Chapter 12 Provenance
The most important problem is that we are trying to understand the fundamental workings of the universe via a language devised for telling one another when the best fruit is.
— Terry Pratchett
We have now developed, automated, and tested a workflow for plotting the word count distribution for classic novels. In the normal course of events, we would include the outputs from that workflow (e.g., our figures and \(\alpha\) values) in a report. Here we use the term “report” to include research papers, summaries for clients, or anything else that is shorter than a book and aimed at people other than its creators.
But modern publishing involves much more than producing a printable PDF and making it available on a preprint server such as arXiv or bioRxiv. It also entails providing the data underpinning the report and the code used to do the analysis:
An article about computational science in a scientific publication is not the scholarship itself, it is merely advertising of the scholarship. The actual scholarship is the complete software development environment and the complete set of instructions which generated the figures.
— Jonathan Buckheit and David Donoho, paraphrasing Jon Claerbout, in Buckheit and Donoho (1995)
While some reports, datasets, software packages, and/or analysis scripts can’t be published without violating personal or commercial confidentiality, every researcher’s default should be to make all these as widely available as possible. Publishing it under an open license (Section 7.4) is the first step; the sections below describe what else we can do to capture the provenance of our data analysis.
Our Zipf’s Law project files are structured as they were at the end of the previous chapter:
zipf/ ├── .gitignore ├── CONDUCT.md ├── CONTRIBUTING.md ├── LICENSE.md ├── Makefile ├── README.md ├── requirements.txt ├── bin │ ├── book_summary.sh │ ├── collate.py │ ├── countwords.py │ ├── plotcounts.py │ ├── plotparams.yml │ ├── script_template.py │ └── utilities.py ├── data │ ├── README.md │ ├── dracula.txt │ └── ... ├── results │ ├── dracula.csv │ ├── dracula.png │ └── ... └── test_data ├── random_words.txt └── risk.txt
Identifying Reports and Authors
Before publishing anything, we need to understand how authors and their works are identified. A Digital Object Identifier (DOI) is a unique identifier for a particular version of a particular digital artifact such as a report, a dataset, or a piece of software. DOIs are written as
doi:prefix/suffix, but are often also represented as URLs like
http://dx.doi.org/prefix/suffix. In order to be allowed to issue a DOI, an academic journal, data archive, or other organiation must guarantee a certain level of security, longevity and access.
An ORCID is an Open Researcher and Contributor ID. Anyone can get an ORCID for free, and should include it in publications because people’s names and affiliations change over time.
12.1 Data Provenance
The first step in documenting the data associated with a report is to determine what (if anything) needs to be published. If the report involved the analysis of a publicly available dataset that is maintained and documented by a third party, it’s not necessary to publish a duplicate of the dataset: the report simply needs to document where to access the data and what version was analyzed. This is the case for our Zipf’s Law analysis, since the texts we analyze are available at Project Gutenberg.
It’s not strictly necessary to publish intermediate data
produced during the analysis of a publicly available dataset either
(e.g., the CSV files produced by
so long as readers have access to the original data and the code/software used to process it.
making intermediate data available can save people time and effort,
particularly if it takes a lot of computing power to reproduce it
(or if installing the required software is complicated).
NASA has published
the Goddard Institute for Space Studies Surface Temperature Analysis,
which estimates the global average surface temperature
based on thousands of land and ocean weather observations,
because a simple metric of global warming is expensive to produce
and is useful in many research contexts.
If a report involves a new dataset, such as observations collected during a field experiment, then that data needs to be published in its raw (unprocessed) form. The publication of a dataset, whether raw or intermediate, should follow the FAIR Principles.
12.1.1 The FAIR Principles
The FAIR Principles describe what research data should look like. They are still aspirational for most researchers, but tell us what to aim for (Goodman et al. 2014, @Mich2015, @Hart2016, @Broc2019, @Tier2020). The most immediately important elements of the FAIR Principles are outlined below.
220.127.116.11 Data should be findable.
The first step in using or re-using data is to find it. We can tell we’ve done this if:
- (Meta)data is assigned a globally unique and persistent identifier (i.e. a DOI).
- Data is described with rich metadata
- Metadata clearly and explicitly includes the identifier of the data it describes.
- (Meta)data is registered or indexed in a searchable resource, such as the data sharing platforms described in Section 12.1.2.
18.104.22.168 Data should be accessible.
People can’t use data if they don’t have access to it. In practice, this rule means the data should be openly accessible (the preferred solution) or that authenticating in order to view or download it should be free. We can tell we’ve done this if:
- (Meta)data is retrievable by its identifier using a standard communications protocol like HTTP.
- Metadata is accessible even when the data is no longer available.
22.214.171.124 Data should be interoperable.
Data usually needs to be integrated with other data, which means that tools need to be able to process it. We can tell we’ve done this if:
- (Meta)data uses a formal, accessible, shared, and broadly applicable language for knowledge representation.
- (Meta)data uses vocabularies that follow FAIR principles.
