Running a 2,000 variable real world data study

Data operations at a scale where process discipline matters more than effort. Over 2,000 variables across a multi-site observational programme.

The scale changes the problem

A study with fifty variables is managed by paying attention. A study with more than two thousand is not. At that size, anything that depends on a person remembering to check something will fail, not occasionally but systematically.

That is the whole lesson of this programme. The work was not harder than a small study, it was differently structured, and the structure had to exist before data collection started rather than being improvised once queries began arriving.

What that meant in practice

The data dictionary is the product. Every variable defined once, with its type, permitted values, units, derivation and relationship to other variables, before a single form was built. Ambiguity resolved at dictionary stage costs minutes. The same ambiguity resolved after data collection costs a protocol deviation and a re-query across every site.

Checks at entry, not at review. With this many fields, catching an error at monitoring means catching it weeks after the visit, when the source may no longer clarify it. Validation sits as close to the point of entry as the workflow allows, so the person who can answer the question is still in the room.

Query management by category, not by queue. Two thousand variables generate queries faster than anyone can work a flat list. Grouping by type surfaces the systematic problems, which is what you actually want. Twenty sites making the same mistake is one training problem, not twenty queries.

Consistency as a continuous check. Cross form and cross visit logic running throughout, not as a pre lock sweep. Finding an impossible date sequence during the study is a correction. Finding it at lock is a finding.

Real world data is not trial data

Observational data arrives from clinical practice rather than from a protocol. Fields that would be mandatory in an interventional study are genuinely missing, not missing in error, and treating the two the same way produces a query log nobody can close.

The distinction had to be built in: what must be present, what may legitimately be absent, and what absence itself means. Getting that wrong in either direction is costly. Too strict and sites drown in queries for data that never existed. Too loose and the dataset cannot support the analysis it was collected for.

What I took from it

The instinct at scale is to work harder. The correct response is to move work earlier. Almost every hour spent on the dictionary and the check library before first patient in saved several hours later, and the hours it saved were the expensive kind, spent at a distance from the source with the clock running.