The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.

fido 1.1.5

This release fixes a long-standing error in how the prior degrees of freedom (upsilon) were used. Results from pibble(), basset() and orthus() fitted with earlier versions may change; please refit.

What was wrong

fido fits models in two steps. It first finds Eta using a “collapsed” model in which Lambda and Sigma have been integrated out. It then draws Lambda and Sigma given Eta (the “uncollapse” step). Both steps should use the same inverse-Wishart prior on Sigma, with upsilon degrees of freedom.

The collapsed step expects its degrees of freedom on a different scale (the matrix-t scale). Converting the inverse-Wishart upsilon to that scale requires subtracting D - 2. Since 2018, pibble() passed upsilon to the collapsed step without this conversion. The collapsed step therefore used a prior on Sigma that was more informative than the one requested. The uncollapse step used the correct prior, so the two steps did not agree.

Who is affected

• pibble() and basset() (including additive basset() models) with three or more categories (D > 2). With D = 2 the two scales coincide and results are unchanged. The larger D is relative to upsilon, the larger the effect.
• orthus() in all cases, because it had a similar conversion error (upsilon + P was used where upsilon - D + 2 is correct).
• maltipoo() (experimental) received the same fix.

Other fixes

• The log marginal likelihood reported by the Laplace approximation (logMarginalLikelihood) was missing several constant terms and is now correctly normalized. Stored values from earlier versions (for any D) should not be compared with new ones.
• pibble(), basset(), orthus() and maltipoo() now stop with an error if upsilon is too small for a valid inverse-Wishart prior (upsilon <= D-2, or upsilon <= D+P-2 for orthus()). Previously such values were accepted silently and produced meaningless fits.
• orthus_sim(true_priors = FALSE) now returns upsilon = D+P (previously D, which is not a valid prior for the (D-1+P)-dimensional covariance).
• orthus_sim(true_priors = TRUE) now uses upsilon = D+P+10 so the inverse-Wishart prior mean equals the simulated Sigma, including with the default dimensions. Previously the scale was zero or negative when P >= 10.
• orthus_sim() now preserves matrix dimensions with two categories or a single Gaussian variable, sample, or covariate. Invalid dimensions are rejected before simulation, and its fixed covariance entry no longer makes the simulated covariance singular when both corresponding variances are one.
• Non-integer upsilon was silently rounded down when drawing Sigma; it is now used as given.
• ret_mean = TRUE now errors when upsilon + N is too small for the conditional posterior mean of Sigma to exist. Use posterior draws (ret_mean = FALSE) with these proper, heavy-tailed posteriors.
• basset() with more than one component (lists of Theta and Gamma) passed ncores into the ret_mean slot of the component sampler. When ncores was set this could return component means instead of posterior draws, and the seed and ncores settings were ignored. Arguments are now passed by name, and each component uses its own random stream, so component uncertainty is now correct.
• Added fido_build_info() to report native runtime threading configuration and compiler support for OpenMP, Eigen, and MKL.
• Default thread limits are capped at two when R sets _R_CHECK_LIMIT_CORES_=TRUE for package checks.
• Added opt-in support for compiling with Intel MKL and GNU OpenMP via --configure-args="--with-mkl" on supported Linux x86_64 systems, with automatic thread safety guards during parallel uncollapsing.
• Fixed error handling in LaplaceApproximation to verify Cholesky factor status and preserve block-diagonal partial Hessian handling when MKL is disabled.

