Posterior samples from several independent Hmsc
objects can be combined as new chains with new method function
c(). This provides an easy alternative for distributed
computing. The user should take care that these independent models are
defined similarly so that they really can be combined. The function
tests for similarity of objects, but it only gives warnings and can
allow combination of incompatible models at user will. The user should
be careful not to start these models from the same
random number seed as these just duplicate your data instead of adding
independent new samples.
sampleMcmc allows the use of fork clusters instead
of socket clusters. Socket clusters are still the only alternative in
Windows, but other platforms can profit from the use of fork clusters
which may have lower memory use and are faster to set up, and also may
be marginally faster. The choice can be made with new argument
useSocket (defaults TRUE).
Updaters in sampleMcmc can occasionally fail in
extreme Hmsc models. This is no longer an error that stops
analysis, but sampling tries to recover from failures. The numbers of
failures for each updater is reported with the result. If there are only
a low number of failures, the sampling is safe to use. If there are
updater errors only in some chains, these chains can be removed, but
other chains can be used. See issue
#123.
New experimental function pcomputePredictedValues
with more aggressive parallelization than
computePredictedValues. In old code chains within each
partition could be run in parallel, but partitions were run serially. In
the new function, all chains and partitions can be run in parallel. The
plan is to replace the old function with this new alternative, but at
the moment both functions are available for testing. See issue
#142.
Implemented longitude-latitude coordinates and user-supplied distance matrices for NNGP spatial models. Sanity checks for spatial model input were improved.
Improved support for spatial models defined via distance matrices instead of spatial coordinates.
constructGradient provides wider choice of
coordinates for centroid of new_unit, including user-set
and infinite (meaning no spatial dependence) coordinates.
Detect cases when user tries to analyse posterior samples of
non-sampled Hmsc object to avoid confusing error messages
such as reported in issue
#125.
sampleMcmc with
initPar = "fixed effects" failed if Y
variates had missing values. The choice "fixed effects" was
undocumented in the package, but was used in several scripts at large.
See issue
#101.
The default number of neighbours in NNGP spatial models was not known in all posterior analysis tools giving very obscure error messages. Reported by Ben Weigel (Uni Helsinki).
Covariate-dependent latent loadings did not have correct alignment.
predict did not honour setting start
and thin which could result in huge output data that
exhausted memory. See issue #86.
predict failed in NNGP spatial models. See issue #96.
predict failed with one-dimensional spatial data.
See comments in unrelated issue #61.
Missing values are handled better in predict, but
they are still not allowed in all cases.
prepareGradient failed with geo-referenced spatial
random levels.
More robust handling of models that were fitted with model matrix
X instead of model frame XData and model
formula XFormula. Concerns functions biPlot,
computeVariancePartitioning and
constructGradient. Fixes issue
#126.
biPlot has improved handling of colour scaling of
continuous variables.
plotGradient gained argument to show the support of
trend for continuous variables. Main title can be shown for factor
variables (earlier it was shown only for continuous variables).
Grids of knots for Gaussian Predictive Process (GPP) are centred
for the coordinates in constructKnots. More knots were
produced than requested.
Prediction failed in spatial models with
predictEtaMean = TRUE.
Prediction failed in spatial NNGP models.
constructGradient (and hence
plotGradient) ignored specified order of factor levels. See
github issue
#63.
Performance inefficiency issues were fixed in NNGP models and some updaters.
User interface is more robust and can handle several inputs that earlier caused errors (often with confusing and obscure error messages). Input data is checked more carefully to avoid misleading results because of wrongly interpreted data.
User interface changes fix several github issues: #65, #66, #68, #70, #71, #78, #80, #81, #82.
Spatial and phylogenetic data are inspected more carefully to avoid errors in sampling.
Updaters are automatically disabled when needed instead of producing an error.
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.