Plan first, results second

A variable is a single per-participant value used in analysis, for example the mean daily step count in the first seven days after surgery, or the length of a hospital stay. Its name is the column name in the exported file and does not change later. The formula lives in immutable versions: every change to the formula creates a new version, while changes to the label or description do not.

The preview computes each participant’s value from the data as of the current cut-off; values are not stored. Because it uses the revision history of clinical records and the version history of measurements, the same version with the same cut-off gives the same value.

When a value cannot be computed, it is left empty and the reason is recorded: no reference record, no end, no data, insufficient data or threshold not reached. No zero or estimated value is written, so “not measured” is never confused with “measured as zero.”

Selected variables become columns in a dataset with one row per participant. The dataset pins the current version of each variable and keeps using it even if the formula changes later. Files are pseudonymized and contain no identity fields.

An export includes variables.csv, missing.csv with the reasons for missing values, and codebook.csv as the data dictionary; Excel, SPSS (.sav) with variable labels and an R script that sets labels can be added. The same variable versions with the same cut-off produce the same content hash.

On the Analyses page, you first write a plan: the test, the variables, the significance level and the hypothesis. A plan that has not been approved is not computed. Approval freezes the variable dataset version and the data cut-off; an approved plan never changes, and changing it means opening a new analysis. Draft plans can be edited and deleted.

If the same dataset has more than one approved hypothesis test, a multiple-comparison warning with the Bonferroni-adjusted significance level is shown. The warning is not part of the result; interpretation stays with the research team.

Calculations use widely used open-source libraries, SciPy and statsmodels; most tests follow R’s default methods. Each result is stored with the input checksum and the library versions. The same plan, version and cut-off give the same result as long as participation records do not change.

Statistics and counts are not shown for a group smaller than 5. In cross-tables, cells smaller than 5 are hidden while the test still uses the full table. Complementary suppression keeps a hidden value from being recovered by subtracting from the total: when needed, the next smallest cell or group is hidden too. Regression requires at least 5 observations per coefficient.

Every result carries the note “for research, not for clinical decisions.”