  0.005 ** 
## t85to65    1     4.959  4.9593  2.6251 0.05121  0.001 ***
## Residuals 47    88.792  1.8892         0.91689           
## Total     49    96.841                 1.00000           
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

4.6.2 geiger::dtt model fitting in dispRity

The dtt function from the geiger package is also often used to compare a trait’s disparity observed in living taxa to the disparity of a simulated trait based on a given phylogeny. The dispRity package proposes a wrapper function for geiger::dtt, dtt.dispRity that allows the use of any disparity metric. Unfortunately, this implementation is slower that geiger::dtt (so if you’re using the metrics implemented in geiger prefer the original version) and, as the original function, is limited to ultrametric trees (only living taxa!)…

## Loading required package: geiger
## Warning in dtt.dispRity(data = geiger_data$dat, metric = c(sum,
## variances), : The following tip(s) was not present in the data: olivacea.

Note that, like in the original dtt function, it is possible to change the evolutionary model (see ?geiger::sim.char documentation).

4.6.3 null morphospace testing with null.test

This test is equivalent to the test performed in Díaz et al. (2016). It compares the disparity measured in the observed space to the disparity measured in a set of simulated spaces. These simulated spaces can be built with based on the hypothesis assumptions: for example, we can test whether our space is normal.

## Monte-Carlo test
## Call: [1] "dispRity::null.test"
## 
## Observation: 9.910536 
## 
## Based on 100 replicates
## Simulated p-value: 0.7524752 
## Alternative hypothesis: two-sided 
## 
##     Std.Obs Expectation    Variance 
## -0.24885893  9.96420000  0.04650127

Here the results show that disparity measured in our observed space is not significantly different than the one measured in a normal space. We can then propose that our observed space is normal!

These results have an attributed dispRity and randtest class and can be plotted as randtest objects using the dispRity S3 plot method:

For more details on generating spaces see the space.maker function tutorial.

4.7 Fitting modes of evolution to disparity data

The code used for these models is based on those developed by Gene Hunt (Hunt 2006, 2012; Hunt, Hopkins, and Lidgard 2015). So we acknowledge and thank Gene Hunt for developing these models and writing the original R code that served as inspiration for these models.

4.7.1 Simple modes of disparity change through time

4.7.1.1 model.test

Changes in disparity-through-time can follow a range of models, such as random walks, stasis, constrained evolution, trends, or an early burst model of evolution. We will start with by fitting the simplest modes of evolution to our data. For example we may have a null expectation of time-invariant change in disparity in which values fluctuate with a variance around the mean - this would be best describe by a Stasis model:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501

We can see the standard output from model.test. The first output message tells us it has tested for equal variances in each sample. The model uses Bartlett’s test of equal variances to assess if variances are equal, so if p > 0.05 then variance is treated as the same for all samples, but if (p < 0.05) then each bin variance is unique. Here we have p < 0.05, so variance is not pooled between samples.

By default model.test will use Bartlett’s test to assess for homogeneity of variances, and then use this to decide to pool variances or not. This is ignored if the argument pool.variance in model.test is changed from the default NULL to TRUE or FALSE. For example, to ignore Bartlett’s test and pool variances manually we would do the following:

## Running Stasis model...Done. Log-likelihood = -58.233

However, unless you have good reason to choose otherwise it is recommended to use the default of pool.variance = NULL:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501
## Disparity evolution model fitting:
## Call: model.test(data = BeckLee_disparity, model = "Stasis", pool.variance = NULL) 
## 
##            aicc delta_aicc weight_aicc
## Stasis 123.1027          0           1
## 
## Use x$full.details for displaying the models details
## or summary(x) for summarising them.

The remaining output gives us the log-likelihood of the Stasis model of -59.501 (you may notice this change when we pooled variances above). The output also gives us the small sample Akaike Information Criterion (AICc), the delta AICc (the distance from the best fitting model), and the AICc weights (~the relative support of this model compared to all models, scaled to one).