What you need to do

• Refit analyses from the original data and priors. Old fits can still be loaded, but there is no way to correct their draws by adjusting upsilon after the fact; predictions from old fit objects reflect the old results.
• The meaning of the upsilon argument has not changed: it is the standard inverse-Wishart degrees of freedom. For pibble() and basset(), the prior mean of Sigma is Xi / (upsilon - D) when upsilon > D. For orthus(), it is Xi / (upsilon - D - P) when upsilon > D + P. Existing scripts need no edits.
• If you call the low-level functions optimPibbleCollapsed(), loglikPibbleCollapsed(), gradPibbleCollapsed() or hessPibbleCollapsed() directly, note that they take matrix-t degrees of freedom. Pass upsilon - D + 2 rather than upsilon (see their help pages). uncollapsePibble() still takes upsilon unchanged.

fido 1.1.4

• Fixed the posterior mean of Sigma returned by uncollapsePibble(..., ret_mean = TRUE) to use the correct inverse-Wishart mean.
• Fixed OpenMP sampling in the uncollapse routines so posterior draws are reproducible across ncores for a fixed seed and no longer depend on overlapping thread RNG streams.
• Added input validation to the multinomial-Dirichlet bootstrap so Dirichlet concentration parameters must be positive and finite.

fido 1.1.3

• Replaced explicit integer sequences with seq_len() and seq_along() for improved robustness.

fido 1.1.1

• Fixing a minor bug in the configure script

fido 1.1.0

• Updating the basset model to allow for additive Gaussian processes.
• Updating the random seed generation.
• Removed maltipoo as an export. There were a few small changes to names of internal maltipoo functions (e.g., “verify.maltipoo” became “verify_maltipoo”).
• Updating the configure script to a compile but that occured when OpenMP is present but R is not configured to use it.
• Fixed a few small bugs (predicting with one sample; wrong constant; etc.).

fido 1.0.4

• changing default optimizer (lbfgs rather than adam), adam is more sensitive to random seed
• adding new r2 methods for calculating Bayesian R2 for pibble and basset models
• predict.bassetfit now accepts a newdata=NULL argument to predict on training set

bug fixes

• fixing a few discrepancies with base R commands (print, summary, etc).

fido 1.0.3

bug fixes

• Updating documentation under the newest version of R (4.2.1)

fido 1.0.2

bug fixes

• Removing a timing test because it is irrelevant and can fail under certain set-ups.

fido 1.0.1

bug fixes

• Changing phyloseq from imports to depends, was causing issues installing from CRAN.
• Fixing bug in configure.ac that was causing openmp parallelization to not work on some machines.

fido 1.0.0

More changes to make the final push for CRAN! Changes include:

• Updating the description file to match CRAN’s standards
• Removing the dependency to driver
• Linking the one remote package (MicrobeDS) to a separate CRAN-like Github repo using the “drat” package. This is to match CRAN’s policy against no use of “Remotes”.
• Precomputing the longer running vignettes, cutting R CMD check times in half. The original vignettes can be found in the vignettes folder with a “.orig” extension for reproducibility.

fido 0.1.14

• More changes to prepare for version 0.2 (and CRAN soon), including:
• Lots of cleaning of small warnings and notes
• Fixing erorrs with the base plot function
• Cleaning up some manual entries
• Fixing the errors posted on Github (to_ilr among others)

fido 0.1.13

• tons of tiny changes to prepare for version 0.2 (and ultimately CRAN) featured changes include:
• plot and coef methods for *fit objects now abide by CRAN generic naming scheme - may cause some problems if prior code had positional arguments to these functions
• now being tested using travis integration (vignettes not tested there due to time constraints)
• merged Kim’s fixes to maltipoo code (will almost certainly change maltipoo results; there had been a bug in prior versions)
• Internally stored mallard and mallard_family data now import as mallard and mallard_family rather than both importing as ps. Solved a R CMD CHECK warning.
• Lots of tiny updates to documentation

fido 0.1.12

• orthus added for jointly fitting multinomial and gaussian data (e.g., 16S microbiome and metabolomics data)

fido 0.1.11

• Kim fixed windows installation problems (Thanks Kim!)
• Fixed bug in predict that was effecting ppc
• Fixed error when pars!=NULL in summary
• lambda_to_iqlr now had default behavior when focus.cov=NULL
• Fixed issue with Xi=NULL in pibblefit when trying to transform
• small changes to make pibble wrapper around c++ functions faster

fido 0.1.10

• basset added for fitting non-linear regression using fido
• numerous changes to make installation easier
• Added a NEWS.md file to track changes to the package.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.