Overview
The quality engine profiles evidence without mutating the source and records every approved transformation as a new version.
Theoretical foundation
Diagnostics describe distributions, covariance and response patterns. Transformations are modelling decisions whose appropriateness depends on the study and instrument.
Algorithm and computation
Pandas/NumPy/SciPy compute missing patterns, sample skewness/kurtosis, Shapiro–Wilk for 3–5,000 nonconstant observations, z/MAD/IQR flags, guarded Mahalanobis distance, exact duplicates, selected-battery long strings and VIF. Available transformations include explicit row exclusion, recoding, reverse scoring, winsorisation, standardisation, mean centring and composite scoring; every operation is previewed.
Required data
A private immutable dataset version and researcher-selected variables for multivariate or battery checks.
Assumptions
Codes and valid ranges are correct, selected batteries are substantively related, and treatments are justified before approval.
Available settings
Selected diagnostics; recode mapping; scale endpoints; winsor percentiles; composite name; variables; row IDs; justification.
Worked example
A synthetic reverse-worded 1–5 item uses new = 6 − old after confirming endpoints. The preview shows changed cells before a new version is approved.
Limitations and non-claims
Little MCAR, multiple imputation, robust covariance distance, response-time analysis and automated case deletion are unavailable.
Validation notes
Quality calculations, safeguards, preview immutability and version creation are covered by repository tests.
How Surv.lytics computes Data-quality diagnostics
{}References and verification status
Little, R. J. A. (1988). A test of missing completely at random for multivariate data with missing values. Journal of the American Statistical Association, 83(404), 1198–1202. https://doi.org/10.1080/01621459.1988.10478722
Unverified draftOpen DOI
Was this article helpful?
Your response helps the methodology team improve this documentation.