How would you decide whether missing source documentation threatens the reliability of a primary endpoint?

Short answer

Ask whether the endpoint value can still be verified from another original record, how many participants and which arms are affected, and whether the gap could bias the comparison. A few unverifiable values spread evenly may be tolerable. A cluster at one site or one arm, or a pattern that hints at fabrication, is a threat to the result.

Why

E6(R3) requires source records to be attributable, legible, contemporaneous, original, accurate and complete, and data sent to the sponsor to match source or have discrepancies explained. When source is missing, the reported value cannot be verified, so the question is how much that uncertainty matters. First, look for alternative originals: an instrument printout, a laboratory system record, an imaging archive, or an audit trail in an electronic clinical outcome tool.

Second, count. Ten unverifiable values out of 600 is different from 60. Third, check distribution.

If missing source is concentrated at one site, in one arm, or around the time of an event, it can bias the comparison, not just add noise. Fourth, ask why. Lost paper after a flood is different from records that never existed.

The statistician can run a sensitivity analysis that excludes or imputes the affected values to see whether the conclusion changes. If it does, or if fabrication is suspected, escalate as a possible serious breach in the EU.

A strong answer covers

Search for alternative original records first
Scale of the gap relative to the dataset
Distribution across sites, arms and time
Cause: loss versus records that never existed
Sensitivity analysis and serious breach assessment

Common mistakes

Recreating source documents after the fact to fill the gap.
Judging only by the count and ignoring whether the gap is unbalanced between arms.
Assuming the value in the case report form is right because it looks plausible.

The interviewer might ask next

What sensitivity analyses would you ask the statistician to run in this situation?