Hmsc is no longer dependent on packages mvtnorm and pdist.
Hmsc is now dependent on the sp package.
Vignettes can be re-built from their sources out of the box. Previously they needed editing by hand to reproduce the pdf version.
It is now possible to use Spatial data in random models. Handling of Spatial data is based on the sp package and follows its conventions. The locations of sampling units can be given as a decimal longitude-latitude matrix, and the Hmsc functions will use great circle distances in spatial models. Projected spatial coordinates will be handled as such and Euclidean distances will be used internally.
User-specified spatial distances can be more widely used in spatial random models. However, some models are more flexible with spatial coordinates. Most importantly, Gaussian Predictive Process (GPP) needs spatial coordinate data.
Species data Y is normally a numeric matrix, but now it is allowed to use numeric data frames, or in univariate models, a numeric vector.
A tibble can be used for measured covariates for fixed effects XData in addition to a data frame (the wish of Github issue #37).
The names of distributions can be abbreviated in Hmsc definition as long as the names are unique.
computeWAIC is more robust against results of poorly fitting models, and it is now possible to evaluate WAIC separately for each species. See GitHub issue #44.
constructGradient argument nonFocalVariables accepts now a single number 1 or 2 as a shortcut of default type for all non-focal variables instead of requesting a list of types of all variables.
plotGradient gained new argument yshow which is a single number or vector of numeric values that must be included on the y-axis. In general, the y-axis is scaled to show the plotted values, but yshow = 0 will always show zero, even when this is not among plotted values, and yshow = c(0,1) will show both zero and one.
plotVariancePartition defaults to plot the original terms instead of single contrast. For instance, only one component is shown for multilevel factors instead of showing each level separately. User can still specify how the components are displayed.
plot functions plotBeta, plotGamma and plotVariancePartitioning allow setting or modifying the plot main title. plotGradient already allowed this.
Random seed is now saved in sampleMcmc models. This allows replication of same random number sequences. However, there is no guarantee of replication across Hmsc release versions or computing platforms.
HmscRandomLevel saves the function call. The call can be inspected with getCall() and the model can be modified with update().
constructGradient could sometimes shuffle spatial locations leading into wrong predictions with spatial models.
plotGradient(..., showData = TRUE) ignored data values in setting plot minimum. See GitHub issue #48. The data values were not always shown with measure = "S" in quantitative linear models.
R release 4.0 will drop the convention to automatically change character variables to factors, and this causes errors in internal working of several Hmsc functions. This version of Hmsc is released principally to accomodate these changes in R. Hmsc will also work in previous versions of R.
Hmsc 3.0-5 was never released to CRAN. It is a snapshot that corresponds to the on-line publication of Tikhonov et al. (2020) Joint species distribution modelling with the R-package Hmsc. Methods in Ecology and Evolution 11, 442–447. (https://doi.org/10.1111/2041-210X.13345).
Shape and rate parameters (aSigma, bSigma) for the prior Gamma distribution for the variance parameter (sigma) changed. The change will influence models with "normal" and "lognormal poisson" distributions. In particular, "lognormal poisson" will more easily tend toward zero sigma if there is no overdispersion to "poisson". However, in such cases it may be wiser to refit models with pure "poisson" distribution. You can changes these parameters with setPriors function.
Cross-validation works also when the test data set has some spatial units that were unseen in the training data.
When calling sampleMcmc with fromPrior = TRUE, the residual variance parameter sigma used Gamma rather than inverse of Gamma distribution. The same error was present when sampling the initial values for the MCMC algorithm. However, the actual MCMC algorithm (and thus the posterior distribution) was correct.
Predictions with spatial NNGP models failed if there was only one unit. Github issue #40.
Reduced-Rank Regression also works for single-species models, and more robust scaling is used for species-specific covariate matrices.
Spatial models with Gaussian Predictive Process now also works when the number of spatial locations is less than the number of sampling units.
Predictions with spatial NNGP and GPP models gave bad estimates.
Several functions failed in the development version of R (to be released as R version 4). The failures were caused by changes in R internals.
Fixed bug with delta for alignPosterior which influences sampleMcmc. See github issue #27.
plotBeta failed with argument plotTree = FALSE together with SpeciesOrder = "Tree".
Spatial models with Nearest Neighbour Gaussian Process (NNGP) failed when the number spatial locations was not equal to the number of sampling units. This could happen, for instance, if there are multiple observations on the same spatial location. The problem still persists in spatial models with Gaussian Predictive Process (GPP).
Hmsc models can be modified using update(<Hmsc model>, <new arguments>). This was achieved by adding a call component per wish in github issue #34.
evaluateModelFit can handle probit models where binary data were given as TRUE/FALSE. Earlier only numeric data (0/1) were accepted. See github issue #30.
biPlot uses equal aspect ratio in ordination biplots.