---
title: "Prediction Contrasts, Rankings, and Groups"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Prediction Contrasts, Rankings, and Groups}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(collapse = TRUE, comment = "#>")
```

Prediction grids can be summarised by observed design variables without
expanding the approved model-family scope.

```{r, eval=FALSE}
grid <- create_prediction_grid(
  fit,
  at = list(condition = c("control", "treatment"))
)

pred <- predict_model(
  fit,
  newdata = grid,
  type = "expected",
  include_group_effects = FALSE
)

prediction_pairwise_contrasts(pred)
prediction_interval_width(pred)
prediction_rank_probabilities(pred)
```

The ranking function is deliberately descriptive. A probability of rank one is
not converted into an automatic selection.

When the prediction data contain multiple rows per substantive group:

```{r, eval=FALSE}
grouped <- group_prediction_summary(pred, by = "condition")
grouped
plot_group_predictions(grouped, "condition")
```

This makes aggregation explicit and reproducible rather than hiding it inside
plotting code.