Nadeem Shafique Butt

Professor of Biostatistics

Department of Family and Community Medicine

King Abdulaziz University, KSA

Contact

Data Quality & Measurement

Data Cleaning Workflow for Health Studies

A reproducible data-cleaning workflow to improve validity and reduce analysis errors in health datasets.

By Updated 7 min read

Quick answer

Data cleaning in health research requires standardized coding, missing-data strategy, outlier review, and reproducible audit trails. A structured pipeline improves model stability, prevents analytical errors, and ensures that published findings reflect true patterns instead of preventable data-quality artifacts.

Key takeaways

  • Preserve raw data as immutable and create analysis data through versioned, repeatable code.
  • Write validation rules from the protocol and data dictionary before reviewing unusual values.
  • Investigate outliers as possible clinical signals or data errors; do not delete them automatically.
  • Document every correction, derivation, exclusion, and unresolved query in an audit trail.

Build a Data Dictionary First

Define variable names, units, allowed values, and transformations before analysis. This reduces ambiguity across collaborators.

Missingness and Outliers

Profile missingness mechanisms and identify implausible observations using domain-aware thresholds rather than automatic deletion.

Auditability and Reproducibility

Maintain versioned scripts and cleaning logs. Reproducible pipelines are easier to validate during peer review and regulatory checks.

Practical method

Step-by-Step Workflow

  1. 1

    Profile the raw extract

    Confirm row counts, identifiers, data types, ranges, category levels, date formats, missingness, and duplicate patterns without overwriting source data.

  2. 2

    Run validation rules

    Flag impossible and implausible values, inconsistent units, broken skip logic, date-order violations, and cross-field contradictions.

  3. 3

    Resolve and derive

    Trace anomalies to their source, document approved corrections, define analysis variables in code, and distinguish structural from unexpected missingness.

  4. 4

    Review and freeze

    Independently review high-impact transformations, reconcile the final cohort, produce a cleaning report, and tag the locked analysis dataset.

Worked example

Cleaning a multicenter health registry

Scenario
A registry combines records from five hospitals with different date formats, weight units, missing-value codes, and patient identifiers.
Approach
Standardize source formats in a staging layer, map all missing codes to documented categories, convert units while retaining originals, resolve duplicate patients with approved rules, and generate site-level discrepancy summaries.
Interpretation
The final dataset should be reproducible from the preserved extracts. Site differences in missingness or implausible values remain findings to investigate, not nuisances to conceal.

Common Mistakes to Avoid

  • Editing the only copy of a spreadsheet
  • Converting missing values to zero
  • Deleting all statistical outliers
  • Applying undocumented manual fixes

Frequently Asked Questions

Is deleting missing records acceptable?

Only when missingness is minimal and plausibly random; otherwise consider imputation or model-based handling.

Why keep a cleaning log?

Logs preserve transparency, support QA review, and allow exact replication of analytical datasets.

References and Further Reading

  1. NIH data management and sharing policy
  2. NIH rigor and reproducibility resources