Team-level EQ scores mostly measure your team size

A plant HR head sent me a slide last year. Twenty-two teams, one EQ score each, ranked from 4.2 down to 3.1, with the bottom four flagged for intervention. The maintenance crew at the bottom had six people. The sales team at the top had nineteen.

The six-person team did not have the lowest emotional intelligence in that plant. It had the smallest sample. Nobody in the room knew that, because the slide looked exactly like a ranking, and a ranking looks like a measurement.

This is the most common error in team-level EQ work, and it is arithmetic, not judgement. An average of six people's scores is a number. It is not a team score, and it is not stable enough to carry a training decision.

What you are actually averaging

There are two different things people call team emotional intelligence, and they are built from different data.

The first is the mean of individual EQ scores. You test twelve people, add up the scores, divide by twelve, and call the result the team's EQ. That measures the average ability of the individuals in the room. Move one person out and another in, and the number changes, but nothing about how the team handles a bad shift has changed.

The second treats the team as the unit, and asks each member to rate the team rather than themselves. The item changes from "I can tell when a colleague is upset" to "In this team, we notice when someone is struggling". This is called a referent shift, and it is a genuine group-level construct: you are measuring a shared perception of how the team operates, not a sum of individual personalities. Druskat and Wolff defined team emotional intelligence this way, as the ability of a group to develop norms that manage emotional processes, and the research team that built the T-TMMS measure used exactly this referent-shift approach rather than aggregating self-ratings.

The distinction is not academic. A meta-analysis of climate studies by Wallace and colleagues found that referent-shift measures were stronger predictors of job performance and customer service performance than direct consensus measures, while direct consensus was the better predictor of job attitudes. Which one you use depends on the question. If it is "will this team deliver", shift the referent. If it is "are these people happy", direct consensus is fine.

Most vendor reports quietly mix the two. They average self-ratings and label the result "team EQ". That is the first thing to check in any report you are handed.

The arithmetic that decides whether the average means anything

Even with the right items, averaging claims that the group mean is a reliable estimate of something real at the group level. That claim is testable, and cheaply.

The relevant statistic is ICC(2), the reliability of the group mean. It depends on two things: how much of the variance in scores is between teams rather than within them, called ICC(1), and how many people are in the team. The relation is:

ICC(2) = k × ICC(1) / (1 + (k − 1) × ICC(1))

where k is team size. Bliese's own worked example in the multilevel literature is a good calibration. In a dataset of 99 groups averaging 54.5 members each, work hours had an ICC(1) of 0.174, which produced an ICC(2) of 0.92. Groups could be reliably ranked on average work hours. In the same dataset, a coping item had an ICC(1) of 0.0098 and an ICC(2) of 0.35. The group averages existed. They were noise.

Now run the same arithmetic at Indian team sizes. Below is ICC(2) against team size for a few plausible ICC(1) values, computed from the formula above.

Read the left column. A team of six with an ICC(1) of 0.10 has a group-mean reliability of 0.40. That is below any defensible threshold. The number is not a slightly imprecise team score. It is mostly measurement error with a decimal point attached.

Turn the formula around and it gives you a stop rule. The team size you need for a reliable group mean is k = r × (1 − ICC(1)) / (ICC(1) × (1 − r)). At a target reliability of 0.70, a team needs about 44 members when ICC(1) is 0.05, 21 when it is 0.10, 13 when it is 0.15, and 9 when it is 0.20. At a target of 0.80 the numbers nearly double: 76, 36, 23 and 16.

Almost every real team in an Indian plant, branch network or services firm is smaller than that. Which is the finding, not a technicality: for the team sizes most L&D teams work with, the averaged EQ score cannot carry a decision.

The method, in the order you do it

Step one: name the construct before you field the survey. Write down whether you are measuring the average ability of the individuals or a shared perception of how the team works. If it is the second, the items must refer to the team, not the self. Do not field self-report items and then relabel the average as a team property.

Step two: compute agreement, then reliability, before you look at any team's score. Agreement and reliability are different questions and you need both. Agreement asks whether members describe the same team. Reliability asks whether the team mean is stable enough to distinguish teams from each other.

