The shape: one record per row, one variable per column
This is the rectangular, or tidy, form, and it is what every statistical package expects. Three rules define it: each variable has its column, each observation has its row, each cell holds a single value.
The most frequent departures, and what they cost
| Departure | What it looks like | What it costs |
|---|---|---|
| Multiple values in one cell | HTN, diabetes, smoking in a "history" column | Every analysis starts with string splitting |
| A variable inside the column name | ipss_m0, ipss_m3, ipss_m12 | Time becomes structure rather than data |
| Mixed units in one column | 0.8 and 70 in the same creatinine | Every statistic is wrong and looks plausible |
| Total rows inside the data | A "mean" row in the middle of the table | Counts come out wrong |
| Two-level headers | Merged cells | Import fails or silently shifts |
The six pieces of a deliverable dataset
- The data — the rectangular matrix itself.
- The dictionary — the meaning, type, unit and bounds of every column. See the data dictionary.
- The summary — counts, completeness per variable, distribution of the main ones.
- The metadata — study identity, protocol version, freeze date and signature, extraction scope.
- The missing-value table — for each variable, how many are missing and for which declared reason.
- The decisions — the choices that made the dataset: inclusion criteria applied, recodings, adjudicated cases.
These six pieces are what the FAIR principles call rich metadata. Their absence is the most frequent reason a publicly deposited dataset is never reused by anyone.
Preparing for SPSS, R and pandas
All three read a CSV happily, but they do not have the same expectations, and a few precautions avoid most of the friction.
- UTF-8 encoding, without exception. It is the leading cause of unreadable characters at import.
- An explicit decimal separator. A French CSV with semicolons and decimal commas imports badly elsewhere; declaring it in the metadata costs one line.
- Column names without spaces or accents.
ipss_totalworks everywhere;Score IPSS (total)needs transforming in all three. - ISO dates —
YYYY-MM-DD, unambiguous across conventions. - Missingness coded one way. An empty cell, not
NA,-99andunknownmixed together — unless every code is documented in the dictionary. - Categories coded and labelled. The code for analysis, the label for reading; both in the dictionary.
The freeze, and why a dataset must be dated
A dataset with no freeze date is not a dataset: it is a transient state. Two analyses run two weeks apart on a live database give two results, and nothing says which corresponds to what will be published.
The freeze fixes the content, timestamps it, and separates it from collection, which may continue. A dated extraction becomes citable, comparable and verifiable — exactly what a reviewer will ask for. For the mechanics of the freeze and its reasoned reversibility, see the audit trail.
How AnzarSeha addresses it
- The export is a six-sheet workbook — data, dictionary, summary, metadata, missing values, decisions — not a bare table.
- The matrix is rectangular by construction: one record per row, one variable per column.
- The available formats are open: Excel, CSV, PDF and Word.
- A STROBE-shaped draft accompanies the export, so writing up starts from what was actually done.
- The freeze is a dated, signed state, and the exported dataset is exactly the one that was frozen.
Frequent questions
Should raw or cleaned data be delivered?
Both where possible, with the transformation log between them. Failing that, the analysed data together with the decisions that produced it: that is what lets a third party retrace the path, and it is more useful than a raw file with no explanation.
Which file format for archiving?
An open format readable without proprietary software — CSV for the data, with the dictionary in an equally open format. A .sav or .xlsx opens today; archiving is judged at ten years.
How long should a research dataset be kept?
The period depends on the applicable framework, the commitment made to participants and institutional policy, and must be set before the study begins. What is certain is that an undecided period becomes indefinite retention by default.