Participant-Reported and Device Data

Part of Clinical Data Management and Integrity: External Data and Reconciliation.

About 5 minutes. Free, in your language, no account needed.

Why this matters

Data reported directly by participants, and data generated by devices they wear or use, has a property nothing else in the trial has: nobody is present when it is created.

A site-recorded measurement has a clinician who can be asked. A laboratory result has a validated process and a technician. A participant completing a questionnaire at home at two in the morning has nobody, and the record is whatever the system captured.

That property drives everything about how this data must be handled.

What makes it different

There is no source behind it. The participant's response, as captured, is the source. There is no earlier record to verify it against, and source data verification in the ordinary sense does not apply.

Compliance is data. Whether a participant completed the assessment, when, and how often they did not, is itself informative and frequently related to the outcome. A dataset of completed responses with no record of the missed ones has discarded a signal.

The timing is the measurement. A symptom diary completed at the scheduled time and one completed retrospectively for four days are different measurements wearing the same label. Systems that permit backfilling without recording that it was backfilled destroy this distinction.

Device data arrives at a volume nobody reads. A wearable producing readings continuously generates more data in a week than the entire case report form does in a year. Nobody inspects it record by record, which means defects persist until an aggregate check finds them, and only if one exists.

The questions to settle in advance

What counts as a valid completion, including partial responses and responses outside a window.

Whether backfilling is permitted, and if so, how it is recorded so the analysis can distinguish.

What happens when the device fails. Devices are lost, break and run out of charge. A defined process produces a recorded gap with a reason; an undefined one produces a gap indistinguishable from non-compliance.

How a participant's identity is established, since nobody is watching. This is usually a device assignment rather than authentication, and the assumption that the assigned participant is the person responding should be stated rather than assumed.

What the analysis does with incomplete data, decided before the data exists rather than after anyone knows its shape.

The compliance trap

Sites and sponsors both want high compliance, and there are legitimate ways to get it: reminders, good instrument design, and a burden proportionate to what is being asked.

There is also an illegitimate route that appears without anyone deciding on it. When a site can see a participant's compliance and is measured on it, and the system permits late entry, entries appear. A coordinator reminding a participant at a visit to complete the four days they missed produces four records that look like daily diary entries and are a single retrospective recollection.

The countermeasure is structural rather than exhortative: record the entry timestamp separately from the reported timestamp, always, and make the difference visible in the data so the analysis can see it.

In practice

A data manager reviewing an electronic outcome setup asks what happens when a participant opens the daily questionnaire and abandons it partway.

The answer is that the partial response is discarded and they may start again.

She asks how often that happens. Nobody knows, because it is not recorded.

The problem is that abandonment is meaningful. A participant too unwell or too fatigued to complete a symptom questionnaire is telling the study something, and the current configuration removes that signal while leaving a dataset that looks complete.

The change is a configuration setting and one variable: retain partial responses, flag them, and record abandonment as an event. It costs a day now. In eight months the analysis would have had a clean dataset with a hole in it whose shape was unknowable.

What this looks like when it goes wrong

A study collects a daily pain score by handheld device, with the protocol requiring entry each evening.

The system permits entry for the previous seven days, so participants who forget can catch up. Only the reported date is stored; the entry timestamp is not retained.

At analysis, the daily pain scores show unusually low day-to-day variability, which is interpreted as a stable clinical picture.

The real explanation is that a substantial proportion of entries were made in batches covering several days, and a person recalling four days of pain produces a smoother series than four independent daily reports. The data measures recall, not daily pain, and there is no way to identify which entries are which.

The variable that would have separated them, the entry timestamp, existed in the device and was not retained, because nobody specified it.

Key takeaways