What "Good" Looks Like in Manager-Development Data

There is a specific moment that trips up almost every People or L&D leader who tries to prove their manager-development programme worked. You open the dashboard. The numbers are green. Completion is high, satisfaction is high, and confidence scores have climbed since the pre-course survey. It reads like success. And you still have no idea whether a single manager is behaving any differently in the conversations that actually matter.

That gap — between data that looks good and data that means something — is the real measurement problem in management development. It is not that organisations don't measure. It is that they measure the things that are cheap to collect and easy to make green, then treat those as evidence of behaviour change. The argument of this piece is narrow and, I think, defensible: "good" manager-development data is not a high score. It is a specific, observable behaviour moving in the right direction, from a known baseline, repeatedly, under realistic pressure. Almost everything else on the typical dashboard is a proxy dressed up as the real thing.

This is a different question from whether you should measure behaviour at all. That case is already made elsewhere. Here I want to assume you're convinced, you're trying, and you're now staring at a screen full of metrics asking the harder question: which of these numbers can I actually trust, and what does a credible one look like?

Why do manager-development dashboards look better than the managers?

Because they mostly measure inputs and feelings, not behaviour. The evaluation model nearly everyone borrows from — Donald Kirkpatrick's four levels — is explicit about this. Level 1 is reaction (did they like it), Level 2 is learning (did they know it afterwards), Level 3 is behaviour (are they doing it differently at work), and Level 4 is results. Behaviour is Level 3, and it is the level almost everyone skips, because it is the only one you cannot capture with a quiz at the end of a workshop [1].

So the dashboard fills up with what is easy: completion rates, attendance, satisfaction scores, self-rated confidence. Completion and attendance remain among the most commonly reported learning measures precisely because they are automatic — the system logs them for you [2]. None of them tells you what a manager did the next time an underperformer got defensive, or a good employee hinted they were leaving. A programme can score beautifully on every one of those metrics and change nobody's behaviour. When the data looks better than the managers, it is usually because the data is measuring the wrong layer.

Isn't a rise in manager confidence a good sign?

It is the single most seductive number on the board, and the one to be most suspicious of. Confidence is the thing training most reliably moves — a day of role-play and encouragement will lift almost anyone's self-belief. The problem is that confidence and competence are not the same variable, and they do not always move together. A manager can leave a course feeling far more assured about difficult conversations while behaving no differently, or worse.

We have hard evidence that self-perception is a weak stand-in for actual behaviour. A meta-analysis of self- versus supervisor ratings of job performance across 115 samples and more than 37,000 people found the two correlated at just r = .22 — and that self-ratings ran consistently more lenient than others' [3]. In plain terms: how a manager rates themselves tells you very little about how they actually perform, and skews optimistic. So a confidence score is a genuine data point — about how the manager feels — but it is not a behaviour metric, and the moment you let it stand in for one, your "behaviour change" is really just mood.

What does a credible behaviour-change signal actually look like?

Strip away the proxies and a trustworthy signal has a recognisable shape. It has four properties, and a metric that is missing any one of them is not measuring what you think it is.

• It is a behaviour, not a feeling or an input. Something a manager did that another person could observe — named checking of understanding, staying on the issue when the other person deflected, actually stating the standard rather than softening it into nothing. "More confident" and "course completed" do not qualify.
• It is observed, not self-reported. The signal comes from watching the behaviour happen — in a real or rehearsed conversation — not from asking the manager afterwards whether they think they've improved. Given the .22 correlation above, self-report and observation are close to different measurements.
• It is a change from a baseline. A single score in isolation is meaningless. You cannot know whether a "7 out of 10" is progress without knowing where this manager started. Behaviour-change data is a delta, not an absolute.
• It is repeated, not a one-off. One good rehearsed conversation can be luck or performance for the camera. A credible signal is a trajectory across several attempts — the slope, not the snapshot.

That last property matters more than it sounds. Skill change, as the deliberate-practice research has shown since Anders Ericsson's foundational 1993 work, is a curve built from repeated effort with feedback, not a step you take once [4]. If your data can only ever show a before and an after, it is structurally incapable of telling you whether a behaviour is actually forming or just briefly performed.

Why do good behaviour metrics go bad the moment you reward them?

Because measurement changes the thing it measures. This is Goodhart's Law, in the formulation the anthropologist Marilyn Strathern gave it in 1997: "When a measure becomes a target, it ceases to be a good measure" [5]. The instant a behaviour score is attached to a bonus, a promotion, or a league table of managers, people optimise for the score rather than the behaviour. Managers perform the tracked behaviours when observed and drop them when not; the number climbs while the reality doesn't.

This is why the collection context is part of the metric, not a footnote to it. Behaviour data gathered where the manager knows they are being graded is data about performance under assessment, which is a different — and less useful — thing than how they behave when it counts. "Good" data is gathered somewhere the manager is trying to get better, not trying to pass. Practise, not audit.

What should an L&D or People leader do differently?

Change what you demand from the measurement before you change the tool. Concretely:

• Define the behaviours first, in observable language. Not "communicates better" but "names the specific behaviour and its impact within the first two minutes", "checks the other person's understanding before moving on", "holds the standard when the conversation gets tense". If you can't see it happen, you can't measure it.
• Insist on a baseline. Measure the same behaviour before development starts. Without it you have a number, not a change.
• Measure at the point of practice, under realistic pressure. The behaviour you care about happens when the other person goes off-script and emotion is in the room. A calm retrospective survey samples the wrong moment entirely.
• Read the slope, not the snapshot. Ask for the trajectory across several attempts. Improvement that survives repetition is the thing you're actually buying.
• Keep confidence and competence in separate columns. Track confidence — it matters for whether managers will start the conversation at all — but never let it substitute for a behaviour score. When confidence rises and observed behaviour doesn't, that divergence is itself a finding worth acting on.
• Keep the data developmental, not punitive. The fastest way to destroy a behaviour metric is to grade managers with it. Measure to help people practise; the moment it becomes a stick, Goodhart takes it from you.

The five questions to ask before you trust a number

You do not need a new framework to read a manager-development dashboard well. You need five questions, and a metric has to survive all of them:

• Is this an observable behaviour — or an input, a reaction, or a feeling?
• Was it observed — or self-reported?
• Is it a change from a baseline — or a floating absolute?
• Is it repeated — or a single moment?
• Could someone game it if it became a target — and is it collected somewhere they'd want to?

Most of what currently sits on manager-development dashboards fails at least two of these. That is not an argument for measuring less. It is an argument for being honest about what the green numbers are telling you, and for holding out for the one metric that is harder to collect and worth infinitely more: a real behaviour, seen changing, again and again, when it mattered.