The Kirkpatrick Model Explained: Four Levels of Training Evaluation

The Kirkpatrick model is a four-level method for judging a training program. The first level asks how participants reacted, the second what they learned, the third whether their behavior at work changed, and the fourth whether the business result the course was meant to move moved. It has long been the most popular framework in the field, and its weak spot is the assumption that each level follows from the one below. Trainees’ liking for a course barely predicted what they learned or how they later worked, a 1997 meta-analysis led by George Alliger found.¹

Anyone who commissions or runs training should decide which result the course should move before it is designed, and plan a separate check for each level instead of reading a stack of good feedback forms as proof that it worked.

Prefer to watch? The video covers the same ground as this article.

VideoThe Kirkpatrick Model Explained: Four Levels of Training EvaluationWiserHours

What the Kirkpatrick model measures at each level

The Kirkpatrick model sorts evidence about a training program into four levels, reaction, learning, behavior and results, first set out by Donald Kirkpatrick in work published in 1959 and 1960. Each level answers a different question, and each sits further from the classroom than the one before.²

Kirkpatrick Partners, the consulting firm that now promotes the model, says he built the levels in the 1950s for his doctoral research on training for supervisors, adapting four steps from the industrial psychologist Raymond Katzell.³ The table shows what each level asks and when published studies tend to measure it.

Take a two-day course that teaches new team leads to give feedback. Level 1 is the form handed out as people leave. Level 2 might be a role play scored against a checklist on the second afternoon. Level 3 asks whether, a few weeks on, the leads hold feedback conversations with their teams. Level 4 asks whether the problem that prompted the course, perhaps new hires leaving early, has eased.

The distance from the classroom is the reason the levels get harder to measure as they climb. By the time results arrive, a new product, a hiring freeze or a good quarter may have moved the same numbers, which a 2003 meta-analysis by Winfred Arthur and colleagues names as a likely reason results-level studies are rare.⁴ Organizations follow the path of least effort: in 1994 to 1996 data from the US-based American Society for Training and Development, cited by Alliger’s team, more than 90 percent of courses were evaluated at level 1, and almost none at level 4.¹

That pattern is understandable. A form costs minutes and arrives the same day, while a behavior check needs managers’ time and a results check needs data owned by someone else. But the cheap measure answers the least important question.

So before the course runs, pick the question you need answered. A feedback form cannot tell you whether behavior changed, however good the scores. Guides on planning learning programs will join this one among the talent development and training guides.

The New World version starts from the result

The most practical change to the Kirkpatrick model is the order of planning, not the levels themselves. In 2010, Jim Kirkpatrick, Donald’s son, and Wendy Kayser Kirkpatrick revised the definitions and called the result the New World Kirkpatrick Model.³

In the firm’s current version, evaluation starts with level 4: the outcome the organization needs, and then the handful of critical behaviors that would produce it. Level 1 now puts most weight on whether participants find the course relevant to their work, and level 2 adds their confidence and commitment to use what they learned. Level 3 includes the support and accountability people get back at work, such as manager check-ins and job aids, and a 2026 update adds the wider performance environment that can help or block every level.⁵

1. 1Resultsthe outcome the organization needs
2. 2Behaviorthe few actions on the job that drive it
3. 3Learningwhat people must know and be able to do
4. 4Reactionis it relevant to their work?

Treat this version as company material: Kirkpatrick Partners sells certification and consulting in the model, its pages are not peer-reviewed, and the changes are recommendations rather than tested findings.⁵ The backward order still earns its place on logic alone. It is the same logic as a hiring process built on criteria agreed before anyone applies: write down what success looks like first, and the evidence you collect afterward has something to answer to, rather than drifting toward whatever is easy to count.

In practice, a sales team that wants fewer lost renewals would start there, name two or three behaviors, such as calling accounts well before the renewal date, and only then design the course and its tests. If nobody can name the behaviors, the problem may not be one that training can fix.

Does a good reaction score predict learning and behavior?

Mostly, no. Once researchers pooled the studies that measured more than one level, reaction scores turned out to be weak guides to both learning and behavior at work. In 1989, George Alliger and Elizabeth Janak identified three problematic assumptions in the model: that each level is more informative than the one below, that the levels are causally linked, and that they rise and fall together.² Alliger’s 1997 meta-analysismeta-analysis: A study that combines the results of earlier studies on the same question into one overall estimate. Pooling makes the estimate more precise, but it cannot repair the studies it pools: a meta-analysis of surveys is still survey evidence.Full entry in the glossary checked the correlations behind the last two with more studies.

The study

Moderate evidence

Liking, learning and doing: how 34 studies lined up (1997)

Across the studies, reactions correlated about 0.08 with learning measured right after the course and about 0.13 with behavior on the job. Reactions about how useful a course was related to later job behavior more strongly than reactions about how much people enjoyed it (about 0.18 against 0.07), and learning test scores themselves correlated only about 0.11 with job behavior.¹

A correlationcorrelation: A measure of how closely two things move together, running from minus one, where one rises as the other falls, through zero, meaning no link, to plus one. It says how strong the relationship is, not what causes it, and it is not a percentage.Full entry in the glossary near zero means one score tells you almost nothing about the other. A course people loved and one they merely tolerated could produce much the same change at work. Some links rest on a handful of studies, and only three studies reported correlations involving results.¹

Bar chart with 5 bars; the largest is Usefulness reactions to job behavior at 0.18.

