NEWS.md

In JenniNiku/gllvm: Generalized Linear Latent Variable Models

Version 1.4.5

• Separated "n.init" functionality into gllvm.iter.R
• Prep for parallelisation
• Enabled parallelisation (see TMB::openmp)
• Largely vectorized "residuals.gllvm", and residuals in "gllvm.aux"
• Added covariance of random effects to summary

Version 1.4.4

• Removed "dependent.row" feature
• Added possibility for multiple random row intercepts
• Added possibility for (correlated) random species random effects
• Can be plotted with "randomCoefPlot"
• Added possibility to Phylogenetically structure the random species effects
• Phylogenetic signal parameter is included as object$params$rho.sp
• Can be covariate specific
• num.RR and num.lv.c can now be larger than the number of predictors if randomB!=FALSE
• Added "iid" option for "randomB"
• Added a "getEnvironCov" function to extract species associations due to random covariate effects

Version 1.4.3

• For CRAN release 1.4.3 see updates for versions 1.4.2 and 1.4.3

Bug Fixes

• Bug in correlated row effects fixed
• Bug in getPredictErr for models fitted with LA fixed, and it returns now prediction errors for random slopes of X covariates as well
• Bug in randomCoefplot fixed

Version 1.4.2

New Features

• Added a correction factor to the second partial derivatives of the canonical coefficients for concurrent and constrained ordination
• Added randomCoefPlot functionality of constrained and concurrent ordination models with random slopes. Currently not supported for models with quadratic responses
• Summary now provides the possibility to calculate wald statistics across LVs or predictors for concurrent and constrained ordination
• coef now renames parameter estimates with more intuitive names and allows to subset the parameter list with names
• Tweedie power parameter is estimated now if set to NULL in `gllvm.
• VA support for Zero-inflated poisson distribution
• Zero-inflated negative-binomial distribution added
• Binomial (Ntrials>1) support added (previously only Bernoulli)
• Now allowed to have (some) NAs in the response data

Bug Fixes

• Fixed an issue with structured row-effects in concurrent and constrained ordination
• Fixed a bug that prevented plotting prediction regions for constrained ordination with structured row-effects
• No standard errors should be returned by optima.gllvm and tolerances.gllvm with randomB != FALSE
• Species names were in the original order with order = TRUE in RandomCoefPlot
• Fixed an issue that arose when {0,1} bounded parameters reached the bounds
• Various bug fixes for constrained/concurrent ordination with random intercepts and random slopes
• Bug in predictions with structured row intercepts was fixed, see issue #86

Version 1.4.1

New Features

• Computational stability of random slopes for constr. and concr. ordination significantly improved
• Computational stability of quadratic model significantly improved
• Unstructured VA covariance matrix for quadratic models with random intercepts
• Added example for se.gllvm

Bug Fixes

• Bugfix in random slopes for concr. ordination with LV-specific variances and random row intercepts
• Bugfix for quadratic model with Poisson, NB, gamma, or exponential responses
• Bugfix in starting values for constrained and concurrent quadratic model

Bug Fixes

• Valgrind error fixed

Version 1.4.0

New Features

• For CRAN release 1.4.0's new features see features described for versions 1.3.2-1.3.3

Bug Fixes

• For bug fixes to CRAN release 1.4.0 see versions 1.3.2-1.3.3

Version 1.3.3

New Features

• The n.init option has been improved, so that it stops if no improved fit has been found after n.init.max (defaults to 10) iterations.
• Row names from the data now carry over to the site scores, so that they can be displayed in ordiplot

Bug Fixes

• Memory allocation problem in development version fixed

• Diagonal elements of loading matrix 'theta' fixed for fourth corner model

• Bug in 'predict' for random slopes fixed, occurred when new x-covariate values were given

Version 1.3.2

New Features

• Ordination with predictors (num.RR,num.lv.c) is now implemented with constrained optimization routines (alabama,nloptr) as long as the canonical coefficients are treated as fixed-effects. This follows from the necessary identifiability constraints.

