v1.49.1

Changes

• For contributors: python3 run_tests.py runs the test files in parallel, a full run in about 5 minutes instead of 30. run_tests.py offset runs only the matching files.

v1.49.0

Features

• New option effect_labels. 'cohen' uses Cohen's values as lower thresholds (negligible, small, medium, large), as most readers expect; 'centred' keeps the seven bins centred on them.

Changes

• The default is now 'cohen', so the effectSize labels in all tables change on rerun: d = 0.45 was "medium", now "small". Numbers are unchanged. Set effect_labels = 'centred' for the previous labels.

v1.48.0

Changes

• Summary.txt lists the ranges of the effect size labels (very small to very large) for the measures it uses, and STATISTICAL_NOTES tabulates them for partial η², Cohen's d and r. The labels are bins centred on Cohen's values, not cutoffs at them: d = 0.45 is already "medium".

v1.47.3

Bugs

• With a nested random term such as (1|subject/rater), Summary.txt called the subjects' ICC the agreement ICC(2,1), which holds only for crossed factors. It now names nested models as such and adds the test-retest ICC (incl. subject). Values were already correct.

v1.47.2

Changes

• Summary.txt and the docs call the one-way ICC ICC(1,1), as ICC(2,1) and ICC(3,1). It was ICC(1), the shorthand for the same form. Values unchanged.

v1.47.1

Bugs

• Summary.txt called every ICC "ICC(1)". It now names the form the model fits: ICC(1) one-way, ICC(2,1) agreement with a random rater, ICC(3,1) consistency with a fixed rater. Values were already correct. See STATISTICAL_NOTES for the three models.

v1.47.0

Features

• y_transform takes one entry per dependent variable as a list, like constraints. Summary.txt names the transform.

Bugs

• With y_transform, data plots showed rows excluded by constraints, each excluded level with its own marker, and ignored x_order and rename. Statistics were unaffected; plots made with y_transform and any of these options must be regenerated.
• An x_order naming only some levels dropped the others from the analysis without a word, or failed with an R error under a constraint. Unnamed levels now follow the named ones. Results from such runs must be recomputed.

v1.46.0

Features

• New distribution 'negative_binomial' (glmmTMB's nbinom2) for counts that vary more than Poisson allows. Poisson understates the standard errors of such counts. Summary.txt reports the estimated θ; a large θ means Poisson would do.

v1.45.0

Changes

• A numeric variable in x with more than ten values, most of them distinct, now stops with ContinuousFactor, a plain message pointing to covariate. As a factor it cost one parameter per value; three such variables crossed could exhaust memory and kill a Jupyter kernel without explanation.
• interaction = 'all' and an order such as 2 now cross the covariates too; before, covariates stayed additive without a word. Analyses with covariates and one of these settings fit a different model from this version on. 'auto', the default, still leaves covariates additive.

v1.44.1

Bugs

• On a JupyterHub, glmmTMB warned on every non-Gaussian fit that it was built with a different TMB version, because conda-forge installs the newest TMB. The installer now installs the matching TMB. Existing installations: run the install line again, then restart the kernel. Results are usually unaffected, but were not guaranteed.