These are all metrics of relative fit, so when we test a single model they are not useful. By using the function summary in dispRity we can see the maximum likelihood estimates of the model parameters:

##         aicc delta_aicc weight_aicc log.lik param theta.1 omega
## Stasis 123.1          0           1   -59.5     2     3.4   0.1

So we again see the AICc, delta AICc, AICc weight, and the log-likelihood we saw previously. We now also see the number of parameters from the model (2: theta and omega), and their estimates so the variance (omega = 0.01) and the mean (theta.1 = 3.4).

The model.test function is designed to test relative model fit, so we need to test more than one model to make relative comparisons. So let’s compare to the fit of the Stasis model to another model with two parameters: the Brownian motion. Brownian motion assumes a constant mean that is equal to the ancestral estimate of the sequence, and the variance around this mean increases linearly with time. The easier way to compare these models is to simply add "BM" to the models vector argument:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501
## Running BM model...Done. Log-likelihood = 123.938
## Disparity evolution model fitting:
## Call: model.test(data = BeckLee_disparity, model = c("Stasis", "BM")) 
## 
##             aicc delta_aicc  weight_aicc
## Stasis  123.1027   366.8774 2.155677e-80
## BM     -243.7747     0.0000 1.000000e+00
## 
## Use x$full.details for displaying the models details
## or summary(x) for summarising them.

Et voilà! Here we can see by the log-likelihood, AICc, delta AICc, and AICc weight Brownian motion has a much better relative fit to these data than the Stasis model. Brownian motion has a relative AICc fit 366 units better than Stasis, and virtually has a AICc weight of 1.

We can also all the information about the relative fit of models alongside the maximum likelihood estimates of model parameters using the summary function

##        aicc delta_aicc weight_aicc log.lik param theta.1 omega
## Stasis  123      366.9           0   -59.5     2   3.403  0.15
## BM     -244        0.0           1   123.9     2      NA    NA
##        ancestral state sigma squared
## Stasis              NA            NA
## BM               2.858         0.003

Not that because the parameters per models differ, the summary includes NA for inapplicable parameters per models (e.g. the theta and omega parameters from the Stasis models are inapplicable for a Brownian motion model).

We can plot the relative fit of our models using the plot function

relative fit (AICc weight) of Stasis and Brownian models of disparity through time

Figure 4.1: relative fit (AICc weight) of Stasis and Brownian models of disparity through time

Here we see and overwhelming support for the Brownian motion model.

Alternatively, we could test all available models single modes: Stasis, Brownian motion, Ornstein-Uhlenbeck (evolution constrained to an optima), Trend (increasing or decreasing mean through time), and Early Burst (exponentially decreasing rate through time)

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501
## Running BM model...Done. Log-likelihood = 123.938
## Running OU model...Done. Log-likelihood = 126.431
## Running Trend model...Done. Log-likelihood = 126.361
## Running EB model...Done. Log-likelihood = 113.081
##        aicc delta_aicc weight_aicc log.lik param theta.1 omega
## Stasis  123      369.6       0.000   -59.5     2   3.403  0.15
## BM     -244        2.7       0.157   123.9     2      NA    NA
## OU     -245        2.0       0.227   126.4     4      NA    NA
## Trend  -247        0.0       0.617   126.4     3      NA    NA
## EB     -220       26.6       0.000   113.1     3      NA    NA
##        ancestral state sigma squared alpha optima.1 trend     eb
## Stasis              NA            NA    NA       NA    NA     NA
## BM               2.858         0.003    NA       NA    NA     NA
## OU               2.835         0.002 0.004    5.707    NA     NA
## Trend            2.839         0.002    NA       NA  0.01     NA
## EB               4.055         0.002    NA       NA    NA -0.014

These models indicate support for a Trend model, and we can plot the relative support of all model AICc weights.

relative fit (AICc weight) of various modes of evolution

Figure 4.2: relative fit (AICc weight) of various modes of evolution

Note that although AIC values are indicator of model best fit, it is also important to look at the parameters themselves. For example OU can be really well supported but with an alpha parameter really close to 0, making it effectively a BM model (Cooper et al. 2016).

