Training metrics and their denominators
A training average can be accurately transcribed and still be unsuitable for comparison. This study reads two public disclosures to distinguish what their denominators establish, what nearby workforce figures leave unresolved, and which calculations a reader should decline to make.
The cases pair one organisation’s staff report with a UK employer survey. They illustrate measurement differences, not training quality. Later numerical exercises are entirely fictional.
A workforce table is not a denominator specification
HM Land Registry’s staff report gives 5.6 training days per employee and 41.6 per apprentice for 2023–24. Nearby figures distinguish 6,993 permanent employees from 6,326 permanent full-time equivalents at 31 March 2024; a separate table gives an annual average of 6,254 FTE.
Our reading is deliberately bounded: these different workforce measures do not establish which count generated the training average. Multiplying 5.6 by the nearest headcount would invent an undisclosed training total. Nor should the apprentice average be added to the employee average. The selected table does not supply the matching numerators and population relationship needed for that calculation.
A survey separates trainees from all employees
The Employer Skills Survey 2024 report, updated June 2026, reports 5.7 training days per trainee and 3.6 per employee. It estimates that 63% of employees received training. Its findings are weighted unless stated otherwise; table bases show unweighted responding sites. The employee-based results must not be reconstructed using that site count.
The report covers employer-funded or arranged training over the previous twelve months. Its population and survey construction differ from a single employer’s records. Read the methodology and training definitions before treating its averages as benchmarks for a company. A simple average of employer averages is not a substitute for the survey’s employment-weighted calculation.
Decide whether a comparison survives the labels
These cases support a comparison of disclosure design, not a performance league table. A near match between a per-employee figure and a per-trainee figure does not make them comparable. The denominators, reporting windows and evidence-generating processes remain different.
For a research note, preserve the source’s unit and denominator wording before creating a standardised column. Give each proposed transformation a separate status: supported by the disclosed method, dependent on additional information, or not attempted. A blank converted value is more useful than a precise-looking estimate that cannot be traced to a defensible population.
An average also leaves the distribution hidden. Two organisations can report the same average while one spreads learning across most employees and the other concentrates long courses among a small group. Ask for participation coverage and a suitable distribution alongside the mean when the decision concerns access. Ask for assessment evidence when the decision concerns competence.
Original experiment one A falling participant average
All quantities in this and the next experiment are invented. Consider a stable workforce of 100 people in each of two complete years. Training days use the same approved duration rule throughout. Year A records 300 training days among 50 unique employees; Year B records 400 days among 80. No activities or employee categories change between years.
Year A: 300 ÷ 50 = 6 days per participant; 300 ÷ 100 = 3 per employee; participation is 50%.
Year B: 400 ÷ 80 = 5 days per participant; 400 ÷ 100 = 4 per employee; participation is 80%.
The participant average falls by one sixth, or 16.7%, while total days and days per employee rise by one third, or 33.3%. Participation rises by 30 percentage points. These movements can coexist because the participant group expands. None demonstrates that learning quality improved or deteriorated; that question requires evidence beyond duration and reach.
The useful management statement preserves all three views: more total learning was recorded, a larger share of employees participated, and the average among those participants was lower. Reporting only the falling participant average could imply contraction; reporting only the growing total could conceal unequal access. The paired measures explain the apparent contradiction without inventing an outcome.
Original experiment two Rounding cannot restore the roster
Suppose a different fictional disclosure reports exactly 900 training days and an average displayed as 3.0 days per employee. Its stated rule rounds the average to the nearest tenth, with a half-tenth rounded upward. The numerator and population are matched, but the exact employee count is withheld. Dividing 900 by 3.0 gives 300; that is only one possible count.
A displayed 3.0 allows an underlying average from 2.95 inclusive to 3.05 exclusive. The compatible whole-person denominators run from 296 to 305. At the endpoints, 900 ÷ 296 is about 3.0405 and 900 ÷ 305 about 2.9508; both display as 3.0. Counts of 295 and 306 instead round to 3.1 and 2.9.
Therefore the published precision cannot identify the roster, even under these unusually helpful assumptions. If the numerator were also rounded, or the denominator were annual-average FTE rather than whole people, this integer range would no longer describe the possibilities. Do not transfer this invented range to either real disclosure.
A practical research decision is to retain the disclosed average, mark the underlying denominator count as unavailable in the selected passage, and request the calculation basis only if the intended comparison requires it. There is no need to fill every database cell. Distinguish an absent detail in the passage you reviewed from a claim that the organisation never disclosed it anywhere.
What the study adds to a reporting review
When someone proposes a cross-company training chart, ask them to reproduce the matching decisions before reproducing the numbers. Record the reporting period, employee or participant population, headcount or FTE treatment, included learning, duration convention, weighting, rounding and any unresolved transformations. Keep original values beside proposed conversions so a reviewer can see where judgment entered.
The training-measures guide covers constructing measures from attendance records. This study addresses a later task: judging whether already-published figures can support the proposed comparison. Source labels are evidence to preserve; a research spreadsheet should not silently supply definitions that the selected disclosure does not establish.
Sources checked 14 October 2026. Contains public sector information licensed under the Open Government Licence v3.0. Analysis and fictional experiments are Nitium’s own; no source table or diagram is reproduced.
Comments