Changes in version 2.1.0 (2026-03-05)                  

  - Breaking change: the backtransform_response argument has been
    replaced by pred_transform, a generic hook to transform predicted
    values (and confidence limits), whether computed on the link or
    response scale (as determined by pred_type). Warnings are issued to
    clarify when the transformation is applied on the link scale or
    after an inverse-link step (for models with non-identity links).
  - Maintenance: improved robustness of random-effects detection for
    mixed models. Refined namespace checks for lme4, glmmTMB, and
    reformulas to ensure optional dependency handling works correctly
    when suggested packages are not installed.
  - Added internal fallback detection of grouping variables in
    mixed-model formulas when optional parsing backends (reformulas,
    lme4) are not available.
  - Improved handling of random-effects conditioning in prediction
    summaries and returned data frames. When re_form = NA
    (population-level predictions), grouping variables used solely for
    random effects are no longer included in the returned prediction
    data frame unless explicitly requested (e.g., via by). In this case,
    a message is retained to clarify that random effects are excluded.
  - Improved handling of confidence intervals when standard errors are
    unavailable (e.g., conditional predictions from mixed-effects
    models). In such cases, confidence bounds are now returned as NA
    rather than producing invalid or misleading values.
  - Improved and expanded examples.
  - Minor internal code clean-up and stability improvements.

                 Changes in version 2.0.2 (2026-02-25)                  

  - Maintenance: updated internal handling of random-effects formulas to
    use reformulas::findbars() following upstream changes in lme4 and
    glmmTMB. Added reformulas to Imports to prevent deprecation
    warnings.

                 Changes in version 2.0.1 (2026-01-24)                  

  - Bug fix: the legend placement preset legend_position = "out" is now
    computed per panel in multi-panel graphics layouts (e.g.,
    par(mfrow=...), par(mfcol=...), layout()). Legends are consistently
    positioned above their relevant plot region rather than between
    panels.
  - Documentation: updated the examples to use exposure as the offset
    variable name (replacing offset_var) for consistency and clarity.

                 Changes in version 2.0.0 (2026-01-08)                  

  - Major internal redesign: the package has been fully modularized to
    improve clarity, maintainability, and extensibility. Core
    functionality is now organized into dedicated components for data
    preparation, newdata construction, prediction, plotting, and output
    assembly.
  - Added support for betareg models: confidence intervals for the mean
    submodel are computed using a delta-method approximation.
  - Added support for survival models fitted with survival::coxph().
    Conditional effects are plotted on the linear predictor (log-hazard)
    scale. Raw survival data (event times and censoring) are not
    plotted, and confidence intervals are based on standard errors of
    the linear predictor.
  - New by_breaks argument: allows users to manually specify
    conditioning values when by is numeric, replacing the previous fixed
    default behavior. If not provided, numeric by variables with more
    than 6 unique values default to their 10th, 50th, and 90th
    percentiles.
  - New rlm_vcov argument: robust standard errors for MASS::rlm() models
    can now be computed using any sandwich estimator type (e.g., "HC3",
    "HC5", "const", etc.), a user-supplied covariance matrix, or a
    custom function.
  - New ci_polygon_alpha argument: allows users to control the
    transparency of polygon-style confidence interval bands when ci_type
    = "polygon". Defaults to 0.5 and accepts values between 0 and 1.
  - New show_conditioning argument: allows users to print a concise
    summary in the console describing how predictions are conditioned.
    When show_conditioning = TRUE, easyViz() reports which variables are
    held fixed and which vary in the prediction grid, whether
    predictions from mixed-effects or GAM models are conditional on
    random effects or marginal, and, for numeric by variables, the
    values at which predictions are evaluated (e.g., quantiles or
    user-specified by_breaks). This option does not affect plots or
    returned values and is intended as a diagnostic aid.
  - New plot argument: allows users to suppress graphical output. When
    plot = FALSE, easyViz() returns only the predictions (as an
    invisible easyviz.pred.df object) without drawing a plot. Useful
    when storing or further processing predicted values.
  - Improved offset handling: offsets specified either in the model
    formula or via offset = are now automatically detected. For log-link
    count models (e.g., Poisson, negative binomial, COM-Poisson,
    generalized or truncated count models), observed responses are
    automatically scaled by the offset for plotting, and the prediction
    grid sets the offset to 1 by default, ensuring predictions and raw
    data are displayed on a common rate scale.
  - Improved handling of numeric by variables: numeric conditioning
    variables with a small number of distinct values (currently ≤ 6) are
    now treated as discrete groups and plotted with one line per value
    (like a factor), instead of always using quantile cross-sections.
    For truly continuous numeric by variables (more than 6 unique
    values), easyViz() still uses default cross-sections (10th, 50th,
    and 90th percentiles) or the user-supplied values from by_breaks.
    This makes the behavior more intuitive for common cases such as 0/1
    or small ordinal scales.
  - Improved stability and consistency in handling numeric conditioning
    variables, interactions, binomial data, and GAM models with
    random-effect smooths.
  - Improved legend behavior: when a by variable is specified and users
    do not explicitly set add_legend, easyViz() now automatically
    enables the legend. This ensures that multi-line or multi-group
    plots always display a clear legend by default, while preserving
    user control when add_legend is manually set.
  - New legend placement preset "out" for legend_position: setting
    legend_position = "out" applies a predefined layout that places the
    legend outside the plotting region above the panel.
  - Documentation and diagnostic messages significantly expanded and
    clarified.