Is this a trend of increasing or decreasing disparity through time? One way to find out is to look at the summary function for the Trend model:

##            aicc      delta_aicc     weight_aicc         log.lik 
##        -247.000           0.000           0.617         126.400 
##           param         theta.1           omega ancestral state 
##           3.000              NA              NA           2.839 
##   sigma squared           alpha        optima.1           trend 
##           0.002              NA              NA           0.010 
##              eb 
##              NA

This show a positive trend (0.01) of increasing disparity through time.

4.7.2 Plot and run simulation tests in a single step

4.7.2.1 model.test.wrapper

Patterns of evolution can be fit using model.test, but the model.test.wrapper fits the same models as model.test as well as running predictive tests and plots.

The predictive tests use the maximum likelihood estimates of model parameters to simulate a number of datasets (default = 1000), and analyse whether this is significantly different to the empirical input data using the Rank Envelope test (Murrell 2018). Finally we can plot the empirical data, simulated data, and the Rank Envelope test p values. This can all be done using the function model.test.wrapper, and we will set the argument show.p = TRUE so p values from the Rank Envelope test are printed on the plot:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501
## Running BM model...Done. Log-likelihood = 123.938
## Running OU model...Done. Log-likelihood = 126.431
## Running Trend model...Done. Log-likelihood = 126.361
## Running EB model...Done. Log-likelihood = 113.081
Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for Trend, OU, BM, EB, and Stasis models

Figure 4.3: Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for Trend, OU, BM, EB, and Stasis models 

##        aicc delta_aicc weight_aicc log.lik param theta.1 omega
## Trend  -247        0.0       0.617   126.4     3      NA    NA
## OU     -245        2.0       0.227   126.4     4      NA    NA
## BM     -244        2.7       0.157   123.9     2      NA    NA
## EB     -220       26.6       0.000   113.1     3      NA    NA
## Stasis  123      369.6       0.000   -59.5     2   3.403  0.15
##        ancestral state sigma squared alpha optima.1 trend     eb
## Trend            2.839         0.002    NA       NA  0.01     NA
## OU               2.835         0.002 0.004    5.707    NA     NA
## BM               2.858         0.003    NA       NA    NA     NA
## EB               4.055         0.002    NA       NA    NA -0.014
## Stasis              NA            NA    NA       NA    NA     NA
##        median p value lower p value upper p value
## Trend      0.97902098   0.978021978     0.9800200
## OU         0.93406593   0.934065934     0.9340659
## BM         0.33266733   0.322677323     0.3426573
## EB         0.06293706   0.000999001     0.1248751
## Stasis     1.00000000   1.000000000     1.0000000

From this plot we can see the empirical estimates of disparity through time (pink) compared to the predictive data based upon the simulations using the estimated parameters from each model. There is no significant differences between the empirical data and simulated data, except for the Early Burst model.

Trend is the best-fitting model but the plot suggests the OU model also follows a trend-like pattern. This is because the optima for the OU model (5.7) is different to the ancestral state 2.835 and outside the observed value. This is potentially unrealistic, and one way to alleviate this issue is to set the optima of the OU model to equal the ancestral estimate - this is the normal practice for OU models in comparative phylogenetics. To set the optima to the ancestral value we change the argument fixed.optima = TRUE:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501
## Running BM model...Done. Log-likelihood = 123.938
## Running OU model...Done. Log-likelihood = 123.938
## Running Trend model...Done. Log-likelihood = 126.361
## Running EB model...Done. Log-likelihood = 113.081
Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for Trend, OU, BM, EB, and Stasis models with the optima of the OU model set to equal the ancestral value

Figure 4.4: Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for Trend, OU, BM, EB, and Stasis models with the optima of the OU model set to equal the ancestral value 

