ofemt() no longer accepts nmin_cell or alpha_bonferroni. Both were kept as deprecated aliases; they are now removed. Use min_per_cell instead of nmin_cell, and p_adjust_method = "bonferroni" instead of alpha_bonferroni = TRUE. Since the package has never been on CRAN, the disruption should be limited to a rename in existing scripts. Note that dropping the two arguments also shifts the positional order of the remaining ones, so calls that relied on argument position (rather than name) need to be revisited.
p_adj is now the median of the per-run adjusted p-values. Previously the median p-value was computed first and adjusted afterwards. Now stats::p.adjust() is applied within each sampling run, across that run’s pairwise comparisons, and the reported p_adj is the median of the resulting distribution. This keeps p_value and p_adj on the same footing — both medians over the n_s runs — and makes the median line in plot_pvalue_hist() agree with the table. Numeric results change when p_adjust_method != "none".
keep_components = "light" now stores the complete ofe_grid object (full grid, selection, per-cell stats and parameters) rather than only the selected cells, so the result can be plotted directly.
Bug fixes
Treatment labels are no longer rewritten internally. Spaces used to be replaced by dots ("Bortrac + Zintrac" became "Bortrac.+.Zintrac"), which leaked into Means comparison, Cells per treatment and the Comparison column, and could produce NA letters when a label collided with the separator used by multcompView::multcompLetters(). Labels are now mapped to internal placeholders for the whole computation and restored on output, so any label — spaces, +, -, parentheses, accents — is reported verbatim. The same rewriting is gone from select_grid(), where it could collapse two genuinely different treatments into one and make a mixed cell look single-treatment.
The compact letter display now follows the ordering of the means: treatments are sorted by decreasing median response before the letters are assigned, so "a" always marks the highest-yielding group.
The upper bound passed to spdep::dnearneigh() is exclusive, so the one cell whose nearest-neighbour distance defineddmax was being dropped from its own neighbour graph. The bound is now nudged to include it. This silently degraded Rho, Moran’s I and the ESS on every analysis; on the bundled ofe_f2 example the corrected values are ESS 97 / Rho 0.652 / Moran’s I 0.548 (previously 99 / 0.646 / 0.529).
ofemt() no longer flips spdep’s global ZeroPolicyOption as a side effect when an empty neighbour set is encountered; the policy is passed explicitly to the calls that need it.
Response and treatment column names are quoted before being used in a formula, so columns whose names contain spaces or other special characters ("Rinde kg/ha (seco)") no longer fail. Point coordinates are stored under dot-prefixed names, so user columns literally named X or Y are no longer shadowed.
Sampling a treatment that contributes a single cell no longer falls into sample()’s length-one trap (sample(5, 1) sampling from 1:5).
Reporting and plotting
spdep’s “neighbour object has N sub-graphs” warnings are muffled. They are expected for OFE trial shapes and carry no consequence for the diagnostics. What is now reported instead is the condition that actually matters: an error when no cell has any neighbour (the distance threshold is too small for the selection), and a warning naming how many cells were left isolated.
plot_grid_selection() accepts an ofemt_result directly, and plot() methods are provided for both ofe_grid and ofemt_result. On a result produced with keep_components = "full" a bare plot(res) draws the full grid, the selected cells and the observations, with no further arguments — the view needed to judge the effect of cellsize, shift, angle_deg and buffer. Selected cells are shaded translucently so the points underneath stay visible, and the plot title repeats the grid parameters so successive calls can be compared.
Both plotting functions gained an engine argument, "ggplot2" (default) or "base". The ggplot2 engine places the legend outside the plotting panel, so it can never sit on top of the data and the result no longer depends on the size of the graphics device — the base engine has to fit the legend inside the panel, which on a dense grid or a small window left it unreadable. ggplot2 stays in Suggests: when it is not installed both functions fall back to "base" with a message.
plot_grid_selection() gained point_size (and legend_pos, which accepts base-style corner keywords under either engine). Yield-monitor data runs to tens of thousands of observations, where the default dot size still reads as a solid mass; lowering it brings the cell boundaries back.
Unselected grid cells are now filled a very light grey instead of left transparent, so they read as cells rather than as page background and their legend key is visible.
plot_pvalue_hist() now draws a solid line at the median of the distribution in addition to the dashed line at alpha, and defaults to the adjusted p-values whenever the analysis used a p_adjust_method (override with which = "raw" or which = "adjusted").
print() reports the total number of grid cells alongside the number selected, states which adjustment produced p_adj, and notes that the means table is sorted by decreasing mean.
?ofemt now documents every component of the returned object — including perm_runs, the per-run empirical p-value distribution the method is built on — and spells out what each keep_components level stores and costs.
Other changes
ofemt() exposes a seed argument that controls reproducibility of the permutation sampling. Default is seed = 7L; pass NULL to let results vary across runs.
ofemt() accepts a pre-built grid via the grid argument. When grid = NULL the function builds one internally with make_ofe_grid(); the two paths are interchangeable when called with matching arguments.
make_ofe_grid() exported as the canonical way to construct and inspect the analysis grid before running the test.
Dependencies reduced: dropped dplyr, magrittr and tidyr from Imports. ggplot2 moved to Suggests.
ofemt() rewritten on top of the new grid pipeline; cells are filtered once in select_grid(), not twice.
Moran’s I is now extracted by name ($estimate["Moran I statistic"]) instead of by position, with a dependency-contract test guarding against future renames in spdep.
Test suite added (testthat 3rd edition) covering grid construction, end-to-end ofemt() on ofe_f2, reproducibility from seed, error paths, and contracts with upstream dependencies.
Continuous integration on GitHub Actions: R CMD check on Ubuntu/macOS/Windows times R release+devel+oldrel; pkgdown site deployed on push to main.