Reproducible research in public health is an operations problem

What I mean by reproducibility when the data still come from the field

Public Health
Research Methods
Reproducibility
Field Systems
Author

Nichodemus Amollo

Published

October 26, 2025

In public health, “reproducible research” is often introduced with crisis statistics and lab metaphors. Those debates matter. In the work I do, the more urgent version is plainer:

Can another analyst on the team re-run the quality checks and the core tables next month without calling me?

If the answer is no, the study is already at risk — especially when data are still streaming from the field.

What reproducibility means on a field study

For household diaries, multi-country evaluations, or clinical chart abstraction, I treat reproducibility as four concrete things:

  1. Raw is sacred — immutable extracts or locked REDCap exports with dates and versions
  2. Cleaning is code — R or Stata scripts with named rules, not a one-off Excel “final_v7”
  3. HFC is scheduled — high-frequency checks run on a rhythm, with logs of what failed
  4. Outputs are rebuildable — Quarto/R Markdown (or equivalent) can regenerate tables and briefs

That is closer to research data management than to a philosophy seminar.

Where studies actually break

Failure mode What I have seen Mitigation
Analyst turnover Only one person knows the “real” cleaning steps Documented scripts + handoff notes
Wave panic QC postponed until closeout Daily/weekly HFC while field is live
Silent schema drift New form versions break old code Versioned instruments + schema checks
Unowned indicators Dashboard numbers nobody can recompute Indicator dictionary + tested code

On the health financial diaries work, live ops apps and HFCs existed precisely so quality debt did not wait for the end of a year-long follow-up.

A minimum bar for health and development teams

If a PI or programme lead asks whether the work is reproducible, I look for:

  • A project folder with raw/, scripts/, outputs/ (names can vary; structure cannot)
  • A short README: how to refresh HFC and main tables
  • Named responsibility for the database and for analysis
  • No patient or household identifiers in public repos

I do not require perfect packaging on day one. I require that the next person is not starting from mythology.