---
title: "Which Function Should I Use?"
description: "Match the outcome and the study design to the right gtregression function and effect measure."
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Which Function Should I Use?}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

Start from two questions: **what kind of outcome do you have**, and **how were
the observations collected**? The tables below point to the function, the
`approach` where one is needed, and the effect measure that appears in the
table. Every function shares the same arguments for exposures, adjustment
variables, formatting and export.

## Binary outcomes

An outcome with two values, such as disease or no disease.

| Question | Function | `approach` | Reported as |
|---|---|---|---|
| Odds of the outcome | `uni_reg()`, `multi_reg()` | `"logit"` | Odds ratio |
| Odds with sparse data or separation | `uni_reg()`, `multi_reg()` | `"firth"` | Odds ratio (Firth) |
| Risk of the outcome, relative | `uni_reg()`, `multi_reg()` | `"logbinomial"` or `"robpoisson"` | Risk ratio |
| Risk of the outcome, absolute (two groups) | `risk_difference()` | none | Risk difference, NNTB/NNTH |
| Matched case-control sets | `conditional_logit_reg()` | none | Matched odds ratio |
| Clustered observations (clinics, schools) | `multilevel_reg()` | `"logit"` | Odds ratio, random intercept |

`"robpoisson"` (Poisson with robust standard errors) is a dependable
alternative when a log-binomial model fails to converge.

## Counts and rates

| Question | Function | `approach` | Reported as |
|---|---|---|---|
| Event counts | `uni_reg()`, `multi_reg()` | `"poisson"` | Incidence rate ratio |
| Over-dispersed counts | `uni_reg()`, `multi_reg()` | `"negbin"` | Incidence rate ratio |

## Continuous outcomes

| Question | Function | `approach` | Reported as |
|---|---|---|---|
| Mean of the outcome | `uni_reg()`, `multi_reg()` | `"linear"` | Beta (difference in means) |
| Clustered observations | `multilevel_reg()` | `"linear"` | Beta, random intercept |

## Outcomes with three or more categories

| Question | Function | Reported as |
|---|---|---|
| Ordered categories (mild, moderate, severe) | `ordinal_reg()` | Common odds ratio (proportional odds) |
| Unordered categories | `multinomial_reg()` | Odds ratio for each category versus the reference category |

## Time to event

| Question | Function | Reported as |
|---|---|---|
| Describe survival over time | `km_plot()`, `survival_summary()`, `survival_prob()` | Kaplan-Meier estimates |
| Compare survival curves | `logrank_test()`, `rmst_table()` | Log-rank p-value, restricted mean survival time |
| Hazard of the event | `cox_reg()` | Hazard ratio |
| Survival time itself | `surv_reg()` | Time ratio (accelerated failure time) |
| One event type in the presence of others | `cuminc_plot()`, `competing_risk_reg()` | Cumulative incidence, subdistribution hazard ratio |

Check the proportional-hazards assumption of a Cox model with `check_ph()`.

## Repeating an analysis within groups

`stratified_uni_reg()` and `stratified_multi_reg()` fit the same model within
each level of a stratifier, for example by sex or study site.

## Beyond a single estimate

| Question | Function |
|---|---|
| Is a variable a confounder or effect modifier? | `identify_confounder()`, `interaction_models()` |
| How much of an effect acts through a mediator? | `mediation_analysis()` |
| Does the model fit and converge? | `check_convergence()`, `check_collinearity()`, `plot_model_fit()` |
| Which of several prespecified models fits best? | `compare_models()`, `select_models()` |

## Next steps

- [Start Here](gtregression-intro.html) walks through a complete analysis.
- [Function Options](function-options.html) lists every argument and default.
- [Merge and Customise Tables](merge-tables.html) combines the results into a
  manuscript table.