Imagine giving your dataset and your paper to another epidemiologist.
– They have the same research question.
– They have the same raw data.
– They know which statistical software you used.
Could they reconstruct the analytical dataset and reproduce your results?
Not necessarily.
Reproducibility can break long before the final statistical model.
Consider what happens between the raw database and the numbers reported in a paper:
– Which patients were excluded?
– How were missing values handled?
– How was an exposure variable defined?
– Were any categories combined?
– Were observations excluded from a particular analysis because of missing covariates?
– Which model specification was finally selected?
These decisions can change the analytical dataset and, consequently, the results. Yet they are not always visible in the final manuscript.
The same applies to the code itself. A script may have worked perfectly on the researcher’s computer but be difficult for another person to understand or run. Software and package versions can change. Data-cleaning steps may exist only in an analyst’s notes. A variable may have been recoded manually without a clear record of how or why.
A 2025 BMJ Open paper on reproducible data analysis in medical research highlights exactly these practical issues. Based on an examination of coding practices in the Rotterdam Study and the existing literature, the authors recommend documenting data cleaning, sample selection, analytical decisions, software and package versions, and making code and data accessible when possible. They also emphasize code review and comprehensible code as practical ways to strengthen reproducibility.
This does not mean every dataset must be publicly available. Patient-level data may be restricted for ethical, legal or privacy reasons.
But even when the data cannot be shared, the process used to produce the analysis can be documented and, where appropriate, the analytical code can be shared.
A useful self-check is simple:
Could someone unfamiliar with your project reconstruct the path from the raw data to your final result?
If the answer is no, the issue may not be the statistical method.
It may be the missing record of everything that happened before the statistical method was applied.
hashtag#Epidemiology hashtag#ResearchMethods hashtag#Reproducibility hashtag#EpidemiologicalResearch hashtag#Biostatistics hashtag#ResearchTransparency hashtag#OpenScience