NEWS

---

CoDaStereo 3.1.1

• The hex logo now stands upright (pointy top and bottom), following the hex sticker standard; the cube no longer overlaps the wordmark. Favicons were regenerated from it.
• Fewer hard dependencies (18 instead of 24, ahead of a CRAN submission): purrr, tidyr, tibble and rlang were replaced by base R, the HTML report no longer needs knitr, and mirai is now optional (without it, models are fitted in the main process).

• The clr-biplot keeps equal axis scales unless the second axis would be squashed (it spans less than a quarter of the first, e.g. 99% vs 0.4% of the variance). In that case the scales are freed and the plot says that angles and lengths are distorted.
• The hex logo (tissue cube) is used in the app, README and website.
• README and website link to the online app; the website lives at https://provetelab.org/CoDaStereo/.

CoDaStereo 3.1.0

Data

• The four case-study datasets of the companion paper (spleen and kidney point counts per field, leukocyte differentials, erythrocyte nuclear abnormalities) are available as examples.
• Nested designs: fields of view can be aggregated into one composition per specimen (counts summed, areas averaged). A block term in the model restricts permutations within blocks.

Visualisation

• Every figure can be exported as PDF, PNG, SVG or TIFF at a chosen size (cm), resolution and font size.
• The clr-biplot is drawn in ggplot2 (factoextra is no longer a dependency). It adds axis choice, shape by a second factor, sample labels, a scree plot and a table of ray lengths. It can include the log-total as an extra variable (T-space biplot; Pawlowsky-Glahn, Egozcue & Lovell 2015).
• New heatmap of the variation matrix.
• Ternary plots can show 95% confidence or predictive regions (ilr normal model) and group geometric centres, computed on the zero-replaced composition.
• Balances tab: boxplots of each balance per group and a permutational test of each balance (F, Z, P).
• Interactive mode: hovering shows the sample on the ternary plot and biplot (ggiraph).

Statistics

• Numeric covariates (ANCOVA), block term, sums of squares of type I, II or III, and 199, 999 or 9999 permutations with a fixed seed.
• The log-total can enter as a covariate ("does composition change with size?") as well as a joint response.
• Pairwise tests also compare group dispersions (compositional disparity).

Speed

• Model fits and pairwise tests run in background workers (mirai), so the interface stays responsive. Plots are cached. The SBP grid and amalgamation inputs are debounced. Progress bars show long steps. The deployment entry point starts faster.

Reproducibility

• New Reproducibility tab with three outputs:
  • an R script that reproduces the analysis (tested to match the app's results);
  • a self-contained HTML report;
  • saving and reloading the session as JSON, with results recomputed on load.

Interface

• The interface is available in English, Portuguese and Spanish (EN / PT / ES in the header), including the guided tour. Figures, the report and the R script stay in English.
• Package website built with pkgdown.

CoDaStereo 3.0.0

New features

• New Balances (SBP) tab. A sequential binary partition is built automatically by Ward clustering of the parts on clr distances. Users can edit any cell (+1 / -1 / 0) or reset to the automatic SBP. Edits are validated on the fly (D - 1 balances, two-sided balances, one root split, nested partitions). The tab shows the balance definitions and a CoDa dendrogram (compositions::CoDaDendrogram()) with per-group boxplots of each balance. The dendrogram downloads as PDF and the SBP as CSV.

• Composition + total in the permutational MANOVA. The log-total can be added as an extra response coordinate next to the ilr coordinates, following Pawlowsky-Glahn, Egozcue & Lovell (2015, Statistical Modelling 15:175-190, doi:10.1177/1471082X14535526). Two totals are available: the sum of parts (space Ts, the default) and the product of parts to the power 1/sqrt(D) (space Tp, isometric to analysing the log of each part). A separate univariate model and a boxplot of the log-total separate size effects from compositional effects. The app warns when the totals are almost constant (CV < 5%), as happens when the number of grid points or cells counted is fixed by the sampling design.

• Selectable colour palettes (Okabe-Ito, ColorBrewer Dark2, viridis) for the ternary plot, clr-biplot and compositional descriptives.

• A collapsible "How to read this biplot" note explains rays, links, link angles and the limits of the planar approximation.

• New hex sticker.

Changes that affect results

• The clr-biplot now runs PCA on centred, unscaled clr coordinates. Earlier versions called FactoMineR::PCA() with its default scale.unit = TRUE, which standardised the clr variables. That equalised ray lengths, so links no longer approximated log-ratio standard deviations. Biplots produced with 2.x will differ.

Bug fixes

• The compositional descriptives and the CoDa dendrogram no longer fail when the app runs from the installed package (run_app()). compositions reads getOption("robust"), which is only set when the package is attached.
• The logo is now shown when the app is launched with run_app().
• The deployment entry point (app.R) loads rintrojs instead of the removed cicerone.

Package infrastructure

• The package passes R CMD check with no errors, warnings or notes: roxygen2 documentation, a complete NAMESPACE, and declared imports. Continuous integration was re-enabled on Linux, macOS and Windows.
• Added a testthat suite (palettes, SBP validity, composition + total geometry, clr-biplot geometry, and module smoke tests on simulated data).
• Added a getting-started vignette.
• manifest.json was regenerated for Posit Connect Cloud.

CoDaStereo 2.1.0

• Guided tour (now based on rintrojs), pre-flight check of selected parts, compositional descriptives per group, convex hulls on ternary plots and biplots, and a logo.