##        aicc delta_aicc weight_aicc log.lik param theta.1 omega
## Trend  -247        0.0       0.745   126.4     3      NA    NA
## BM     -244        2.7       0.189   123.9     2      NA    NA
## OU     -242        4.8       0.066   123.9     3      NA    NA
## EB     -220       26.6       0.000   113.1     3      NA    NA
## Stasis  123      369.6       0.000   -59.5     2   3.403  0.15
##        ancestral state sigma squared alpha trend     eb median p value
## Trend            2.839         0.002    NA  0.01     NA     0.97402597
## BM               2.858         0.003    NA    NA     NA     0.29870130
## OU               2.858         0.003     0    NA     NA     0.35864136
## EB               4.055         0.002    NA    NA -0.014     0.06393606
## Stasis              NA            NA    NA    NA     NA     1.00000000
##        lower p value upper p value
## Trend    0.973026973     0.9750250
## BM       0.285714286     0.3116883
## OU       0.347652348     0.3696304
## EB       0.000999001     0.1268731
## Stasis   1.000000000     1.0000000

The relative fit of the OU model is decreased by constraining the fit of the optima to equal the ancestral state value. In fact as the OU attraction parameter (alpha) is zero, the model is equal to a Brownian motion model but is penalised by having an extra parameter. Note that indeed, the plots of the BM model and the OU model look nearly identical.

4.7.3 Multiple modes of evolution (time shifts)

As well as fitting a single model to a sequence of disparity values we can also allow for the mode of evolution to shift at a single or multiple points in time. The timing of a shift in mode can be based on an a prior expectation, such as a mass extinction event, or the model can test multiple points to allow to find time shift point with the highest likelihood.

Models can be fit using model.test but it can be more convenient to use model.test.wrapper. Here we will compare the relative fit of Brownian motion, Trend, Ornstein-Uhlenbeck and a multi-mode Ornstein Uhlenbck model in which the optima changes at 66 million years ago, the Cretaceous-Palaeogene boundary.

For example, we could be testing the hypothesis that the extinction of non-avian dinosaurs allowed mammals to go from scurrying in the undergrowth (low optima/low disparity) to dominating all habitats (high optima/high disparity). We will constrain the optima of OU model in the first time begin (i.e, pre-66 Mya) to equal the ancestral value:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running BM model...Done. Log-likelihood = 123.938
## Running Trend model...Done. Log-likelihood = 126.361
## Running OU model...Done. Log-likelihood = 123.938
## Running multi.OU model...Done. Log-likelihood = 126.469
Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for BM, Trend, OU, and multi OU models with a shift in optima allowed at 66 Ma

Figure 4.5: Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for BM, Trend, OU, and multi OU models with a shift in optima allowed at 66 Ma 

##          aicc delta_aicc weight_aicc log.lik param ancestral state
## Trend    -247      0.000       0.580   126.4     3           2.839
## multi.OU -245      1.922       0.222   126.5     4           2.836
## BM       -244      2.742       0.147   123.9     2           2.858
## OU       -242      4.845       0.051   123.9     3           2.858
##          sigma squared trend alpha optima.2 median p value lower p value
## Trend            0.002  0.01    NA       NA      0.9920080     0.9920080
## multi.OU         0.002    NA 0.005    5.526      0.6308691     0.6263736
## BM               0.003    NA    NA       NA      0.3506494     0.3396603
## OU               0.003    NA 0.000       NA      0.3666334     0.3576424
##          upper p value
## Trend        0.9920080
## multi.OU     0.6353646
## BM           0.3616384
## OU           0.3756244

The multi-OU model shows an increase an optima at the Cretaceous-Palaeogene boundary, indicating a shift in disparity. However, this model does not fit as well as a model in which there is an increasing trend through time. We can also fit a model in which the we specify a heterogeneous model but we do not give a time.split. In this instance the model will test all splits that have at least 10 time slices on either side of the split. That’s 102 potential time shifts in this example dataset so be warned, the following code will estimate 105 models!

As well as specifying a multi-OU model we can run any combination of models. For example we could fit a model at the Cretaceous-Palaeogene boundary that goes from an OU to a BM model, a Trend to an OU model, a Stasis to a Trend model or any combination you want to use. The only model that can’t be used in combination is a multi-OU model.