                 Changes in version 1.2.0 (2025-11-30)                  

  - Renamed re.form to re_form for consistency with other arguments.
    Argument re.form is deprecated. Please use re_form instead.
  - Added ci_level argument to control the confidence interval level
    (default is 0.95). The confidence level used for the interval is
    included in the returned prediction data frame.
  - Documentation improvements.

                 Changes in version 1.1.0 (2025-08-21)                  

  - Added support for the new pred_resolution argument, which controls
    the number of prediction points used for numeric predictors. The
    default is 101 (matching visreg), but higher values can be helpful
    for predictors with a wide range or when visualizing nonlinear
    relationships.
  - Added support for ci_type = NULL to suppress confidence intervals
    for numeric predictors.
  - Added support for custom legend titles using the new legend_title
    argument.
  - Added support for adjusting legend title text size using the new
    legend_title_size argument.
  - Added support for horizontal legends using the new legend_horiz
    argument.
  - Added support for fine-tuned legend customization via the new
    legend_args argument, which allows passing any additional parameters
    to legend() (e.g., pt.cex, bg, box.lwd, inset).
  - Renamed legend_text_size to legend_label_size for consistency with
    legend_title_size and legend_labels. The argument legend_text_size
    is deprecated. Please use legend_label_size instead.
  - Added support for categorical predictors: users can now customize
    x-axis text size, color, and font using plot_args (e.g., cex.axis,
    col.axis, font.axis) when the predictor is a factor.
  - Added finite-sample adjusted t-based confidence intervals for lm(),
    rlm(), gls(), and nls() models. For models with asymptotic inference
    (e.g., glm(), mgcv::gam(), glmer(), glmmTMB()), confidence intervals
    are based on standard normal approximations.
  - Improved visual clarity: when by is specified for interactions, all
    prediction lines are now drawn on top of their associated confidence
    intervals, ensuring they remain visible and are not obscured by
    overlapping CI polygons.
  - Improved detection of binomial-style responses.
  - Improved internal handling of link functions: the logit inverse now
    uses plogis() for greater numerical stability, and the cloglog
    inverse is now more robust to extreme values.
  - Added a diagnostic warning when user-supplied aesthetic vectors
    (e.g., point_col) have a length mismatch with the cleaned data used
    for predictions, helping users avoid misaligned plot aesthetics when
    the original data contains missing values.
  - Added an example demonstrating how to visualize 3-way interactions
    using the by and fix_values arguments in a multi-panel layout.
  - Documentation improvements.