• Usefulness reactions to job behavior0.18
• Reactions to job behavior0.13
• Learning to job behavior0.11
• Reactions to learning0.08
• Enjoyment reactions to job behavior0.07

Why would enjoyment say so little? A 2008 meta-analysis by Traci Sitzmann and colleagues found that reactions mainly reflect features of the course itself, with the style of instruction the strongest predictor, so a form mostly rates how the course was delivered.⁶ Alliger’s team suggests that questions about usefulness do slightly better because trainees answer them with their own job in mind, including whether they will get the chance to apply what they learned.¹ Sitzmann’s team complicates that point: they found reactions linked to gains in motivation and self-efficacyself-efficacy: A person's belief in their ability to carry out a particular task or reach a particular goal. It is specific to the task, so someone can feel capable at work and unsure with money, and it goes with how readily people start and keep going.Full entry in the glossary, but no difference between enjoyment and usefulness questions in how they related to learning.⁶

1. The assumed chain: a good reaction leads to learning, learning to changed behavior, and behavior to business results
2. What the measures show: in pooled studies the levels were only weakly related, and liking a course was the weakest link of all

The gap between the two stacks is the practical problem. Reaction data are the cheapest and fastest to collect, so it is tempting to let them stand in for the rest, and the pooled studies say they cannot.¹ A trainer who is funny and well organized can earn high scores whether or not the skill sticks, and a demanding course that makes people practice until they fumble could score lower without teaching any less.

Myth

High satisfaction scores show that a course is working.

Fact

In pooled studies, how much trainees liked a course barely predicted what they learned or how they worked afterward.

Keep the end-of-course form short, ask whether the course was relevant and useful to the work, and never report its average as proof of impact. Alliger’s team also calls for research on asking again one to six months later, and on asking supervisors, once people can judge usefulness from experience.¹

Training can work even when the levels do not line up

Weak links between the levels do not mean training fails. Arthur’s 2003 meta-analysis of 162 sources found medium-to-large average effects at every level, from reactions through results, in evaluations that allowed an effect size to be calculated.⁴

The same paper shows where effects leak away. In studies that measured more than one level, the improvement was clearly smaller for behavior at work than for learning. The authors point to the transfer problem, where people may have no chance to use a skill or lose it before they do.⁴ A 2010 meta-analysis of 89 studies by Brian Blume and colleagues found that a supportive work environment was linked to more transfer of training to the job, alongside trainees’ motivation and cognitive ability.⁷

Picture sales reps who learn a new way to discuss price, pass the role play and return to a manager who still rewards quick discounts. Level 2 looks fine, level 3 looks like failure, and rewriting the course would miss the cause. Skill loss, the other cause Arthur’s team names, is the familiar case of a tool taught in March for a project that starts in September. Their pooled data have limits too: they come mostly from journal articles on off-site training and leave out single-group before-and-after designs.⁴ Whether corporate training pays off as a whole is a bigger question, which a planned guide on the overall evidence will take up.

When level 3 disappoints, look at the workplace before the classroom: does anyone expect, ask about or reward the new behavior?

How to evaluate a training course with the four levels

The Kirkpatrick model works best as a planning checklist and a vocabulary, with each level measured directly and read alongside what happened in the workplace. The table sums up what the evidence says about each link.

For a manager commissioning a course, that comes down to a short sequence. Write the result first. Name the behaviors. Test learning by having people do the thing, not by a quiz about it. Then check behavior at set points with the people who see the work, so a gap shows up early enough to add support.

A plan like this also shows its gaps early. If the result line stays blank, the course has no agreed purpose yet, and one question tests whether anyone will notice.

What the four levels leave out

The Kirkpatrick model lists outcomes but does not explain them, Elwood Holton argued in 1996 in Human Resource Development Quarterly. He called it a taxonomy of outcomes rather than an evaluation model, and proposed one that also accounts for the variables that affect those outcomes, such as motivation to learn, trainability, job attitudes, personal characteristics and the conditions for transfer back at work.⁸

That gap matters when a result is bad. The four levels tell you where the chain broke, but not why, so the next step is a guess unless someone also asked about motivation, workload or a manager’s support.

The ROI Methodology promoted by Jack Phillips’s ROI Institute goes the other way: its first four levels closely parallel Kirkpatrick’s, and a fifth, return on investment, compares a program’s monetary benefits with its costs after first isolating the program’s effect from other influences.⁹ Kirkpatrick Partners itself now warns that results rarely have a single cause and proposes crediting training with a contribution instead.⁵ If a sponsor asks for a return figure, ask first how the course’s effect will be separated from everything else that moved the number. Putting a money value on learning is the subject of a planned guide on measuring the return on L&D.

The bottom line

Feedback forms are the easiest part of the Kirkpatrick model to collect and the least informative, because the four levels do not predict each other well. Plan from the result backward, measure learning and behavior directly, and when behavior does not follow, look at the workplace as hard as at the course.