Where Did This Number Come From

Part of Decentralized Trials (DCT) Training: Vendor Oversight and Data Provenance.

About 2 minutes. Free, in your language, no account needed.

Provenance is the defining question of decentralized data

For any value in the dataset you should be able to say what produced it, who transcribed or transmitted it, and whether anything changed it in between.

At a site, most data has one short provenance. Decentralized data has several long ones.

Four kinds, and they are not equivalent

Clinician-observed, by a nurse in a home or a provider in a clinic.

Device-produced, measured or inferred by hardware and an algorithm.

Participant-reported, entered by the participant.

Coordinator-recorded, what you wrote from a remote visit.

Each has different failure modes. Treating them as one kind of number is the mistake that decentralized studies make most often.

The transformation chain is part of the provenance

A device measures, an app aggregates, a vendor platform derives a daily value, an export maps it into the dataset.

A value can change shape at each step. A daily average is not a measurement, and knowing which you are looking at changes what it can support.

Algorithm versions matter

A vendor updating their sleep algorithm mid-study changes every derived value from that date. Nothing in the data itself announces it.

Vendor change notifications are therefore study documents, not IT emails. Read them and record which version applied to which period.

Ask what happens to data the vendor considers invalid

Many platforms drop or impute readings they judge unreliable. If that happens silently, your dataset is already filtered by somebody else's rule.

Find out what the rule is and whether the excluded data is retrievable.

Time zones and clocks again

Devices, apps, platforms and your own records may each stamp time differently. Reconciling an event across three systems with three clocks is the most common data provenance problem.

If you cannot describe the path, you cannot defend the number

That is the test. Not whether the number looks right.