- (Meta)data includes qualified references to other (meta)data.
126.96.36.199 Data should be reusable.
This is the ultimate purpose of the FAIR Principles and much other work. We can tell we’ve done this if:
- Meta(data) is described with accurate and relevant attributes.
- (Meta)data is released with a clear and accessible data usage license.
- (Meta)data has detailed provenance.
- (Meta)data meets domain-relevant community standards.
12.1.2 Where to archive data
Small datasets (i.e., anything under 500 MB) can be stored in version control. If the data is being used in several projects, it may make sense to create one repository to hold only the data; the R community refers to these as data packages, and they are often accompanied by small scripts to clean up and query the data.
For medium-sized datasets (between 500 MB and 5 GB), it’s better to put the data on platforms like the Open Science Framework, Dryad, and Figshare, which will give the dataset a DOI. Big datasets (i.e., anything more than 5 GB) may not be ours in the first place, and probably need the attention of a professional archivist.
While archiving data at a site like Dryad or Figshare (following the FAIR Principles) is usually the end of the data publishing process, there is the option of publishing a journal paper to describe the dataset in detail. Some research disciplines have journals devoted to describing particular types of data (e.g., the Geoscience Data Journal) and there are also generic data journals (e.g., Scientific Data).
12.2 Code Provenance
Our Zipf’s Law analysis represents a typical data science project in that we’ve written some code (in the form of a series of Python scripts) that leverages other pre-existing software packages (matplotlib, numpy, etc) in order to produce the key results of a report (the word distribution plots). Documenting the details of a computational workflow like this in an open, transparent and reproducible manner typically requires that three key items are archived:
- A copy of any analysis scripts or notebooks used to produce the key results presented in the report.
- A detailed description of the software environment in which those analysis scripts or notebooks ran.
- A description of the data processing steps taken in producing each key result, i.e., a step-by-step account of which scripts were executed in what order for each key result.
Unfortunately, librarians, publishers, and regulatory bodies are still trying to determine the best way to document and archive material like this, so there is not yet anything like the FAIR Principles. The best advice we can give is presented below. It involves adding information about the software environment and data processing steps to a GitHub repository that contains the analysis scripts/notebooks, before creating a new release of that repository and archiving it (with a DOI) with Zenodo.
12.2.1 Software environment
In order to document the software packages used in our analysis, we should archive a list of the names and version numbers of each software package. We can get version information for the Python packages we are using by running:
$ pip freeze
alabaster==0.7.12 anaconda-client==1.7.2 anaconda-navigator==1.9.12 anaconda-project==0.8.3 appnope==0.1.0 appscript==1.0.1 asn1crypto==1.0.1 ...
Other command line tools will often have an option like
to access the version information.
Archiving a list of package names and version numbers would mean that our software environment is technically reproducible, but it would be left up to the reader of the report to figure out how to get all those packages installed and working together. This might be fine for a small number of packages with very few dependencies, but in more complex cases we probably want to make life easier for the reader (and for our future selves looking to re-run the analysis). One way to make things easier is to export a description of a complete conda environment (Section 13.2; Appendix I.2), which can be saved as YAML using:
$ conda env export > environment.yml $ cat environment.yml
name: base channels: - conda-forge - defaults dependencies: - _ipyw_jlab_nb_ext_conf=0.1.0=py37_0 - alabaster=0.7.12=py37_0 - anaconda=2019.10=py37_0 - anaconda-client=1.7.2=py37_0 - anaconda-navigator=1.9.12=py37_0 - anaconda-project=0.8.3=py_0 - appnope=0.1.0=py37_0 - appscript=1.1.0=py37h1de35cc_0 - asn1crypto=1.0.1=py37_0 ...
That software environment can be recreated on another computer with one line of code:
$ conda env create -f environment.yml
We can go ahead and add the
environment.yml file to our GitHub repository:
$ git add environment.yml $ git commit -m "Adding the conda environment file" $ git push origin master
More complex tools like Docker can install our entire environment (down to the precise version of the operating system) on a different computer (Nüst et al. 2020). However, their complexity can be daunting, and there is a lot of debate about how well (or whether) they actually make research more reproducible in practice.
12.2.2 Data processing steps
The second item that needs to be added to our GitHub repository is a description
of the data processing steps involved in each key result.
Assuming the author list on our report is Amira Khan and Sami Virtanen (Section 1.2),
we could add a new Markdown file called
KhanVirtanen2020.md to the repository
to describe the steps:
The code in this repository was used in generating the results for the following paper: Khan A and Virtanen S (2020). Zipf's Law in classic english texts. *Journal of Important Research*, 27, 134-139. `environment.yml`. The code was executed in the software environment described by [conda](https://docs.conda.io/en/latest/): It can be installed using $ conda env create -f environment.yml Figure 1 in the paper was created by running the following at the command line:$ make all
We should also add this information as an appendix to the report itself.
