The goal of reproducible data analysis is, according to the CRAN task view, to tie specific instructions to data analysis and experimental data so that scholarship can be recreated, better understood and verified.
In order to prepare a small talk on this subject for the next Strasbourg’s R user group meeting, I’m starting a series of blog post. Today I will focus on the goals and constraints of my personal workflow.
Goal of my workflow
Beside some side projects, I use mainly R in an academic context. For this, I must be able to:
- Check my analysis and track logical errors.
- Share them to be enhanced or curated by colleagues.
- Send my code with papers. In order to struggle with the non-reproducible results crisis, this practice tends to become a standard.
- As an epidemiologist, my raw data may change (e.g.: due to some quality data check or because a new year of data is available). Then I need to regenerate my reports with the new data.
Furthermore I have some constraints.
- My workflow has to be platform agnostic. For academic works, I use various Linux distributions on my workstation and some remote machines. My hospital’s computer with non-anonymous data run exclusively on Windows and I don’t have root access (then limited choice of my software tools). My personal setting is a mix between macOS and Linux boxes.
- For the same reason, it must be easy to deploy. I don’t want to spend hours setting up a virtual or remote machine. Furthermore, peoples who wants to reproduce my analysis don’t have either the time or the ressources to install some complicated software suite.
- Free open source software (FOSS) is mandatory. Peoples who needs to run my analysis don’t want to pay for software, academics are not rich. Furthermore, open-source software is the only way to control really what the software does. The reproducibility is unattainable with close-source software black boxes.
- Input must be the real raw data, whatever the format is (including csv, Microsoft Excel and Access files, SAS data or direct connection to RDBMS).
- Output could be html files (intermediate reports), PDF (for printing) or web app (Shiny).
- It must be IDE agnostic. I love Rstudio but I have to work on headless remote machines and I don’t want to set up an Rstudio server each time. Script should be launch simply by a
R CMDand edited in vim (or nvim).
Summary of the implementation
With these goals and constraints, R is a natural choice:
- R is light (± 70 Mo) and easy to install compared to Python with data science modules or SAS with its awkward licence checking system.
- R’s package system give me all the tools needed, on every OS.
- R is FOSS.
- R tends to become the lingua franca in data science and statistics.
- Rstudio is a tremendous IDE, also easy to deploy
After several years of R practice, I developed a simple workflow based on
- R with several packages: rmarkdown, knitr
- Coding style
This is basic, doesn’t cover all reproductibility problems (e.g. no software archive) but is for me a good balance.
If you have your own reproducible data analysis workflow, please feel free to describe it in the commentaries or send me a link!