Learning transfer is the extent to which trained knowledge and skills are actually used on the job. Measuring it means collecting evidence of changed behaviour at work — not test scores, not satisfaction ratings — from a source other than the trainee, far enough after training that the behaviour has had a chance to appear.
That last clause is where most programmes quietly stop. In a large meta-analysis of training design and evaluation features, Arthur, Bennett, Edens and Bell (2003) coded 397 data points from 162 sources covering 33,325 participants. Only 122 of those 397 data points — 31% — used a behavioural criterion at all. A 2026 benchmark survey of 101 US HR managers by TalentLMS found 37% measuring L&D by business impact. The questions below are the ones that come up once a team decides to actually do it.
What counts as evidence of transfer rather than evidence of learning?
A score on an end-of-course test is learning. A rating of how often someone performs the trained behaviour at work, or a scored sample of their actual work output, is transfer.
The distinction matters because the headline numbers hide it. Arthur et al. report almost identical sample-weighted effect sizes across criterion types: reactions d = 0.60, learning d = 0.63, behavioural d = 0.62, results d = 0.62. But when the authors restricted the comparison to the same samples measured both ways (k = 17 each), learning came in at d = 0.66 and behaviour at d = 0.44. Different studies populate each column of the overall table, so those averages conceal a real drop.
How long after training should I measure?
The field's working norm is roughly four months. Arthur et al. report average measurement intervals of 26.34 days for learning criteria, 133.59 days for behavioural criteria, and 158.88 days for results.
Usefully, the interval was essentially unrelated to the size of the effect found (r = .03). Waiting four months rather than four weeks did not systematically shrink what studies detected. That removes the main excuse for measuring early: an early measurement is not a more flattering one, it is just a measurement of something else.
Can I just ask trainees whether they are using what they learned?
You can, but you should not treat the result as the same evidence. Blume, Ford, Baldwin and Huang (2010) pooled 89 studies, 325 independent correlations and 12,496 participants, and compared transfer measured by self-report against transfer measured by someone else or objectively:
- Trainee motivation and transfer
- Self-reported transfer: .33 (k = 10)
- Other-rated or objective transfer: .11 (k = 13)
- Pre-training self-efficacy and transfer
- Self-reported transfer: .29 (k = 8)
- Other-rated or objective transfer: .14 (k = 10)
- Work environment and transfer
- Self-reported transfer: .28 (k = 8)
- Other-rated or objective transfer: .20 (k = 12)
The sharpest finding concerns same-source, same-moment measurement. Where the predictor and the transfer rating both came from the trainee at the same sitting, the work-environment-to-transfer correlation was .54 (k = 13); without that contamination, .23. A post-course form asking people both how supportive their manager is and how much they are applying the training manufactures a relationship roughly twice the real size.
Who should rate the behaviour instead, and how much do raters agree?
Blume et al. report agreement between sources rating the same transfer at the same time: self with peer .26 (k = 4), self with supervisor .28 (k = 4), and all different-rater pairs pooled .28 (k = 9). Repeated ratings from the same source over time correlated .57 (k = 6).
Two things follow. No single rater is the answer — a supervisor rating is a different measurement, not a corrected one. And the apparent stability of a repeated self-rating is largely rater consistency, not accuracy. If you can only get one source, get the one that is not the trainee, and name it in the report.
Doesn't the "only 10% of training transfers" figure make this pointless?
That figure has no study behind it. It traces to a 1982 article by David Georgenson in Training and Development Journal (36(10), 75–78) which opened by asking readers how often they had heard training directors say only 10% of classroom content shows up in behaviour. Fitzpatrick's reconstruction of its spread, originally published in SIOP's TIP, notes that Georgenson cited no evidence or authority for the number; the often-attached dollar figure was added by later writers.
The measured picture is better than 10% and worse than the brochure: behavioural effects around d = 0.44 to 0.62, varying with what was trained and who did the rating.
Does the skill decay before I get to measure it?
Not all of it, and not uniformly. Taylor, Russ-Eft and Chan (2005), a meta-analysis of 117 studies of behaviour modelling training, found that declarative knowledge gains decayed over time while skill and job-behaviour effects held steady or increased.
That is a measurement instruction as much as a finding: what fades fastest is what a knowledge test captures. A late behavioural measurement is not automatically a weaker one.
What do I need to capture before training starts?
The same behaviour, on the same instrument, from the same source type you will use afterwards. Without a pre-measure you cannot separate a trained change from a selection effect, and matching the rater source matters as much as matching the items.
Worth noting how rarely the groundwork happens: in Arthur et al., only 22 of 397 data points — 6% — reported conducting a needs assessment at all.
What predicts transfer, so I know what else to fix?
Blume et al.'s corrected correlations with transfer, after excluding the same-source inflation described above:
- Cognitive ability: .37 (k = 10, N = 1,452)
- Voluntary participation: .34 (k = 5, N = 1,413)
- Conscientiousness: .28 (k = 5, N = 433)
- Transfer climate: .27 (k = 8, N = 870)
- Motivation: .23 (k = 24, N = 2,574)
- Work environment generally: .23 (k = 22, N = 2,085)
- Support: .21 (k = 12, N = 1,075)
The authors also report supervisor support at .31 against peer support at .14, flagging both as resting on small samples. Treat the ordering as a hypothesis about where to intervene, not a settled ranking.
Do open and closed skills need different measurement?
Yes, and the difference is large. Blume et al. found the work environment correlated with transfer at .26 for open skills (k = 18) versus .04 for closed skills (k = 4). Cognitive ability correlated .41 with closed-skill transfer (k = 8).
Closed skills — a fixed procedure with one correct execution — transfer largely as a function of whether the person learned them, so a scored work sample is adequate evidence. Open skills such as coaching or negotiation have no single correct execution and depend heavily on conditions back at work, so behavioural ratings in context do necessary work a test cannot.
What if supervisor ratings are genuinely unavailable?
Substitute a scored performance, not a self-report. A scenario or work sample at the four-month mark is an other-than-self measurement: the trainee produces the response, something else scores it against fixed criteria. AssessAll's AI-graded scenarios are built for this, and pay-as-you-go credits at ₹30/US$0.50 mean a second administration costs roughly what the first one did — which is usually the real reason post-measures get skipped.
Document the scoring criteria before administration either way, in line with the *Standards for Educational and Psychological Testing* requirement that intended interpretations be specified in advance.
The minimum defensible setup
- Define the behaviour in observable terms before the programme is designed.
- Pre-measure it with the instrument and rater source you will use again.
- Wait roughly four months — the field norm, and not a penalty on effect size.
- Measure from a source other than the trainee, or with a scored work sample.
- Never collect predictor and outcome from the same person at the same sitting.
- Report the rater source and the interval alongside the number.
If you measure transfer with a post-course form, you will get a number that is both higher than the truth and uninterpretable. The two changes that matter most cost nothing conceptually: move the measurement four months out, and take it from someone other than the person who sat in the room.