Interviewing the Spreadsheet: A One-Hour Data Dictionary Drill for Incoming Lab Members
A new graduate student opens the shared drive and finds a file named final_data_v3_USE_THIS.xlsx. The columns are labeled T_avg, flag, conc2, and date_run. One column contains the values 4, 9, and the string NA. Another contains −999 in roughly one row out of twenty. The student needs a single summary statistic for a rotation report and has no way to know whether the missing values are missing, censored, or never measured.
This is the ordinary condition of a working lab. Data outlives the person who collected it, and the knowledge required to read it usually does not. The usual remedy — a documentation requirement imposed at the end of a project — arrives too late and reads as paperwork. The drill described here moves documentation to the moment of inheritance, when there is a newcomer with questions and an owner with answers in the same room.
The setup
Book one hour. Two people: the incoming member, who has never used the dataset, and the owner, who collected or assembled it. One screen, shared if the pair is remote.
The rule that makes the hour work is a division of labor. The newcomer asks the questions and writes the dictionary. The owner answers and does not touch the keyboard. Owners who type produce the documentation they already had in mind; newcomers who type produce the documentation a future reader will need.
A visible timer helps. Sixty minutes is short enough to force prioritization and long enough to cover a moderate dataset.
Pass one: the story of the data (fifteen minutes)
The owner narrates the experiment from instrument to file. What was measured, in what order, on what days. Which runs failed and were repeated. What was renamed, merged, or discarded along the way. Where the raw files live relative to this spreadsheet.
The newcomer takes notes on the narrative, not the columns. Most of the ambiguity later encountered in a dataset is a residue of a decision made during collection — a sensor swapped mid-series, a dilution series redone, a batch relabeled — and those decisions surface naturally in a chronological account.
Pass two: column by column (twenty-five minutes)
Work left to right. For each column, five questions:
- What is this quantity, in plain language?
- In what unit, and is the unit recorded anywhere else?
- Which values are allowed, including every code used for missing or excluded data?
- What does a blank cell mean here, and does it differ from a coded missing value?
- Has this column ever been wrong, and how would we know?
Apply one discipline throughout: if the owner says "I think," the entry is written and flagged as unverified with the date. Uncertainty is information; a dictionary that hides it transfers a false confidence to the next reader.
Pass three: read-back (twenty minutes)
The newcomer writes the dictionary entry and reads it aloud for each field in a condensed form — name, definition, unit, allowed values, missing convention. The owner corrects. Read-back is the step most teams skip and the one that catches the gaps everyone assumed were obvious. Phrases such as "everyone knows that column is log-transformed" rarely survive being spoken to a person who did not know.
What belongs in the dictionary
| Field | Example | Why it matters |
|---|---|---|
| Column name as it appears | conc2 |
Readers search by the literal string |
| Plain definition | Nominal peptide concentration, second dilution series | Resolves ambiguity across versions |
| Unit | micromolar | Prevents the most common silent error |
| Allowed values and codes | 0–500; −999 = not determined; NA = not collected |
Makes missingness legible |
| Derivation | Computed from abs_raw using the 2026 calibration curve |
Distinguishes measured from calculated |
| Instrument or software version | Plate reader serial 2214; export script v2.1 | Supports reanalysis |
| Known quirks | Rows 40–58 collected on a replacement lamp | Prevents a false trend |
| Owner and last verified | A. Chen, 18 Nov 2026 | Tells readers whom to ask |
The dictionary lives beside the data file, versions with it, and carries a date. An undated dictionary is treated as a draft.
The seven-day follow-up
The hour is not complete at the end of the hour. Within a week, the newcomer performs one small real task using only the dictionary as a guide — one plot, one summary table, one filtered subset. Every question the dictionary failed to answer becomes a ticket against it. This is the actual test of the document, and it is far more informative than a review by the owner, who cannot see what is missing because they never needed it.
Why this is onboarding rather than compliance
Data dictionaries are usually produced under duress, at the end of a grant, by whoever is left. The result is a document written to satisfy a requirement and read by no one. The drill inverts the incentive: the newcomer produces it because the newcomer needs it, and the owner participates because an hour now prevents a semester of email questions later.
The exercise also surfaces tacit knowledge that would otherwise leave with a departing member. Labs that have run it report side benefits — protocols corrected, unit errors found in older analyses, one instrument sent for service after a quirk turned out to be a fault rather than a characteristic.
Keeping it alive
Two habits maintain a dictionary between drills. Every column added or renamed gets an entry in the same commit or the same week. Every handover — a student graduating, a postdoc leaving, a technician rotating off a project — includes a thirty-minute read-back of the dictionary with the next owner. Neither habit takes long, and together they keep the dataset usable after the original collector has moved on.