Goleman's five domains are a vocabulary, not a measurement

A manufacturing company in Coimbatore ran an emotional intelligence workshop for 60 team leaders in June. In September the HR head put a slide in front of the plant leadership: average score 3.4 out of 5, up from 3.1. The plant head asked the only question that mattered. Which twenty leaders changed, and what did they do differently on the floor?

Nobody could answer. The vendor had a score per person, written by supervisors who had attended a two hour briefing and rated each leader on five domains on a five point scale. The numbers moved. Nothing behind them could be inspected.

That is not a failure of the workshop. It is a failure of using a vocabulary as if it were an instrument. The five domains are a good way to talk about a manager. They are a poor way to measure one, and the difference matters the week someone asks you to defend a score.

A domain defined by an internal state cannot be observed

Read the five domains as Goleman originally set them out: self-awareness, self-regulation, motivation, empathy and social skill. Four of the five are defined by something happening inside a person. Self-awareness is "knowing your emotions". Empathy is "understanding what others feel". Motivation is "a drive to achieve for its own sake".

A rater cannot see any of that. What the rater sees is a person not raising their voice after a shift handover went badly. From that single act the rater infers self-regulation, and the inference is where the error enters. Two supervisors watching the same thirty seconds will attach it to different domains, weight it differently, and walk away with scores two points apart on the same leader.

This is not a training gap. It is what a rating scale built on unobservable constructs does. Explaining the domains more carefully will not fix it, because the problem is not that raters misunderstand the words, but that the words point at something the rater has to imagine.

The labels moved, which tells you what they are

Goleman's own model has been reorganised since the 1995 book. In later work the five components became four domains (self-awareness, self-management, social awareness and relationship management) carrying twelve competencies, which is the version most leadership programmes now teach. The vocabulary is more useful in that form, not less: twelve named behaviours are easier to discuss than five abstractions.

Notice what the revision implies. A measurement instrument does not quietly re-cut its own factors and keep the same name on the axis. A communication framework does. The five domains, and the four that followed, are a shared language for describing how someone works with other people, and treating a language as a scale is the error that produces a 3.4 on a slide.

What the research actually supports

The evidence is narrower than the training decks suggest, and the shape of it matters.

Joseph and Newman's 2010 meta-analysis in the Journal of Applied Psychology found a cascading pattern among ability based facets, from emotion perception to understanding to regulation to performance. It also carries the finding left out of the workshop: the relationship between ability based emotional intelligence and job performance was inconsistent, positive in jobs with high emotional labour demands and negative in low ones.

In a role where the job is largely managing other people's feelings, an emotional ability helps. In a role where the job is accounts payable, being highly tuned to other people's emotional states predicted slightly worse performance. The domain acts like a fit between person and role, not a general virtue. A claim that a team leader's empathy score should be high regardless of the work is a claim the evidence does not support.

The second thing the research supports is about the rating process itself, and it is the useful part. Frame of reference training, where raters are taught the same standard for what counts as a 1, a 3 and a 5 and then practise on the same recorded examples, is the best supported way to make behavioural ratings comparable. A 2024 analysis of four frame of reference studies found the training was most effective at identifying low performance, and that the setting of the training made little difference. The mechanism is agreement, not insight. Raters get better because they are shown the same behaviour and told what the shared answer is.

That is the whole method in one sentence: stop scoring the domain, score the incident, and check that two people score it the same way.

Step one: choose two domains, not five

Take the five domain names off the assessment form. Then pick the two that the role actually punishes people for getting wrong. Not the two that sound most important. The two where a bad week produces a complaint, a missed shift, an escalation or an exit.

Find them by asking three managers for six incidents each from the last six months: what the person did, and what happened next. Tag the twelve to eighteen incidents. Two tags will dominate. In a plant that is usually self-management under pressure, and social awareness in how a supervisor handles a worker about to lose an argument. Motivation and self-awareness rarely survive, because nobody brings you an incident about them.

Two domains is deliberate. Four domains rated by a supervisor who has seen each leader for ten minutes produces four numbers with the same error and half the attention.

Step two: write the cue so a camera could catch it

Each domain becomes one cue, written as something that could appear in a video with the sound off. The test is simple: if the sentence needs a thought bubble to make sense, it is not a cue.

