Workflow for data analysis (Greg)

Greg's notes on how data moves from collection, to the filesystem, then through the data analysis process.

See also: General programming page.

Filesystem

  • ~/data/ - on its own partition, regularly backed up
  • current/ - contains directories, organized by project, storing metadata, initial field/lab measurements (usually in spreadsheet form), reports, papers in progress, etc. Each project directory may also contain a data_analysis directory storing data analysis functions and scripts, symlinks to the project's rawdata directory, and any other processed data needed for data analysis. THIS DIRECTORY MUST HAVE VERSIONING

    • project_1/
    • Lab and field spreadsheets...
    • Maps of plot locations, etc
    • Documentation files for dataloggers, etc
    • Papers in progress
    • data_analysis/
      • symlink to ../rawdata/project_1/
      • m/
      • matlab scripts and functions
      • py/
      • python scripts and functions
      • processed_data/
      • textfiles with (for example) data exported from spreadsheets in parent directory, data from processing scripts, plot locations converted to UTM (from map), etc.
    • project_2/
    • ...
    • data_analysis/
      • symlink to ../rawdata/project_2/
      • m/
      • ...
    • archive/ - same as above but for projects that are completed/not maintained
    • oldproject_1/
    • rawdata - contains directories for each project, typically storing raw data downloads, datalogger files, or data that has been minimally processed from the initial measurements. This data is usually large in size, is suitable for sharing among many projects, or comprises some contextual information not directly measured in the current project (environmental monitoring data for example).
    • project_1/
    • textfile1 (long-term climate measurments from a nearby weather station)
    • textfile2 (datalogger .dat file)
    • project_2/
    • ...

General Workflow

This refers to current projects, each of which is stored in a data/current/projectname/ directory.

Data collected in the field or lab

  • In general, this data must be transcribed from field sheets, downloaded from instruments, formatted, etc, and this is most often done using spreadsheets or text files.
  • Some formatting, editing, or summary creation of textfiles can be done with shell scripts, awk, etc, but be sure to document and save the scripts in the appropriate projectname/data_analysis/ directory.
  • Processed data can be exported as a textfile to the projectname/data_analysis/processed_data/ directory, or put in the rawdata/ directory (if it fits criteria above).
  • Data is then processed and analyzed with python, matlab, or other scripts located in the projectname/data_analysis/ directory
  • This analysis generates outputs(datafiles, figures) that can be stored in projectname/data_analysis/processed_data/ for further use in analysis, or in the projectname/ directory, the wiki, or elsewhere for interpretation, publication, etc.

The generalized workflow looks like this:

Field/Lab measurements => projectname/ => Process to text => (QC/data munging =>) projectname/processed_data/ => Analysis with matlab/python/r (calculation, summary data, plots) => projectname/processed_data/ (OR projectname/, papers, wiki, reports, etc for finished products)

Data from elsewhere

  • Often this is downloaded as text data (ie, from SNOTEL, etc) in (hopefully) quality-checked form. However some munging may be necessary to trim, format, or otherwise make the data easier to use
  • Probably use shell scripts, awk, etc, and again, be sure to document and save the scripts in the appropriate projectname/data_analysis directory.
  • Data can then be directly placed in the ~/data/rawdata/projectname/ directory
  • Follow steps 3 & 4, as above.

The generalized workflow looks like this:

Download data => (QC/Data munging =>) rawdata/projectname/ => Analysis with matlab/python/r (calculation, summary data, plots) => /projectname/data_analysis/processed_data (OR /projectname/, wiki, paper, etc for finished products)

Coding and version control

Code pieces

In terms of actual pieces of code, there are generally several types that recur in each project.

  • a loading function
  • a data cleaning function
  • the function(s) that actually do the analysis (calculating, transforming, selecting data)
  • the script(s) that produce reports or plots using the analysis functions above

In practice, this means that in each data analysis project folder, there will be (at least) one version of each of these types of functions. Depending on the language in use and the maturity of the project, there may be many function files and reporting/plotting scripts. See more discussion of about these types of issues here, here, here and here

Versioning

Within each project folder in data/data_analysis there is a versioning repository (.hg or .git). These repositories are initialized early in the development of the project and commits are made to the repository when substantial changes are made to the codebase. For example, if a new plotting script is written one day, it should be added and committed to the repository. If a function is changed to use a different equation or constant, the change should be committed. Branching is also possible for testing new analyses or generating publication subsets of the analysis, and these branches can later be merged into the main, or kept separate. In most cases, data is not versioned, it is stored in the rawdata directory and changes to that directory are logged in a README file.

Versioning is currently done with mercurial, but git has a page here too.