A college research tool can connect programme records across directories and appear more complete when it finds more links. But a link to the wrong campus or qualification can be worse than an unresolved record. The number of links made and the quality of those links are different measures.

The ONS data linkage policy distinguishes false links from missed true links and recommends reporting precision and recall rather than treating match rate as a quality metric. This article applies those definitions to an invented directory exercise. The figures are not an audit of CollegeVerse, ONS or any university database.

Establish a checked reference set

Suppose an analyst independently verifies that one hundred records in Directory A each have exactly one corresponding programme record in Directory B. That verified set is the reference for this small demonstration. A conservative matching method produces eighty links: seventy-six correct and four incorrect. It misses twenty-four true links.

Measure Conservative method Broader method
Links produced 80 100
Correct links 76 85
Incorrect links 4 15
True links missed 24 15
Precision: correct divided by produced 95% 85%
Recall: correct divided by 100 true links 76% 85%

The broader method finds nine additional correct links, but also eleven additional wrong ones. Its recall improves, while its precision falls. Saying only that coverage rose from eighty links to one hundred would conceal the trade-off. Neither method is automatically best for every use.

For an application shortlist, wrongly attaching a course's fee or entry requirements to another programme can materially mislead a student. An unresolved match instead signals that more checking is needed. The acceptable balance should depend on what the linked information will be used for, with uncertainty visible in the result.

Define what counts as the same programme

The reference decision needs a clear unit: institution, campus, award, delivery route, subject and relevant year. Two records with similar names can represent different programmes. Two with different names can sometimes describe the same entity after an official change. Names alone are insufficient to settle either case.

Prefer verified official identifiers and documented relationships where they exist. Check whether an identifier belongs to a provider, a location or an individual programme before using it as a join key. A provider-level match does not automatically establish a course-level match.

Keep the decision evidence and the reason a candidate link was accepted or rejected. If a programme genuinely maps to several records, the one-to-one assumption in the toy example is inappropriate. Adapt the reference design rather than forcing the real data into a convenient denominator.

Review unmatched records as well as accepted links

Checking only links that were produced can help estimate false positives, but it cannot reveal how many true matches were missed. A quality review also needs a defensible way to investigate unlinked records. Sampling design and uncertain reference decisions affect the precision of any resulting estimates.

The demonstration table assumes the full true-link set is known. In a real large directory it often is not. Report how the reference labels were created, which records were reviewed and which limitations remain. Do not turn an estimate from a small convenience sample into a guaranteed error rate for every subject or country.

The campus reporting scope guide helps identify the unit before linking. This audit adds a quality question after candidate matches are made: how many are right, and how many genuine relationships remain undiscovered? A useful research interface should preserve unresolved status and the original source links rather than making every record look confidently connected merely to increase apparent coverage.