Reproducible analysis

Last updated on 2024-11-19 | Edit this page

Estimated time: 60 minutes

Overview

Questions

  • How do I make my research project reproducible?
  • How do I include packages as dependencies of my project?
  • What features are related to reproducible research?

Objectives

  • Add dependencies of a project using the DESCRIPTION file.
  • Create an isolated and specific reproducible environment with renv
  • Identify your project features related to reproducible software.

How do I make my analysis reproducible?

The reproducible environment

Any analysis with R needs packages. These packages on which your project relies are called dependencies. To make an analysis reproducible, we need to register these packages (and their versions) somewhere as your project’s dependencies. That place is the DESCRIPTION file.

In the DESCRIPTION file, dependencies are registered at the end of the file with the package names only and usually with a minimum version (dplyr (>= 1.0.0)). We can add dependencies using functions (rcompendium::add_dependencies()), and also use this file to automate version recovery (devtools::install_deps()). However, DESCRIPTION files are most useful for R packages.

For non-package projects we can use renv. It registers specific dependencies by implementing project-specific environments, which means that renv registers even the SHA/hash from GitHub packages, feature that the DESCRIPTION file can not do. Also, renv isolates your project packages from your computer packages. Lastly, renv can detect new dependencies automatically, apart from adding them with functions (renv::snapshot()), and it can also automate the recovery of the whole project (renv::restore()).

Callout

The renv package:

  • Isolates the dependencies of your project from your computer.
  • Registers the specific version of packages from CRAN or GitHub.
  • Provides an automated package management solution to restore an external project.

The analysis workflow

Complementary to the dependencies, your analysis workflow must follow some good practices in scientific computing.

First, for Data management, we need to save input data as originally created and, preferably, configure it as a read-only file. In your project, you can differentiate raw-data from derived-data

Second, for Project organization, we need to store analysis and generated files in specific and isolated folders. In your project, you can differentiate analyses files (like .R scripts and .Rmd files) from figures and other outputs.

Automate your analysis

The make.R file helps automate your analysis project. This file includes a script line to automatically restore your dependencies (renv::restore()) and run all the analysis scripts in your preferred order. The make.R file is the only .R file stored in the project’s root given by the rcompendium template. You can use the make.R file as the only script to run and regenerate all your project outputs.

Callout

The make.R file is inspired but not equivalent to GNU Make file.

GNU Make files can identify out-of-date files and re-execute any downstream code that needs to be updated, usually used for bash scripts.

To use this functionality for your R project, you can use the {targets} package.

Let’s code

We need to play under the rules of the rcompendium template.

The reproducible environment

We will use renv instead of DESCRIPTION files for this.

Usually, to initiate a reproducible environment with renv, we need to run renv::init().

Source: https://rstudio.github.io/renv/
Source: https://rstudio.github.io/renv/

However, when working in a rcompendium template, your first step must be to run:

R 

rcompendium::add_renv()

OUTPUT 

This project contains a DESCRIPTION file.
Which files should renv use for dependency discovery in this project?

1: Use only the DESCRIPTION file. (explicit mode)
2: Use all files in this project. (implicit mode)

Write 2 and press ENTER to use renv instead of DESCRIPTION file.

Question

Why not to use renv in addition to DESCRIPTION?

We can use renv in addition to DESCRIPTION.

However, we opt to use renv instead of DESCRIPTION because the rcompendium::add_dependencies(".") function because it assumes that all packages to add to DESCRIPTION are from CRAN. If you want to add GitHub packages, you need to add them manually in a different section called Remotes: and write repository/package. The renv package solves this automatically.

We need to fix the DESCRIPTION file manually. Packages like {cfr} and {epiparameter} are on GitHub.
We need to fix the DESCRIPTION file manually. Packages like {cfr} and {epiparameter} are on GitHub.

However, this still needs to be assessed with different scenarios to confirm this as the final best decision.

If you decide to use renv in addition to DESCRIPTION run:

R 

rcompendium::add_dependencies(".")

Note that this function requires one argument specification ".", which means that your working directory must be at the root of the R project.

The output below details which packages were included in the description file

✔ Scanning 'Imports' dependencies
  (*) Found 2 package(s)
  (*) Adding the following line in 'DESCRIPTION': `Imports: devtools, here`

If you get an error message like:

ERROR 

Error in renv_snapshot_validate_report(valid, prompt, force) :
  aborting snapshot due to pre-flight validation failure

Run again the rcompendium::add_renv() function. You may get the following message:

OUTPUT 

This project already has a private library. What would you like to do?

1: Activate the project and use the existing library.
2: Re-initialize the project with a new library.
3: Abort project initialization.

Write option 1 and press ENTER.

This step creates a renv/ folder and modifies the content of the make.R in line 15, replacing the default devtools::install_deps() by renv::restore.

Second, to get the status of the project run:

R 

renv::status()

OUTPUT 

This project does not contain a lockfile.
Use renv::snapshot() to create a lockfile.