These can be introduced by changing the input for the models into a list, and supplying a vector with the two models. This is easier to see with an example:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running BM:OU model...Done. Log-likelihood = 115.367
## Running Stasis:OU model...Done. Log-likelihood = 80.16
## Running BM:Stasis model...Done. Log-likelihood = 34.278
## Running OU:Trend model...Done. Log-likelihood = 118.166
## Running Stasis:BM model...Done. Log-likelihood = 80.16
Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for a variety of models with a shift in optima allowed at 66 Ma

Figure 4.6: Empirical disparity through time (pink), simulate data based on estimated model parameters (grey), delta AICc, and range of p values from the Rank Envelope test for a variety of models with a shift in optima allowed at 66 Ma 

##           aicc delta_aicc weight_aicc log.lik param ancestral state
## OU:Trend  -228        0.0       0.943   118.2     4           3.051
## BM:OU     -222        5.6       0.057   115.4     4           3.052
## Stasis:BM -154       73.9       0.000    80.2     3              NA
## Stasis:OU -150       78.2       0.000    80.2     5              NA
## BM:Stasis  -60      167.8       0.000    34.3     4           2.858
##           sigma squared alpha optima.1 theta.1 omega trend median p value
## OU:Trend          0.003 0.039       NA      NA    NA 0.015      0.4840160
## BM:OU             0.003 0.000    4.055      NA    NA    NA      0.7857143
## Stasis:BM         0.004    NA       NA   3.109 0.025    NA      0.9860140
## Stasis:OU         0.004 0.000    4.055   3.109 0.025    NA      0.9635365
## BM:Stasis         0.002    NA       NA   3.648 0.113    NA      0.9615385
##           lower p value upper p value
## OU:Trend      0.4685315     0.4995005
## BM:OU         0.7792208     0.7922078
## Stasis:BM     0.9850150     0.9870130
## Stasis:OU     0.9610390     0.9660340
## BM:Stasis     0.9560440     0.9670330

4.7.4 model.test.sim

Note that all the models above where run using the model.test.wrapper function that is a… wrapping function! In practice, this function runs two main functions from the dispRity package and then plots the results:

  • model.test and
  • model.test.sim

The model.test.sim allows to simulate disparity evolution given a dispRity object input (as in model.test.wrapper) or given a model and its specification. For example, it is possible to simulate a simple Brownian motion model (or any of the other models or models combination described above):

## Disparity evolution model simulation:
## Call: model.test.sim(sim = 1000, model = "BM", time.span = 50, variance = 0.1, sample.size = 100, parameters = list(ancestral.state = 0)) 
## 
## Model simulated (1000 times):
## [1] "BM"

This will simulate 1000 Brownian motions for 50 units of time with 100 sampled elements, a variance of 0.1 and an ancestral state of 0. We can also pass multiple models in the same way we did it for model.test This model can then be summarised and plotted as most dispRity objects:

##   subsets   n var      median      2.5%       25%       75%    97.5%
## 1      50 100 0.1 -0.09312806 -1.718481 -0.712478 0.5854576 1.772700
## 2      49 100 0.1 -0.01881396 -2.697017 -0.967495 0.9105681 2.712452
## 3      48 100 0.1 -0.06914146 -3.443629 -1.205211 1.1171606 3.185465
## 4      47 100 0.1 -0.01584249 -3.897544 -1.474240 1.3762412 3.639836
## 5      46 100 0.1 -0.04118287 -4.402526 -1.496096 1.5347916 4.239536
## 6      45 100 0.1 -0.17175251 -4.764863 -1.697076 1.5894314 4.349500
A simulated Brownian motion

Figure 4.7: A simulated Brownian motion

Note that these functions can take all the arguments that can be passed to plot, summary, plot.dispRity and summary.dispRity.

4.7.4.1 Simulating tested models

Maybe more interestingly though, it is possible to pass the output of model.test directly to model.test.sim to simulate the models that fits the data the best and calculate the Rank Envelope test p value. Let’s see that using the simple example from the start:

## Evidence of equal variance (Bartlett's test of equal variances p = 0).
## Variance is not pooled.
## Running Stasis model...Done. Log-likelihood = -59.501
## Running BM model...Done. Log-likelihood = 123.938
## Running OU model...Done. Log-likelihood = 126.431
## Running Trend model...Done. Log-likelihood = 126.361
## Running EB model...Done. Log-likelihood = 113.081
##        aicc delta_aicc weight_aicc log.lik param theta.1 omega
## Stasis  123      369.6       0.000   -59.5     2   3.403  0.15
## BM     -244        2.7       0.157   123.9     2      NA    NA
## OU     -245        2.0       0.227   126.4     4      NA    NA
## Trend  -247        0.0       0.617   126.4     3      NA    NA
## EB     -220       26.6       0.000   113.1     3      NA    NA
##        ancestral state sigma squared alpha optima.1 trend     eb
## Stasis              NA            NA    NA       NA    NA     NA
## BM               2.858         0.003    NA       NA    NA     NA
## OU               2.835         0.002 0.004    5.707    NA     NA
## Trend            2.839         0.002    NA       NA  0.01     NA
## EB               4.055         0.002    NA       NA    NA -0.014

As seen before, the Trend model fitted this dataset the best. To simulate what 1000 Trend models would look like using the same parameters as the ones estimated with model.test (here the ancestral state being 2.839, the sigma squared beeing 0.002 and the trend of 0.01), we can simply pass this model to model.test.sim:

## Disparity evolution model simulation:
## Call: model.test.sim(sim = 1000, model = disp_time) 
## 
## Model simulated (1000 times):
##       aicc log.lik param ancestral state sigma squared trend
## Trend -247   126.4     3           2.839         0.002  0.01
## 
## Rank envelope test
##  p-value of the test: 0.987013 (ties method: midrank)
##  p-interval         : (0.987013, 0.987013)

By default, the model simulated is the one with the lowest AICc (model.rank = 1) but it is possible to choose any ranked model, for example, the OU (second one):

## Disparity evolution model simulation:
## Call: model.test.sim(sim = 1000, model = disp_time, model.rank = 2) 
## 
## Model simulated (1000 times):
##    aicc log.lik param ancestral state sigma squared alpha optima.1
## OU -245   126.4     4           2.835         0.002 0.004    5.707
## 
## Rank envelope test
##  p-value of the test: 0.8971029 (ties method: midrank)
##  p-interval         : (0.8941059, 0.9000999)

And as the example above, the simulated data can be plotted or summarised:

##   subsets n        var   median     2.5%      25%      75%    97.5%
## 1     120 5 0.06056490 2.837869 2.614855 2.760862 2.908493 3.059927
## 2     119 5 0.07453663 2.858233 2.598562 2.756114 2.945448 3.106885
## 3     118 6 0.07556947 2.856971 2.589279 2.769942 2.941288 3.129721
## 4     117 6 0.07556947 2.870782 2.584536 2.777017 2.965582 3.147397
## 5     116 6 0.07556947 2.869144 2.589171 2.769528 2.978955 3.173132
## 6     115 7 0.06590243 2.883672 2.593479 2.790683 2.978769 3.179178
##   subsets n        var   median     2.5%      25%      75%    97.5%
## 1     120 5 0.06056490 2.832068 2.618984 2.759091 2.908017 3.051097
## 2     119 5 0.07453663 2.849006 2.592719 2.759753 2.940149 3.115460
## 3     118 6 0.07556947 2.858304 2.579423 2.767239 2.946567 3.125425
## 4     117 6 0.07556947 2.872844 2.585720 2.782702 2.971421 3.142925
## 5     116 6 0.07556947 2.881609 2.599897 2.779863 2.985630 3.168541
## 6     115 7 0.06590243 2.897410 2.602577 2.790346 2.998301 3.182402
The best fitted model (Trend) and the observed disparity through time

Figure 4.8: The best fitted model (Trend) and the observed disparity through time

4.8 Disparity as a distribution

Disparity is often regarded as a summary value of the position of the all elements in the ordinated space. For example, the sum of variances, the product of ranges or the median distance between the elements and their centroid will summarise disparity as a single value. This value can be pseudo-replicated (bootstrapped) to obtain a distribution of the summary metric with estimated error. However, another way to perform disparity analysis is to use the whole distribution rather than just a summary metric (e.g. the variances or the ranges).

