What Counts as Clean Data in a Messy Study

The idea of clean data is central to clinical trials. But in digital, hybrid, or real-world contexts, things rarely stay tidy. Data comes in late, arrives incomplete, or conflicts with other sources. So what does clean data actually mean in this setting? And how should teams define it when perfect consistency is out of reach?

‍

Let’s begin with a useful distinction: clean does not mean flawless. It means clear. Understandable. Trustworthy. Clean data is data that has a known origin, a documented flow and a defensible reason for being the way it is.

‍

Take this example:

A participant records their daily symptom scores five days in a row. On the sixth day, they skip. On the seventh, they enter two scores, one for today and one for the day before.

‍

That dataset is not perfect. But it can still be clean... if the system captured the timestamps, recorded the back-entry clearly and flagged the late submission as part of the audit trail. What would make it messy is if:

• The second entry overwrote the first without explanation
• The dates were changed manually but not logged
• The monitor cannot tell which data point was entered when

‍

So, what does clean data look like in the real world? It often means:

✔ Entries that reflect real participant behaviour, even if that includes gaps

✔ A system that captures not only the data, but the context

✔ Visibility into how and when corrections were made

✔ Acknowledgement of ambiguity, not the removal of it

‍

The danger is when cleaning becomes erasing. When systems overwrite values instead of versioning them. When CRAs edit entries to “fit the model.” When analysis depends on assumptions that aren’t backed by documentation.

‍

In messier studies - like nutritional interventions, lifestyle trials, or long-term observational designs - variability is not a flaw. It is part of the signal. Trying to normalise it away can erase what makes the data useful.

‍

That is why clarity is more valuable than control. If you know how the data got into the system, how it was reviewed and what decisions were made along the way, you can work with it. Even if it is imperfect.

‍

A clean dataset in this context is one where you can answer questions like:

• What did the participant do?
• When did they do it?
• Did anything change after entry, and if so, why?
• Can we trust this number to represent what actually happened?

‍

If the answer is yes - even if the number arrived late, was revised, or came with caveats - then the data is clean enough to work with. And often, that is what matters most.

‍