Callout

Always follow the suggestions of the renv::status() output. You can also get a message from it each time you reopen your project.

Third, to create the lockfile run:

R 

renv::snapshot()

This step creates a renv.lock file detailing the following:

  • R version on top and
  • specific version details of all the packages in the project’s dependency tree (including SHA/hash for GitHub packages).
{
  "R": {
    "Version": "4.2.2",
    "Repositories": [
      {
        "Name": "CRAN",
        "URL": "https://packagemanager.posit.co/cran/latest"
      }
    ]
  },
  "Packages": {
    "R6": {
      "Package": "R6",
      "Version": "2.5.1",
      "Source": "Repository",
      "Repository": "RSPM",
      "Requirements": [
        "R"
      ],
      "Hash": "470851b6d5d0ac559e9d01bb352b4021"
    },
    ...

Now, you have completed your reproducible environment configuration.

The analysis workflow

The workflow will follow these three paths:

  • Read raw-data/ to clean.R it and save it to derived-data/.
  • Read derived-data/ to make a plot.R and save it to figures/.
  • Read derived-data/ to make a table.R and save it to outputs/.

First, download the sample data set.

Since this is raw data, save it in the data/raw-data/ folder.

Second, create the analysis script to clean this raw data set. Name it 01-clean.R. Save it in the analyses/ folder. Copy and paste these lines of code:

R 

# Load packages
library(readxl)
library(tidyverse)

# Read raw data
dat <- readxl::read_xlsx("data/raw-data/linelist_20140701.xlsx")

# Clean raw data
dat_clean <- dat %>% 
  select(case_id,date_of_onset,date_of_outcome,outcome) %>% 
  mutate(across(.cols = c(date_of_onset,date_of_outcome),
                .fns = as.Date)) %>% 
  mutate(outcome = fct(outcome,level = c("Death","Recover"),na = "NA"))

# Write clean data
dat_clean %>% 
  write_rds("data/derived-data/linelist_clean.rds")

Notice that we are writing a new cleaned data set in a different path: data/derived-data/.

Callout

  • The default folder to save R scripts will be R/. This path is the place to write your Modular functions. Go to the analyses/ folder to save your analysis script.

  • Yes, it is named analyses/ not “analysis”.

Rstudio will invite you to install new packages. Press Install. Always run renv::status() after installing new packages:

R 

renv::status()

OUTPUT 

The following package(s) are in an inconsistent state:

 package       installed recorded used
 backports     y         n        y
 bit           y         n        y

In this case, we need to follow the instructions in the section of Missing packages from the ?renv::status() documentation.

R 

renv::install()

OUTPUT 

- There are no packages to install.
- Automatic snapshot has updated '~/0projects/projectname/renv.lock'.

Third, create an analysis script to create an incidence plot for this cleaned data set. Name it 02-plot.R. Save it in the analyses/ folder. Copy and paste these lines of code:

R 

# Load packages
library(tidyverse)
library(incidence2)

# Read data
ebola_dat <- read_rds("data/derived-data/linelist_clean.rds")

# Create incidence2 object
ebola_onset <- 
  incidence2::incidence(
    x = ebola_dat,
    date_index = c("date_of_onset"),
    interval = "epiweek"
  )
  
# Read incidence2 object
ebola_onset

# Plot incidence data
plot(ebola_onset)

# Write ggplot as figure
ggsave("figures/02-plot_incidence.png",height = 3,width = 5)

Notice that we are writing the new plot in a different path: figures/.

Challenge

  • Explore the i2extras::fit_curve() to fit a model to the incidence curve.
  • Save the output table in the corresponding folder.
  • You can reuse the incidence2 object as input in the same file.
  • Remember to update the renv status if you need to install and use a new package for this task

Automate your analysis

The easiest step to forget!

Lastly, list all .R scripts and .Rmd in a sequential order in the make.R file after line 32:

R 

## Run Project ----

# List all R scripts in a sequential order and using the following form:
# source(here::here("analyses", "script_X.R"))

source(here::here("analyses", "01-clean.R"))
source(here::here("analyses", "02-plot.R"))

Checklist

Reproducible research features

We defined Reproducible research as a practice that wants to ensure that anyone with access to data inputs and software can feasibly generate the outputs to check or build on them.

A key feature of this practice is the combination of renv with the make.R file. With this file, and any other more sophisticated alternatives like GNU Make or targets, we are sure that we:

  • Can feasibly regenerate the outputs.
  • Can inform about the reliability of the project.
  • Have an isolated time-proof capsule of dependencies.

Key Points

  • A dependency is a package that your project needs to run.

  • Use the DESCRIPTION file to register your project dependencies.

  • Use renv to isolate and create package-specific reproducible environments for your dependencies.

  • Use the folder template to differentiate your raw-data/ and derived-data/.

  • Save analysis and generated files in isolated folders like analyses/, figures/, and outputs/.

  • Use the make.R to list your analysis scripts and facilitate the regeneration of all your outputs.

  • Reproducible environments and Make files are features related to Reproducible research.