12.2.3 Analysis scripts
Later in this book we will package and release our Zipf’s Law code so that it can be downloaded and installed by the wider research community, just like any other Python package (Chapter 13). Doing this is especially helpful if other people might be interested in using and/or extending it, but often the scripts and notebooks we write to produce a particular figure or table are too case-specific to be of broad interest. To fully capture the provenance of the results presented in a report, these analysis scripts and/or notebooks (along with the details of the associated software environment and data processing steps) can be archived with a repository like Figshare or Zenodo, which specialize in storing the long tail of research projects (i.e., supplementary figures, data, and code). Uploading a zip file of analysis scripts to the repository is a valid option, but more recently the process has been streamlined via direct integration between GitHub and Zenodo. As described in this tutorial, the process involves creating a new release of our repository in GitHub that Zenodo copies and then issues a DOI for (Figure 12.1).
12.2.4 Reproducibility versus inspectability
In most cases, documenting our software environment, analysis scripts, and data processing steps will ensure that our computational analysis is reproducible/repeatable at the time our report is published. But what about five or ten years later? As we have discussed, data analysis workflows usually depend on a hierarchy of packages. Our Zipf’s Law analysis depends on a collection of Python libraries, which in turn depend on the Python language itself. Some workflows also depend critically on a particular operating system. Over time some of these dependencies will inevitably be updated or no longer supported, meaning our workflow will be documented but not reproducible.
Fortunately, most readers are not looking to exactly re-run a decade old analysis: they just want to be able to figure out what was run and what the important decisions were, which is sometimes referred to as inspectability (Gil et al. 2016; Brown 2017). While exact repeatability has a short shelf-life, inspectability is the enduring legacy of a well-documented computational analysis.
Your Future Self Will Thank You
Data and code provenance is often promoted for the good of people trying to reproduce your work, who were not part of creating the work in the first place. Prioritizing their needs can be difficult: how can we justify spending time for other people when our current projects need work for the good of the people working on them right now?
Instead of thinking about people who are unknown and unrelated, we can think about newcomers to our team and the time we will save ourselves in onboarding them. We can also think about the time we will save ourselves when we come back to this project five months or five years from now. Documentation that serves these two groups well will almost certainly serve the needs of strangers as well.
The Internet started a revolution in scientific publishing that shows no sign of ending. Where an inter-library loan once took weeks to arrive and data had to be transcribed from published papers (if it could be found at all), we can now download one another’s work in minutes: if we can find it and make sense of it. Organizations like Our Research are building tools to help with both; by using DOIs and ORCIDs, publishing on preprint servers, following the FAIR Principles, and documenting our workflow, we help ensure that everyone can pursue their ideas as we did.
If you don’t already have an ORCID, go to the website and register now. If you do have an ORCID, log in and make sure that your details and publication record are up-to-date.
12.4.2 A FAIR test
An online questionnaire for measuring the extent to which datasets are FAIR has been created by the Australian Research Data Commons. Fill in the questionnaire for a dataset you have published or that you use often.
12.4.3 Evaluate a project’s data provenance
This exercise is modified from Wickes and Stein (2016) and explores the dataset from Meili (2016). Go to the dataset’s page http://doi.org/10.3886/E17507V2 and download the files. You will need to make an ICPSER account and agree to their data agreement before you can download.
Review the dataset’s main page to get a sense of the study, then review the spreadsheet file and the coded response file.
- Who are the participants of this study?
- What types of data was collected and used for analysis?
- Can you find information on the demographics of the interviewees?
- This dataset is clearly in support of an article. What information can you find about it, and can you find a link to it?
12.4.4 Evaluate a project’s code provenance
- Where is the software environment described? What files would you need to recreate the software environment?
- Where are the data processing steps described? How could you re-create the results included in the manuscript?
- How are the scripts and data archived? I.e. Where can you download the version of the code and data as it was when the manuscript was published?
12.4.6 Create an archive of your Zipf’s analysis
A slightly less permanent alternative to having a DOI for your analysis code
is to provide a link to a GitHub release.
Follow the instructions on GitHub to create a release for the current state of your
Once you’ve created the release, read about how to link to it. What is the URL that allows direct download of the zip archive of your release?
What about getting a DOI?
Creating a GitHub release is also a neccessary step to get a DOI through the Zenodo/Github integration (Section 12.2.3). We are stopping short of getting the DOI here, to avoid many DOIs pointing to the same code, that is associated with different authors (you), and that isn’t associated with a publication
12.4.7 Publishing your code
Think about a project that you’re currently working on. How would you go about publishing the code associated with that project (i.e., the software description, analysis scripts, and data processing steps)?
12.5 Key Points
- Publish data and code as well as papers.
- Use DOIs to identify reports, datasets, or software release.
- Use an ORCID to identify yourself as an author of a report, dataset, or software release.
- Data should be FAIR: findable, accessible, interoperable, and reusable.
- Put small datasets in version control repositories; store large ones on data sharing sites.
- Describe your software environment, analysis scripts, and data processing steps in reproducible ways.
- Make your analyses inspectable as well as reproducible.