• The reduced-rank approximated predictor slopes of a multivariate regression can now be plotted (with confidence intervals) using coefplot. Not available yet for quadratic effects.

• Separate checks are put in place to warn users if the constraints on the canonical coefficients (orthogonality of the columns) have not converged.

• Separate checks are put in place to warn users if the coefficients of a quadratic model have not converged

• Canonical coefficients in ordination with predictors (num.RR,num.lv.c) can now be treated as random-effects using the 'randomB' argument. For the moment, all need to be either random or fixed, no mixing. Prediction intervals can be retrieved with the getPredictErr function.

• An extended version of the spider dataset has been made available

• Added an option to magnify the x-axis labels in coefplot

• Site names present as row labels in the response data are now shown in the ordination plot

Bug Fixes

• The order of the quadratic coefficients was wrong when num.RR, num.lv, and num.lv.c were all used in the same model.

• Fixed a bug in the calculation of starting values for constrained ordination (num.RR) where the residuals were not re-calculated if num.lv.c>0

• Fixed a bug in coefplot for when only one predictor was included in the model

• Fixed a bug that would prevent using a gllvm with quadratic response model as starting values for another model

• Changed import/export of various functions as requested in github issue #65

• Various minor tweaks to the summary function

Version 1.3.1

New Features

• Structured row parameters are implemented, including a possibility for between or within group correlations for random row effects.

• Constrained ordination model is implemented.

• NB and binomial (with probit and logit) response model implemented using extended variational approximation method.

Bug Fixes

• Vignettes are removed from the CRAN version of the package, can be seen at the package's website only.

Version 1.3.0

New Features

• Quadratic latent variables allowed, that is term - u_i'D_j u_i can be included in the model using 'quadratic = TRUE'. In addition, functions 'optima()', 'tolerances()' and 'gradient.length()' included.

• Beta response distribution implemented using Laplace approximation and extended variational approximation method.

• Tweedie response model implemented using extended variational approximation method.

• Ordinal model works now for 'num.lv=0'.

• Residual covariance adjustment added for gaussian family.

Bug Fixes

• Estimation of the variances of random slopes of the X covariates didn't work properly when 'row.eff = FALSE' or 'row.eff = "fixed"'.

• Problems occurred in calculation of the starting values for ordinal model.

• Problems occurred in predict() and residuals(), when random slopes for X covariates were included.

• Problems occurred in predict() when new X covariates were given.

• Problems occurred in predictLVs() for fourth corner models.

Version 1.3.1

New Features

• Structured row parameters are implemented, including a possibility for between or within group correlations for random row effects.

• Constrained ordination model is implemented.

• NB and binomial (with probit and logit) response model implemented using extended variational approximation method.

Bug Fixes

• Vignettes are removed from the CRAN version of the package, can be seen at the package's website only.

Version 1.3.0

New Features

• Quadratic latent variables allowed, that is term - u_i'D_j u_i can be included in the model using 'quadratic = TRUE'. In addition, functions 'optima()', 'tolerances()' and 'gradient.length()' included.

• Beta response distribution implemented using Laplace approximation and extended variational approximation method.

• Tweedie response model implemented using extended variational approximation method.

• Ordinal model works now for 'num.lv=0'.

• Residual covariance adjustment added for gaussian family.

Bug Fixes

• Estimation of the variances of random slopes of the X covariates didn't work properly when 'row.eff = FALSE' or 'row.eff = "fixed"'.

• Problems occurred in calculation of the starting values for ordinal model.

• Problems occurred in predict() and residuals(), when random slopes for X covariates were included.

• Problems occurred in predict() when new X covariates were given.

• Problems occurred in predictLVs() for fourth corner models.

JenniNiku/gllvm documentation built on May 3, 2024, 2:15 a.m.