Advanced Recordkeeping and Data Integrity

Lesson 4.1 · 12 min

Separating raw data from interpretation

Raw data is what an instrument read, what you observed directly, or what a subject reported, recorded as close to the moment of observation as possible, in its original units and without adjustment. Interpretation is everything you conclude from that data afterward. Mixing the two in the same entry makes it impossible later to tell what was actually measured versus what was believed.

Good practice keeps a raw log that is never edited after the fact, only appended to, alongside a separate analysis or interpretation document that can be revised as thinking develops. If a raw entry turns out to be wrong, it is corrected with a dated note explaining the correction, not silently overwritten.

This separation matters most exactly when you are wrong about something. A record where raw data has been quietly adjusted to match a later, better understanding is no longer a record of what happened, it is a record of what you now believe should have happened.

Raw data:

The original observation or instrument output, recorded without adjustment or interpretation.

Audit trail:

A record of who changed what, when, and why, sufficient to reconstruct the history of a document or dataset.

Lesson 4.2 · 11 min

Version control and change logs

Protocols and datasets change over the life of a project. A dated version number, even a simple one, and a short note describing what changed and why, turns a moving target into a traceable history.

When a protocol deviation happens during execution, record it in place, at the time, rather than adjusting the written protocol retroactively to match what actually happened. The gap between plan and execution is itself useful information.

Tools do not need to be sophisticated. A dated notebook entry, a spreadsheet with a change log tab, or timestamped files with consistent naming all accomplish the same goal: an outsider can reconstruct the sequence of decisions without asking you.

Lesson 4.3 · 11 min

Common integrity failures

Silently excluding inconvenient data points without recording the exclusion and its stated reason is one of the most common and most damaging integrity failures, even when the researcher genuinely believes the excluded point was an error.

Backfilling a protocol after seeing results, rounding numbers in a direction that favors a preferred conclusion, and reporting only the most favorable of several attempts are all forms of the same underlying problem: the record has stopped being neutral to the outcome.

None of this requires deliberate fraud to happen. Fatigue, hope, and time pressure are enough. Structuring your process so that these choices are made in advance, before you know the outcome, is the practical defense.