---
title: "Continuous outcomes"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Continuous outcomes}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

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

# Example

The example uses the `cancer` data from the survival package. To compare patient age by ECOG performance status, recode the variable `ph.ecog` to be categorical (a `factor`) and exclude patients with rare elevated `ecog.ps`:

```{r ex_continuous, message = FALSE}
library(dplyr)
library(rifttable)
data(cancer, package = "survival")
cancer <- cancer |>
  filter(ph.ecog < 3) |>
  mutate(ph.ecog = factor(ph.ecog))
attr(cancer$ph.ecog, which = "label") <- "ECOG performance status"

tribble(
  ~label,                                       ~type,
  "**Absolute estimates**",                     "",
  "Observations",                               "total",
  "Sum",                                        "sum",
  "Range",                                      "range",
  "Mean",                                       "",
  "  Mean (i.e., arithmetic mean)",             "mean",
  "  Mean (95% CI)",                            "mean (ci)",
  "  Mean (standard deviation)",                "mean (sd)",
  "  Geometric mean",                           "geomean",
  "Median",                                     "median",
  "Median (interquartile range)",               "median (iqr)",
  "",                                           "",
  "**Comparative estimates**",                  "",
  "Mean difference (95% CI)",                   "diff",
  "Median difference (95% CI)",                 "quantreg",
  "Mean ratio",                                 "",
  "  of arithmetic means",                      "fold",
  "  of arithmetic means, empirical SE",        "irrrob",
  "  of geometric means",                       "foldlog"
) |>
  mutate(
    exposure = "ph.ecog",
    outcome = "age"
  ) |>
  rifttable(
    data = cancer,
    diff_digits = 1, # Suppress unnecessary precision in most estimates
    # Show extraneous digits to highlight (minor) differences in ratio  models:
    ratio_digits = 3,
    overall = TRUE
  ) |>
  rt_gt() # obtain formatted output
```


# Absolute estimates per exposure category

`type` | Description | Options (`arguments = `)
-------+-------------+--------------
`"total"` | Count of observations.
`"sum"` | Sum.
`"range"` | Range: Minimum to maximum value.
`"mean"` | Mean (arithmetic mean).
`"mean (ci)"` | Mean and CI (default: 95%).
`"mean (sd)"` | Mean and standard deviation.
`"geomean"` | Geometric mean.
`"median"` | Median.
`"median (iqr)"` | Median and interquartile range.


# Comparative estimates with confidence intervals

`type` | Description | Options (`arguments = `)
-----+------------------+---------
`"diff"` | Mean difference from linear model. |
`"fold"` | Fold change from generalized linear model with log link (i.e., ratio of arithmetic means).
`"foldlog"` | Fold change from linear model after log transformation of the outcome (*i.e.*, ratio of geometric means).
`"irrrob"` | Fold change from generalized linear model with Poisson distribution and log link and robust (sandwich) standard errors
`"quantreg"` | Quantile difference from quantile regression using `quantreg::rq()` with `method = "fn"`. By default, this is the difference in medians. | `list(tau = 0.75)` to change the quantile to, e.g., the 75th percentile.
   

# More on ratios of continuous outcomes

Three types of ratios for continuous outcomes are implemented in `rifttable()`:

`type =` | Ratio of ... | Model | Use
------|----------|-----------------------|-----------------------------
`"fold"` | Arithmetic means | Generalized linear model with Gaussian distribution and log link: `glm(y ~ x, family = gaussian(link = "log"))` | Tolerates `0` in the `outcome` variable. May be informative if outcome is normally distributed without transformation.
`"irrrob"` | Arithmetic means | Generalized linear model with Poisson distribution and log link: `glm(y ~ x, family = poisson())`, with robust (sandwich) standard errors | Tolerates `0` in the `outcome` variable. May be informative if outcome is normally distributed without transformation.
`"foldlog"` | Geometric means | Linear model with log-transformed outcome: `lm(log(y) ~ x)` | Does not tolerate `0` in the `outcome` variable, as `log(0)` is undefined (R returns `-Inf`). May be informative if outcome is normally distributed after log transformation.

In all models, after exponentiation, beta coefficients can be interpreted as ratios. rifttable automatically does all necessary transformations.

In the `cancer` data, ratios of usual (arithmetic) means of age could be considered informative, given that `hist(cancer$age)` etc. does not show a major skew in this outcome.