The right column has a property the left does not: two people can disagree about what they saw, but not about whether it happened. That is what turns a rating into a record. Write the cue in the negative direction where you can: raters agree faster on an interruption than on a virtue, and the training evidence says agreement is easiest at the low end.

Step three: three anchors, written as incidents

Do not use a five point scale with word labels. Use three anchors, each one a specific incident you have actually seen, and rate on 1, 3 or 5 with nothing in between.

Anchor 1: in the handover meeting he told the shift supervisor the plan was "the usual mess" while two operators stood there, and the supervisor stopped contributing for the rest of the meeting.

Anchor 3: in the same meeting he said the changeover time was two hours over, asked the supervisor what was blocking it, and wrote the answer down.

Anchor 5: he opened by saying the previous week's changeover had gone badly and he wanted to understand why before assigning blame, then let the supervisor talk for four minutes without interrupting.

Those are not illustrative virtues. They are three different films, and a rater who has watched a fourth has to decide which of the three it most resembles. Forcing a choice between concrete incidents is what makes agreement possible.

Step four: calibrate on recorded incidents

Now run the session that most programmes skip. Record or collect eight to ten short incidents, either real footage with consent or re-enacted by volunteers. Have every rater score all of them independently, then reveal the agreed answer and argue through the disagreements.

The argument is the training. When a rater marks the four minute listen as a 3 and the agreed answer is 5, the gap is not knowledge, it is a different theory of what good supervision looks like. Frame of reference training replaces that private theory with a shared one, which is why everyone practises on the same examples.

Budget three hours: two domains, ten incidents, six to eight raters.

Step five: test agreement before you rate anyone real

Before the instrument touches a live leader, run the check that decides whether the domain survives.

Take ten fresh incidents. Two raters score all ten independently. Count how often they land within one point of each other, and check the direction. If a domain gets eight or more of ten within one point, it is usable. If it gets five or six, the cue is too vague or the domain is too broad, and you go back to step two rather than proceeding. If it gets four, drop the domain. That saves a year of reporting a number that is mostly rater taste.

Run the check on the anchors too. A domain often fails because one anchor describes two different behaviours, and that is a five minute fix.

What this looks like on a real set of numbers

Illustrative: 40 team leaders, two domains, two raters, ten calibration incidents. Self-management survives with nine of ten within one point. Social awareness survives with eight of ten, after the anchor for "names the person affected" is rewritten to cover written messages as well as meetings. Motivation, which someone insisted on adding back, fails at five of ten and is dropped. The report then says: in four of six observed handovers, this leader interrupted the supervisor within the first minute.

Six hours of rater time buys that. Five domains at 3.4 buys a slide.

The limit of this method

This measures behaviour that was observed, in contexts the rater happened to see. It is not a measure of capacity, and it is not stable across contexts: a leader who is calm with the shift supervisor may be volatile with the vendor, and your rater saw one of those rooms.

It is also reputation shaped. Who gets observed, and by whom, decides the score more than the person's inner life does. If two leaders behave identically and one has a manager who likes them, the scores will differ, and no anchor fixes that.

So keep it out of selection and pay. To compare candidates, use a structured exercise with trained assessors or an ability instrument with published validity, and accept that you are buying something narrower. Use the behavioural method for development conversations, where a specific incident and a shared standard are what the conversation needs and being wrong costs a coaching hour, not a career.

One more: do not compute a total. Adding two domains into an "EI index out of 10" recreates the problem you just spent six hours solving, because the sum implies the parts were measured on the same ruler.

Before the next workshop

Open the deck and find the slide with the five domains. For each one, write the single sentence a rater could watch happen. If you cannot write it in two minutes, that domain belongs in the discussion part of the programme, not the assessment part.

Then take the domain you care about most, collect ten incidents, and have two raters score them blind. Agree on eight, and you have an instrument. Disagree, and you found out in an afternoon rather than in a review meeting next quarter.

If you want the assessment side built rather than assembled, our 360 and behavioural assessment tools structure rater groups and anchors with the same logic. For why self-report EI scores behave so differently from observed behaviour, self-report EQ tests covers that measurement problem. For anchors across a wider competency set, start with building a framework managers will actually use.