Word is better than Quarto?

NHS RPySOC 2025

Jacqueline Grout

Matt Dray

13 November 2025

Hello 👋

• We’re from The Strategy Unit
• I am Jacqueline
• This is Matt

• From the website: ‘The Strategy Unit is a specialist NHS internal consultancy team. We produce high-quality, multi-disciplinary analytical work – and we help people apply the results. Our proposition is simple: better evidence, better decisions, better outcomes.’
• We have different skills that were both needed on this project and could learn a lot from each other.

The New Hospital Programme 🏥

• The government is committed to building hospitals
• How to meet needs/get good value for money?
• Answer: we built a demand model

• The government has committed to building new hospitals.
• New hospital infrastructure must meet the future needs of the population and have good value for money.
• NHP is a DHSC/NHSE partnership.
• The Strategy Unit is helping by building and maintaining a demand model.
• The Strategy Unit supports end-users to create model scenarios and and interpret the outputs.

Reporting 📒

• End-users get model outputs and we co-author a report
• Requires automation/reproducibility
• So use Quarto, right? Right?

• End-users are ‘schemes’, usually hospitals.
• Once they’ve completed modelling, they are given a draft report containing model outputs and calculations.
• It’s a standard template report, so it would be wise to automate its generation.
• We should be able to recreate these documents in future, if needed.
• Ideal solution: create a Quarto document and populate it for a named organisation with images and calculations derived from data read from Azure.
• In other words, an obvious Reproducible Analytical Pipeline (RAP).

Existing Word template 📑

• However, the pre-existing Word template was complicated and was frequently being changed as it was passed between the SU and the scheme.
• It turned out to be easier to use the Word template as an input and insert our figures and values into it instead.
• This zoomed-out view of part of the Word document template highlights the complexity.
• The document uses a pre-existing theme; contains many specifically-formatted text chunks, tables and page orientations; and uses colour-coded highlighting to indicate who needs to edit which parts of the document.
• The document was being changed and updated quite a lot. The customers were used to amending Word documents quickly, as and when required and did not have skills or software to work with R/Quarto.

Existing Word template 📑

It looks like you’ve accepted your fate.

Decision 🧐

• Meet users where they are: in Word
• Editability and flexibility trumps total automation
• How can we maximise reproducibility?

Doing it

Process ✏️

1. Record a request with a GitHub issue template.
2. Tag model-run JSONs on Azure, update site-selections file.
3. Read SharePoint template with {Microsoft365R}.
4. Insert content with {officer}.
5. Write timestamped folder with doc, results, log.
6. Check output, return to recipient.

• We designed a process to programmatically generate the reports.
• How do we identify chosen runs? On Azure, with run_stage metadata on results files and a list of chosen sites.
• How to confirm with model relationship managers? Expose tagged scenarios and sites via a scheduled Quarto-report.
• Basic interface: supply a scheme code, at simplest level.
• Underlying functions save figures and values independently.
• Images inserted by moving the imaginary ‘cursor’ to a known string of text in the report.
• Value-insertion is more complicated: insert ‘field’ codes into the report, insert the values to custom document properties, refresh to insert these into the fields.
• The output is a timestamped directory with the report, standalone images and values, and a log.

Issue template ✅

• We have an issue template in the repo that allows us to assign and manage tasks for creating a new report.
• This makes sure we follow the same steps from receiving a request through to returning a populated report.
• We use it to reference further guidance and links, which should help improve our ‘bus factor’: other members of the SU should be able to run a new report without us.

Generate a report 🖨️

In R:

populate_template(
  scheme_code = "XYZ",                 # create a report for this scheme
  site_codes = NULL,                   # NULL to fetch from Azure
  result_sets = get_nhp_result_sets(), # fetch results metadata from Azure
  run_stages = list(                   # list the tagged model runs
    primary = "final_report_ndg3",     # main data source for the report
    secondary = "final_report_ndg2"    # used as a comparator
  )
)

• The interface for us is quite simple: a single function where we set the scheme code and identify the metadata ‘tag’ for the results that will form the main basis of the report and a secondary set used for comparative purposes.
• Alternatively we can provide a scenario filename directly using the parameter scenario_files.
• An option also exists to choose different templates.
• A bunch of sub-functions then read the data, calculate values, generate figures, read the Word template from SharePoint and then insert the content.

Generated report 📁️

Output file structure:

output/
└──YYYY-MM-DD-HHMMSS_scheme/
   ├──YYYY-MM-DD-HHMMSS_scheme.log                        # log of printed metadata
   ├──YYYY-MM-DD-HHMMSS_scheme_outputs-report_draft.docx  # populated report
   ├──figures/                                            # standalone PNG files
   └──values/                                             # CSVs of calculated values
    

• The output is a timestamped folder that contains the populated Word doc, subfolders of the charts and calculated values, and a log file that contains information printed to the console (details things like the scheme name and site selections for the run, as well as warning and error messages if present).

Insert images with {officer} 👮

1. Add unique target strings, e.g. [Insert Figure 9.2].
2. Find/replace with cursor_reach() and body_add_img().

• The {officer} package is crucial for inserting our content into the template.
• We use different methods to insert values and images.
• To insert images, programmatically find unique target strings with {officer}’s cursor_reach() and then overwrite with images using body_add_img(); we use and a custom insert_figure_on_cursor() to do this.

Insert values with {officer} 👮

1. Add doc-property fields, e.g.

   { DOCPROPERTY item_54 }.

2. Add custom doc properties.
3. Update values programmatically with set_doc_properties().
4. Refresh to insert values to fields.

• We need something smarter for inserting values.
• Images have obvious places for insertion in the document, but calculated values could be in the middle of sentences that get edited later.
• We can instead put placeholders into the document that reference a given value.
• To do this, first add ‘document-property fields’ to the Word doc (ctrl+F9); then write your calculated values to the custom document properties with officer::set_doc_properties(); then open and refresh the document (F9) to replace the fields with the corresponding value for the named property.

Reflecting on it

Doing it right 💯

• Github: ‘learn by doing’, challenge and improve things
• Documentation/version history have saved us!
• The repo is a springboard for parallel/future work

Flexibility 🤸

The repository and code continue to help with:

• ad hoc work
• upcoming ‘refresh’ reports
• changes to inputs and back-end processes

• The functions created for this purpose have served as useful tools for related work that doesn’t involve report generation.
• The approach is flexible enough to allow for other types of template report to be included.
• Changes to the template and model, or the way we store results files, can be handled easily and shared between us with version-control and good documentation.

Reflections 🪞

• Challenge your thinking, be adaptable to user needs
• Embrace agility and acknowledge fragility
• ‘The power of friendship’ keeps things working

• The default mode of ‘make a Quarto doc’ wasn’t the right fit here, but we recognised that and pivoted quickly.
• We know the solution isn’t perfect, but it is good enough.
• The process has some frailties, but we recognise them. For example, the report path being changed on SharePoint means we can’t retrieve it!
• Talking between us on the code side, but also with the user, is really important to handle changing needs and to keep the process and outputs fit for purpose.

Actions 🎬

1. Consider user needs.
2. Collaborate with GitHub.
3. Stay flexible.

Check out:

• these slides and open code
• {officer}, the officeverse book and {Microsoft365R}