Citation info now points to the bioRxiv paper.
All modeling functions now completely rely on the
phylolm package, and no longer use
ape. This is a major change, that will possibly change the outcomes of some of your existing analyses (as can happen when chaning the modeling package). There are, however, several good reasons to make this change, which I think make it worth the trouble. Firstly, the package is much faster for large trees, and this effect is compounded in
phylopath because one may have to fit a few dozen models. Secondly, I think it is important to have confidence intervals around the regression coefficients, and those were not available for
phylolm makes it easy to use a larger variety of models of evolution, including two versions of OU and early burst, which can be simply set using the
model parameter. Lastly, the
phyloglm() functions give more uniform results, which makes it easier to code for situation where you may use both.
phylo_path and all related methods now deal automatically with both continuous and binary data. All separate binary functions and methods have disappeared as they are no longer needed. Mixing of binary and continious data in the same models is now allowed.
The variable order in d-seperation statements now better follows the causal flow of the DAG.
plot() method for
phylopath.summary objects, that shows the weights and p-values for the different models.
to arguments. The first allows the user to choose between confidence invervals and standard errors, the second to order the paths by several methods, and the last two can be used to select only certain paths.
Plotting methods of causal models now support a manual layout.
Plotting of fitted DAG’s now uses edge width instead of color to indicate, the standardized regression coefficient strength, but this can be reverted using the
define_model_set() convenience function for building models, that avoids repeated calls to
DAG() and has an argument to supply paths that are shared between all your models. It is not needed to specify isolate variables. Old code using
DAG() continues to work as normal.
Added support for additional arguments passed to
phylo_path. This can be helpful, for example, for setting the fitting method to maximum likelihood (
method = "ML").
Added support for completely binary models, that are fitted with
phylo_path_binary() to compare models.
choice() are now S3 generics and will handle both continuous and binary versions. Usage is designed to be as close to the continuous version as possible.
est_DAG_binary() powers the binary S3 methods.
All plot functions that used
DiagrammeR now use
ggraph instead. This gives much more control over the positioning of the nodes, and allows to plot multiple models at once. Exporting plots also becomes much easier.
You can now plot a list of causal models with
plot_model_set(). This creates a faceted plot where all nodes are kept in the same location, which makes it easier to spot how models are different.
If there are any
NA values in
data for the variables in
models, these rows are now dropped from
data with a message. Use
na.rm = FALSE to revert to the old behavior.
When PGLS models fail, an informative error is now returned to the user.
phylo_path() now checks for row.names that line up with the tree tip labels. If the tree contains surplus species, it gets pruned to size with a message. This includes cases where species are dropped due to missing values.
citation() now correctly refers to the methods paper by Von Hardenberg & Gonzalez-Voyer first and the package second.
Fewer models are now fitted when using
phylo_path(), since any duplicated independence statements are now only fitted once. This leads to a significant reduction in running time in many cases, especially when many models are considered.
Implemented support for parallel processing in
phylo_path() using the
phylo_path() now shows a progress bar.
New function added (
choice()) that is a very simple wrapper around
est_DAG(). It adds to
average() by allowing for choosing any model as the final model, and encourages users to not always pick the lowest CICc model.
Prepared plotting functions for new release of
DiagrammeR, v0.9 now required.
IMPORTANT: Faulty model averaging has been fixed. This was often introduced due to differences in matrix ordering. Averaging results from versions before 0.2.1 should NOT be trusted.
ape::corBrownian() no longer returns an error.
Averaging is less likely to fail due to errors in
phylo_path() has become more streamlined with functionality moved to other functions. The
phylopath object now contains all necessary models and data,
summary() is used to obtain the results table, and
average() are used to extract and fit the best or average model. See the vignette for details.
Model averaging for arbitrary models is now possible with
Model averaging now supports both conditional and full model averaging.
Model averaging now returns standard errors and confidence intervals based on the
MuMIn package (issue #1).
A new function
plot_coefs for plotting regression coefficients and their confidence intervals has been added.