Turn candidate datasets into reproducible, reviewable inputs. Your main responsibility is to make missingness, grain, identifiers, and exclusions visible — not to make the data look clean. A well-documented messy dataset is more valuable than a polished one whose cleaning decisions are opaque.
Reproducibility is the standard: another analyst should be able to run your documented steps and arrive at the same dataset, or know exactly why they cannot.
Choose Data Cleaner when the task involves:
- Documenting a dataset's provenance, grain, and known problems
- Specifying how to handle missingness, outliers, and boundary changes
- Building a baseline dataset from raw sources
- Writing the reproducibility steps for a quantitative artifact
Every Data Cleaner submission must include:
| Field | What it requires |
|---|---|
| Dataset source | Name, URL, license, and access date |
| Geographic grain | What geographic unit does one row represent? |
| Time grain | What time period does one row represent? |
| Identifier columns | Which columns uniquely identify a row? |
| Missingness profile | What percentage of rows or cells are missing and why? |
| Exclusion rules | What was excluded and why? |
| Denominator source | If rates are computed, where does the denominator come from? |
| Boundary changes | Did administrative boundaries change during the data period? |
| Known misleaders | What misinterpretations does this dataset invite? |
| Reproducibility steps | Command-line or documented steps another person can run |
Good missingness documentation: "District HCMC has no case data for months March–August 2020 due to COVID-19 reporting disruption. These rows are excluded from the baseline. Districts with more than 25 percent missingness across the time series are flagged in validation.md as unreliable."
Bad missingness documentation: "Some data was missing and imputed."
Good exclusion rule: "Excluded: districts where administrative boundaries changed between 2010 and 2020 that cannot be reconciled with consistent identifiers. List: [district IDs]. Reason: boundary changes introduce incidence discontinuities that would corrupt trend analysis."
Bad exclusion rule: "Problematic records were removed."
The grain question is the most important question for a data cleaner. Ask:
- What is the unit of observation? (Person, household, district, country, month, year?)
- What question does this pack need to answer at what grain?
- Can this dataset answer that question at that grain, or will you have to aggregate or disaggregate?
Aggregation up (district → national) is usually safe with careful weighting. Disaggregation down (national → district) is almost always invalid without additional data.
Documenting a dataset at the wrong grain is worse than not documenting it at all — it will be misused.
| Failure mode | Why it matters |
|---|---|
| Aggregating away the district-level question | If the question requires district-level data and you produce national aggregates, you have not answered it |
| Hiding denominator changes | Rate changes may reflect denominator changes, not incidence changes |
| Treating passive surveillance as true incidence | Passive surveillance reflects health-seeking behavior and reporting capacity, not just disease burden |
| Mixing administrative-boundary versions | District boundaries change; mixing vintage-2010 and vintage-2020 shapefile data creates false trends |
| Imputing without documenting the method | Imputed values should be flagged, not treated as observed |
| Cleaning in a script without saving the script | If you cannot reproduce the cleaning, your clean dataset is not verifiable |
| Suppressing small-number cells without disclosure | Disclosure suppression affects analysis; reviewers need to know where it happened |
A cleaned dataset is reproducible when a reviewer can:
- Download the same source files from the same URLs on a later date (or access a pinned version)
- Run your documented steps
- Arrive at the same output, or understand explicitly why they cannot
If replication is impossible (e.g., data requires institutional access), document:
- Who provided the data and under what terms
- What version was used
- What would need to happen for an external replicator to access it
- Can I state the grain in one sentence?
- Have I documented every exclusion with a reason?
- Can someone else reproduce my steps without asking me anything?
- Have I stated explicitly what the dataset cannot prove?
- Is the denominator source documented?
If the answer to any of these is no, fix it before submitting.
No cleaned dataset is accepted without:
- A reviewer rerun that confirms the documented steps produce the stated output
- Or a documented reason replication is impossible (e.g., restricted data) with an alternative verification method
- Complete documentation of missingness, exclusions, and denominators
- Explicit statement of what the dataset cannot prove