This is possible in the dispRity package by calculating disparity as a dimension-level 2 metric only! Let’s have a look using our previous example of bootstrapped time slices but by measuring the distances between each taxon and their centroid as disparity.

The resulting disparity object is of dimension-level 2, so it can easily be transformed into a dimension-level 1 object by, for example, measuring the median distance of all these distributions:

And we can now compare the differences between these methods:

##   subsets  n obs.median bs.median  2.5%   25%   75% 97.5%
## 1     120  5      1.508     1.360 0.812 1.144 1.553 1.839
## 2      80 19      1.790     1.699 1.405 1.591 1.812 1.945
## 3      40 15      1.689     1.681 1.336 1.563 1.843 2.064
## 4       0 10      1.910     1.803 1.372 1.699 1.956 2.098
##   subsets  n   obs bs.median  2.5%   25%   75% 97.5%
## 1     120  5 1.508     1.421 0.789 1.219 1.479 1.615
## 2      80 19 1.790     1.699 1.570 1.647 1.731 1.781
## 3      40 15 1.689     1.680 1.520 1.621 1.721 1.791
## 4       0 10 1.910     1.807 1.550 1.772 1.852 1.945

We can see that the summary message for the distribution is slightly different than before. Here summary also displays the observed central tendency (i.e. the central tendency of the measured distributions). Note that, as expected, this central tendency is the same in both metrics!

Another, maybe more intuitive way, to compare both approaches for measuring disparity is to plot the distributions:

We can then test for differences in the resulting distributions using test.dispRity and the bhatt.coeff test as described above.

## Warning in test.dispRity(disparity_centroids_median, test = bhatt.coeff): Multiple p-values will be calculated without adjustment!
## This can inflate Type I error!
##          bhatt.coeff
## 120 : 80   0.1121007
## 120 : 40   0.2407057
## 120 : 0    0.1507107
## 80 : 40    0.9308723
## 80 : 0     0.5340463
## 40 : 0     0.5241821

In this case, we are looking at the probability of overlap of the distribution of median distances from centroids among each pair of time slices. In other words, we are measuring whether the medians from each bootstrap pseudo-replicate for each time slice overlap. But of course, we might be interested in the actual distribution of the distances from the centroid rather than simply their central tendencies. This can be problematic depending on the research question asked since we are effectively comparing non-independent medians distributions (because of the pseudo-replication).

One solution, therefore, is to look at the full distribution:

## Warning in test.dispRity(disparity_centroids, test = bhatt.coeff): Multiple p-values will be calculated without adjustment!
## This can inflate Type I error!
##          bhatt.coeff
## 120 : 80   0.6296718
## 120 : 40   0.6853292
## 120 : 0    0.5733186
## 80 : 40    0.9400499
## 80 : 0     0.8685040
## 40 : 0     0.9277124

These results show the actual overlap among all the measured distances from centroids concatenated across all the bootstraps. For example, when comparing the slices 120 and 80, we are effectively comparing the 5 \(\times\) 100 distances (the distances of the five elements in slice 120 bootstrapped 100 times) to the 19 \(\times\) 100 distances from slice 80. However, this can also be problematic for some specific tests since the n \(\times\) 100 distances are also pseudo-replicates and thus are still not independent.

A second solution is to compare the distributions to each other for each replicate:

## Warning in test.dispRity(disparity_centroids, test = bhatt.coeff, concatenate = FALSE): Multiple p-values will be calculated without adjustment!
## This can inflate Type I error!
##          bhatt.coeff       2.5%       25%       75%     97.5%
## 120 : 80   0.2875529 0.00000000 0.1777047 0.4006176 0.5681645
## 120 : 40   0.3134755 0.00000000 0.2232051 0.4309401 0.6264552
## 120 : 0    0.2198913 0.00000000 0.0000000 0.3464102 0.6319158
## 80 : 40    0.6224908 0.27975876 0.5215252 0.7500784 0.9247555
## 80 : 0     0.4593145 0.07254763 0.3317231 0.5685736 0.7538037
## 40 : 0     0.5118333 0.03878359 0.3999051 0.6473892 0.8229464

These results show the median overlap among pairs of distributions in the first column (bhatt.coeff) and then the distribution of these overlaps among each pair of bootstraps. In other words, when two distributions are compared, they are now compared for each bootstrap pseudo-replicate, thus effectively creating a distribution of probabilities of overlap. For example, when comparing the slices 120 and 80, we have a mean probability of overlap of 0.28 and a probability between 0.18 and 0.43 in 50% of the pseudo-replicates. Note that the quantiles and central tendencies can be modified via the conc.quantiles option.