Biemann, Cole and Voelpel make two points that are routinely ignored. The 0.70 cut-off for rWG has no theoretical basis and should be set from the context and the response scale, not copied. And a high rWG does not show you are measuring the construct you think, because agreement can come from a shared bias rather than a shared reality. Agreement is a permission slip for aggregation, not a validity certificate.

Step three: if the tests fail, do not report the average. This is the whole discipline. Low agreement means there is no team-level score to report, only a distribution of individual views. Averaging a bimodal team is how you get an average of 3.5 that describes nobody.

Step four: choose your two metrics and stop adding more. For team development, two are enough, and they answer the two questions that actually lead to action.

The first is median rWG(J) on the emotion items, per team. It tells you whether the team has a shared way of handling emotional events or whether each member is describing a different team. A team that disagrees about itself has a fault line, and that is coachable. A team that agrees it handles conflict badly needs a different intervention.

The second is the within-team standard deviation on the regulation items. Two six-person teams can both average 3.5 out of 5. In one, the responses are 5, 5, 5, 2, 2, 2: a split camp, standard deviation 1.64. In the other they are 4, 3, 4, 3, 4, 3: a mild shared view, standard deviation 0.55. Identical means, opposite diagnoses, opposite actions. The mean destroyed the only information a manager could have used.

Step five: set a minimum team size for reporting, and put it in the rules document. Compute it rather than guess. Take the ICC(1) from your own last cycle, put it in the formula, and pick the size that gets ICC(2) to 0.70. Below that, report the distribution or the individual scores, never the team average.

What this looks like on real numbers

Illustrative figures for a 240-person manufacturer with 22 teams, so the sequence is visible.

The team sizes run from 5 to 19, median 9. Last cycle, ICC(1) on the emotion-regulation items came out at 0.12, and median rWG(J) was 0.61, below the conventional line. Applying the formula, teams needed 17 members for a reliable group mean, and only three of the 22 teams were that large.

The report that came out of this had three columns instead of one: team size, median rWG(J), and within-team SD. The score column was struck. Three teams with rWG(J) above 0.70 and SD under 0.60 were given their team-level profile, because for those teams the aggregate was defensible. Eleven teams got a note saying the responses were too dispersed for a team score, with the dispersion shown. Eight teams below eight members got individual profiles only.

Two things happened next. The plant head stopped ranking teams, because the report no longer ranked them. And the three teams with genuinely low agreement turned out to have a real split: a supervisor and the shopfloor crew describing different teams, which is exactly the fault line an intervention can address.

Where this advice stops working

Four limits, honestly.

ICC(1) is estimated from your own data and carries real uncertainty. If you have fewer than about 15 teams, treat the number as indicative and do not compute a precise size threshold from it. Run it again next cycle and see whether it holds.

Agreement statistics assume people answer similarly for similar reasons. In a team where one member is new and one has been there nine years, some disagreement is a true description of different experiences, not a measurement failure. Excluding recent joiners is legitimate, but record it, because it changes what the score means. Deciding which voices count is the same problem that shows up in how to read a self-report EQ test honestly.

A group-level score is still not a diagnosis. It tells you a team's shared perception of its own process, not whether that perception is right. A team can agree confidently on a distorted picture of itself.

And none of this fixes the larger problem that EQ data at any level is easy to turn into a label. The reporting discipline that stops that is separate work, set out in how to use EQ data in manager development without turning it into a label. Team scores need it more than individual scores do, because a team cannot argue with a number it was never shown the arithmetic for.

The action for this week

Pull last cycle's team-level EQ report. For each team, add two columns: team size and number of respondents. Compute ICC(1) and ICC(2) with team size in the formula above, and strike every team average whose ICC(2) lands below 0.70. You will lose most of the ranking. What is left will be the only part of that report that was ever measuring a team, and it will be a much shorter list of things worth acting on.

If the tool you use cannot produce ICC(1) and within-team dispersion, that is a product decision you can now make on evidence. What the assessment platform reports is either enough to run this check or it is not, and knowing which is the difference between a development plan and a slide that ranks six people against nineteen.