A spreadsheet filter can turn a soft preference into a hard exclusion. If the rule is “keep colleges with at least a 70% outcome rate,” an institution displayed at 69.9% disappears while one at 70.0% remains. That sharp difference in treatment may have little relationship to the precision or practical importance of the underlying evidence.

The Office for Students' data FAQs explain that published data can involve rounding and disclosure controls. The invented boundary exercise below does not reverse-engineer protected records. It asks a different question: is the reader's screening rule more precise than the information being used?

Examine the boundary before applying it

Suppose a fictional publisher rounds rates to one decimal place using ordinary nearest-value rounding. A displayed 70.0% corresponds approximately to underlying values from 69.95% up to, but not including, 70.05%, with endpoint treatment depending on the exact rounding convention. A displayed 69.9% corresponds approximately to 69.85% through values below 69.95%.

Displayed rate Illustrative underlying value Filter at displayed 70.0%
69.9% 69.94% Excluded
70.0% 69.96% Included
70.1% 70.06% Included

The first two underlying values differ by only 0.02 percentage points. A binary screen treats that tiny separation as decisive. This does not make the arithmetic filter incorrect; it shows that the chosen decision rule may be stronger than the evidence warrants.

Now suppose the rates are survey estimates with additional uncertainty. The plausible ranges may be much wider than the rounding intervals. Increasing the display to two decimal places would not remove that uncertainty. Precision of presentation and accuracy of inference are different properties.

Distinguish hard requirements from soft indicators

Some thresholds genuinely are hard rules: a verified application deadline, a formal prerequisite or the maximum available budget can make an option infeasible. A descriptive outcome statistic usually does not function in the same way. A personal screening cutoff should therefore have an explicit purpose and a tolerance justified by the decision context.

One approach is a review band. Instead of excluding everything below 70%, automatically retain clearly stronger cases, manually review values near the boundary and exclude clearly unsuitable cases only after examining other evidence. The band width should reflect the available precision and the consequences of error, not an arbitrary desire to retain a particular institution.

Do not fill suppressed or missing cells with zero so that the filter can run. That creates false exclusions. Give unknown values their own status and identify whether another reliable source can resolve them. If not, a decision may need to acknowledge uncertainty rather than manufacture a comparable number.

Audit the exclusions, not just the survivors

After applying a shortlist rule, inspect a small boundary report showing the closest excluded and included options. Record the original values, rounding notes, reference years and population definitions. A mismatched year or denominator can matter much more than the difference that triggered the cutoff.

Keep the cutoff recorded before reviewing results, so an appealing outcome does not silently redefine the selection rule.

Then test nearby rules. If moving the threshold from 70% to 69% replaces half the shortlist, the screen is sensitive and deserves review. If the feasible options remain stable, the cutoff may be a convenient organising tool, though it still does not establish a causal or personal outcome guarantee.

The output should explain the judgment: “These fictional boundary cases require review because their displayed differences are smaller than the precision relevant to our decision.” A shortlist is a way to manage investigation. Treating every final decimal as a meaningful distinction can prevent the most useful investigation from happening at all.