4.9 Disparity from other matrices

In the example so far, disparity was measured from an ordinated multidimensional space (i.e. a PCO of the distances between taxa based on discrete morphological characters). This is a common approach in palaeobiology, morphometrics or ecology but ordinated matrices are not mandatory for the dispRity package! It is totally possible to perform the same analysis detailed above using other types of matrices as long as your elements are rows in your matrix.

For example, we can use the data set eurodist, an R inbuilt dataset that contains the distances (in km) between European cities. We can check for example, if Northern European cities are closer to each other than Southern ones:

##           Athens Barcelona Brussels Calais Cherbourg
## Athens         0      3313     2963   3175      3339
## Barcelona   3313         0     1318   1326      1294
## Brussels    2963      1318        0    204       583
## Calais      3175      1326      204      0       460
## Cherbourg   3339      1294      583    460         0
## Warning: custom.subsets is applied on what seems to be a distance matrix.
## The resulting matrices won't be distance matrices anymore!

We can compare this approach to an ordination one:

## Warning in cmdscale(eurodist, k = nrow(eurodist) - 2): only 11 of the first
## 19 eigenvalues are > 0

And visualise the differences:

As expected, the results are pretty similar in pattern but different in terms of scale. The median centroids distance is expressed in km in the “Distance differences” plots and in Euclidean units of variation in the “Ordinated differences” plots.

References

Beck, Robin M, and Michael S Lee. 2014. “Ancient Dates or Accelerated Rates? Morphological Clocks and the Antiquity of Placental Mammals.” Proceedings of the Royal Society B: Biological Sciences 281 (20141278): 1–10. https://doi.org/10.1098/rspb.2014.1278.

Cooper, Natalie, Gavin H. Thomas, Chris Venditti, Andrew Meade, and Rob P. Freckleton. 2016. “A Cautionary Note on the Use of Ornstein Uhlenbeck Models in Macroevolutionary Studies.” Biological Journal of the Linnean Society 118 (1): 64–77. https://doi.org/10.1111/bij.12701.

Díaz, Sandra, Jens Kattge, Johannes HC Cornelissen, Ian J Wright, Sandra Lavorel, Stéphane Dray, Björn Reu, et al. 2016. “The Global Spectrum of Plant Form and Function.” Nature 529 (7585). Nature Publishing Group: 167. http://dx.doi.org/10.1038/nature16489.

Guillerme, T., and N. Cooper. 2018. “Time for a Rethink: Time Sub-Sampling Methods in Disparity-Through-Time Analyses.” Palaeontology 61 (4): 481–93. https://doi.org/10.1111/pala.12364.

Hunt, Gene. 2006. “Fitting and Comparing Models of Phyletic Evolution: Random Walks and Beyond.” Paleobiology 32 (4). Cambridge University Press: 578–601. https://doi.org/10.1666/05070.1.

Hunt, Gene. 2012. “Measuring Rates of Phenotypic Evolution and the Inseparability of Tempo and Mode.” Paleobiology 38 (3). GeoScienceWorld: 351–73. https://doi.org/10.1666/11047.1.

Hunt, Gene, Melanie J Hopkins, and Scott Lidgard. 2015. “Simple Versus Complex Models of Trait Evolution and Stasis as a Response to Environmental Change.” Proceedings of the National Academy of Sciences. National Acad Sciences, 201403662. https://doi.org/10.1073/pnas.1403662111.

Murrell, David J. 2018. “A Global Envelope Test to Detect Non-Random Bursts of Trait Evolution.” Methods in Ecology and Evolution 9 (7). Wiley Online Library: 1739–48. https://doi.org/10.1111/2041-210X.13006.