mlmoderator 0.3.0

Corrections to inference (results change)

Degrees of freedom. All tests and intervals now use Satterthwaite degrees of freedom by default (via ‘lmerTest’). Versions up to 0.2.1 used N - p (level-1 rows minus fixed effects), which treats observations as independent and is anti-conservative for cross-level interactions: in a simulation with 10 clusters and no true interaction it rejected at about twice the nominal rate. The new df_method argument offers "satterthwaite" (default), "kenward-roger", "between" (J - q - 1; Snijders and Bosker, 2012), and "residual" (the old rule, kept only for reproducing earlier results).
Johnson-Neyman boundaries are now exact: closed-form roots of the Bauer-Curran quadratic for constant-df methods, and root-finding on the exact criterion for Satterthwaite and Kenward-Roger. Previously they were interpolated from p-values on a grid. mlm_jn() also returns jn_bounds_all (all real roots, including those outside the data) for constant-df methods.
Confidence bands in mlm_plot() now use the full fixed-effects design row. Previously covariates other than the predictor and moderator were omitted from the standard error of the fitted values.
Moderator probe values for a cluster-level moderator are now the mean and SD across clusters, not across observations (new modx.level argument; "observation" restores the old behaviour).

The ICC-shift analysis and robustness_index were removed. The design-effect rescaling it used applies to means under a random-intercept model, not to cross-level interactions, and the adjusted Johnson-Neyman boundary it reported was not correct. icc_range and icc_grid now give a warning and are ignored; loco = FALSE is an error.
Leave-one-cluster-out refits now use update() on the original data, so they keep the original REML/ML setting, weights, and formula. Previously they always used ML (so every “change” included the ML/REML difference) and failed silently for formulas with transformed variables.
The influence measure is now DFBETA, the change in the interaction divided by its full-data SE, flagged at 2/sqrt(J). The previous “Cook’s-distance-style” measure divided by the SD of the changes and so flagged clusters by construction. Failed refits are reported with their error messages.

pct_random was removed: it divided a population variance by a sampling variance and so grew with sample size. The output now reports tau11 (variance), tau11_sd, and df explicitly, and the documentation describes the result as confidence intervals for the average slope versus prediction intervals for a new cluster.
The random slope is taken from the grouping factor that carries pred.

Tests now cover every exported function, including checks against ‘lmerTest’, ‘emmeans’, and ‘pbkrtest’.
New vignette with the public High School and Beyond data (‘mlmRev’).
Removed a reference to a non-existent vignette.
URLs point to github.com/subirhait/mlmoderator.