Skip to content

Scripted actions

The most common kind of action is a scripted action.

Generally-speaking, you can write whatever code you like as long as it will run successfully on server, and it is possible to test this locally.

However, note the following restrictions and guidance:

  • Write analyses in Python, R, or Stata. You can can use more than one language in a single project if necessary. You can find more information about the available libraries below.

  • Do not write code that requires an internet connection to run. Any research objects (datasets, libraries, etc.) that are retrieved via the internet should be imported to the repo locally first. If this is not possible (for instance if the object size is too large to be transferred via GitHub) then get in touch.

  • Avoid code that consumes a lot of time or memory. The server is not an infinite resource. We can advise on code optimisation if run-times become problematic. A good strategy is to split your processing into separate project pipeline actions; the job runner can then choose to run them in parallel if sufficient resources are available.

  • Write code that runs in the OpenSAFELY platform. Code will be run within a Linux-based Docker environment. In practice, this just means ensuring you use forward-slashes / for directories.

  • Structure your code into discrete chunks, both within scripts, and by splitting into different pipeline actions. This helps with:

    • readability
    • bug-finding
    • parallelisation via the project pipeline

Reading and Writing Outputs🔗

Scripted actions can read and write output files that are saved in the workspace. These generally fall into two categories: * large pseudonymised patient-level files of highly_sensitive data for use by other actions * smaller moderately_sensitive aggregated patient-data (this should never be patient-level data) files for review and release

Large highly_sensitive output files🔗

Outputs labelled highly_sensitive will not be visible to researchers. This is a deliberate design feature of OpenSAFELY, intended to reduce the risk of disclosure of sensitive information. Outputs should always be classed as highly_sensitive if they are:

  • Pseudonymised patient-level outputs derived from queries run against Level 1 and 2 data, i.e., a specific study dataset generated by a dataset definition.
  • Pseudonymised patient-level intermediate outputs for a study derived from queries run against Level 3 data which output pseudonymised patient-level data i.e., a processed study dataset with certain filters/formatting applied.

These types of outputs are considered potentially highly-disclosive, should not be pushed to Level 4, and are never intended for publishing outside the secure environment.

Pseudonymised patient-level outputs tend to be large in size and therefore it is important that the right files formats are used for these large data files. The wrong formats can waste disk space, execution time, and server memory. The specific formats used vary with language ecosystem, but they should always be compressed.

Note

The template sets up the ehrql command to produce csv.gz outputs. This is the current recommended output format, as CSV files compress well, and this reduces both storage requirements and improves job execution times on the backend.

If you need to view the raw CSV data locally, you can unzip with opensafely unzip dataset.csv.gz.

# read compressed CSV output from ehrql
pd.read_csv("output/dataset.csv.gz")

# write compressed feather file
df.to_feather("output/model.feather", compression="zstd")

# read feather file, decompressed automatically
pd.read_feather("output/dataset.feather")
# read compressed CSV output from ehrql
df <- readr::read_csv("output/dataset.csv.gz")

# write a compressed feather file
arrow::write_feather(df, "output/model.feather", compression = "zstd")

# read a feather file, decompressed automatically
df <- arrow::read_feather("output/dataset.feather")
// stata cannot handle compressed CSV files directly, so unzip first to a plain CSV file
// the unzipped file will be discarded when the action finishes.
!gunzip output/dataset.csv.gz
// now import the uncompressed CSV using delimited
import delimited using output/dataset.csv

// save in compressed dta.gz format
gzsave output/model.dta.gz

// load a compressed .dta.gz file
gzload output/dataset.dta.gz

Smaller moderately_sensitive output files🔗

Files that are labelled moderately_sensitive should only ever be aggregated data such as summary tables, images, and the outputs from statistical models. These files and will be available to view with Level 4 access. These (and the corresponding automatically created log files of each action/script) will be the only output files that users will have access to; users do not have unfettered access to any patient-level data and only see aggregated outputs derived from their analysis code, which satisfies the GDPR principle of confidentiality.

File type restrictions for moderately_sensitive outputs🔗

There are restrictions on the type of file that are transferred to Level 4. This is to reduce the risk of making pseudonymised patient-level data available for researchers to view.

If a file labelled as moderately_sensitive does not meet the below allowed file types, it will be replaced on Level 4 with a .txt file with the same filename, which explains why the file was not allowed on Level 4.

File format

These are restricted to types of file that are likely to contain summary data, rather than patient-level data, and so reviewers can properly examine the outputs on the secure server.

Type Formats
Text .txt, .log, .md
Data .csv, .tsv, .json
Images .png, .jpeg, .svgz
Reports .html, .pdf

File size

There is a maximum file size of 16 MB to:

  • prevent large patient-level data files being accessed via Level 4
  • allow a thorough review of the outputs in a reasonable time

Files with patient_id in the header

Any CSV file with a patient_id header will not be made available in level 4.

Execution environments🔗

OpenSAFELY currently supports Stata, Python, and R for statistical analysis.

For security reasons, available libraries are restricted to those provided by the framework, though you can request additions.

The framework executes your scripts using Docker images which have been preloaded with a fixed set of libraries. These Docker images have yet to be optimised; if you have skills in creating Dockerfiles and would like to help, get in touch!

Stata🔗

We currently package version 16.1, with datacheck, safetab, and safecount libraries installed; when installed, new libraries will appear in the stata-docker GitHub repository.

As Stata is a commercial product, a license key is needed to use it.

If you are a member of the opensafely GitHub organisation🔗

  • If you are using Windows, you must have the Github Desktop app installed and be logged into it. Then the opensafely command line software will use that app to obtain the OpenSAFELY Stata license automatically.
  • If you are using macOS:
  • Download and install GitHub's command-line tool (gh)
  • Run gh auth login --web. Select the "HTTPS" option, and follow the instructions
  • The opensafely command line software will now automatically use the OpenSAFELY Stata license

All other external users🔗

If you are not a member of the opensafely GitHub organisation, you must provide your own Stata/MP license. Unfortunately other Stata flavours are not yet supported; let us know if this is a problem.

  1. Locate your Stata license string as follows: Locate a text file, called STATA.LIC (on Windows) or stata.lic (macOS and Linux) which is usually at the top level of the folder of your Stata installation:

    • On Windows machine it's usually somewhere like C:\Program Files\Stata17
    • On Linux, somewhere like /usr/local/stata17/
    • On macOS it's usually in /Applications/Stata/
    • Within that file, locate a license string of the format SerialNumber!Code!Authorization!User!Organisation!VersionCode.
    • Set it as an environment variable using a method appropriate to your operating system. The name of the environment variable should be STATA_LICENSE, and its contents should be the entire license string. The opensafely command line software should now automatically use this Stata license after opening a new terminal.

Python🔗

The Docker image provided is Python 3.8, with this list of packages installed.

R🔗

The R image provided is R 4.0, with this list of libraries installed.