Correlation Isn't a Driver

The slide is a clean horizontal bar chart, and it is titled Top Drivers of Engagement. Manager support sits at the top, then growth opportunities, then recognition, each with a tidy percentage next to it. The survey team built it by regressing the engagement score on the other survey items and ranking what came out. The room reads it the way it was meant to be read: these are the levers. A quarter later there is a manager-coaching initiative, a recognition program, and a reorg of the development budget — all aimed at the top of that chart.

Here is the problem, and it is not a problem of effort. Every number on that slide is a correlation, and the title calls them drivers. A driver is a causal claim. A correlation is not. The chart quietly promoted one into the other, and the company just spent a year and a budget on the promotion.

They say: run the key-driver analysis

The genre is everywhere, because it is easy and it sells. Take the outcome you care about — engagement, intent to stay, performance — regress it on everything else you measured, and rank the predictors by how much variance each one accounts for. Dress it up with relative-weights math if you want to be careful about correlated predictors.¹ Out comes a ranked list, and the list gets a verb: these factors drive the outcome. It feels like a finding. It looks like a lever you can pull.

What it actually is, is a description of who tends to answer the survey similar ways. People who rate their manager highly also tend to rate their engagement highly. That is real, and it is not nothing. But it is silent on the question the slide pretends to answer: if you change the manager, does engagement move?

A driver is a rung you haven't climbed to

Judea Pearl gave this confusion the cleanest picture anyone has — a ladder with three rungs.² The bottom rung is seeing: association, what a regression on observational data can tell you. Who is correlated with whom. The middle rung is doing: what happens if I intervene, if I change this thing on purpose. The top rung is imagining: what would have happened otherwise. The key-driver chart lives entirely on the bottom rung. The word driver is a claim from the middle one. There is no statistical operation — no fancier regression, no machine-learning upgrade, no bigger sample — that climbs the ladder for you. You climb it with a design, or you don't climb it at all.

The statisticians put it even more bluntly decades ago: no causation without manipulation.³ If nothing was changed — if you only watched — you have measured an association, full stop. And associations in observational survey data are exactly where the traps live. The arrow may run backward: engaged people are more generous about their managers, so engagement drives the rating, not the reverse. A third thing may cause both: a team that just shipped something great rates the manager and reports high engagement, and neither caused the other. The "driver" can be entirely an artifact of who happened to be having a good quarter when the survey went out.

What the honest version looks like

None of this means throw out the survey or stop looking for what matters. It means be honest about which rung you're on, and design for the rung you actually need.

If you only have observational survey data, then say associated with, not drives — and use the association for what it's good at: generating hypotheses and flagging where to look, not handing leaders a list of levers. If you need the causal claim — and for a budget decision you do — then you have to do something that earns it. Run an actual experiment where you can: pilot the coaching on some teams and not others, and compare. Where you can't randomize, reach for the quasi-experimental toolkit built for exactly that — a difference-in-differences across teams that changed versus teams that didn't, a before-and-after with a real comparison group. At minimum, measure longitudinally and model the confounders you can name, so the claim has a fighting chance of surviving them. The move that turns a correlation into something close to a driver is always the same shape: change one thing on purpose, hold what you can, and re-measure.

That is more work than a ranked bar chart, and it returns fewer confident answers. You will end up saying we don't yet know whether this is causal; here's the pilot that would tell us more often than the slide deck wants. That is not the analysis failing. That is the analysis being honest about what it can and can't support — which is the only kind of analysis worth betting a budget on.

Why it's worth raising your voice about

This isn't a methodologist's quibble, because the stakes are concrete. A correlation dressed as a driver doesn't just mislead a meeting; it reallocates real money and real careers toward a lever that may not be connected to anything. The recognition program gets funded because recognition correlated with engagement, and a year later engagement hasn't moved, and nobody can say whether the program failed or was never a driver in the first place — because the original analysis couldn't have told you either way. You bought a coefficient and called it a cause.

And the temptation is only getting stronger, because the tools are getting better at producing confident-looking rankings from observational data faster than ever. A model will hand you a beautiful importance plot in seconds. It is still, every time, a description of the bottom rung. The speed doesn't change the logic; it just makes it cheaper to be confidently wrong.

So before you act on the drivers slide, ask the one question that separates the rungs: did anything change, or did we only watch? If you only watched, you have a map of correlations and a list of good hypotheses — genuinely useful, and not a set of levers. Correlation isn't a driver. Calling it one is how an organization spends a reorg on a regression coefficient.

---

Measurement-first, in the principal-issues register: a method piece that stands on its own whether or not you ever work with us. If you do want the causal version done right — designed, confounder-aware, honest about what it can support — that discipline is the Principia measurement program's whole posture, and its companion The Error Bar Is the Product takes up how much to trust a number once you have it. Every footnote names a real, checkable work.

Footnotes

1. Jeff W. Johnson, "A Heuristic Method for Estimating the Relative Weight of Predictor Variables in Multiple Regression," Multivariate Behavioral Research 35, no. 1 (2000): 1–19 — relative weights analysis, the standard refinement behind "key driver" charts when predictors are correlated. It improves the apportioning of explained variance; it does not make the result causal. ↩

2. Judea Pearl & Dana Mackenzie, The Book of Why: The New Science of Cause and Effect (Basic Books, 2018) — the "ladder of causation": association (seeing), intervention (doing), and counterfactuals (imagining). Observational regression answers only the first; no amount of it reaches the second. ↩

3. Paul W. Holland, "Statistics and Causal Inference," Journal of the American Statistical Association 81, no. 396 (1986): 945–960 — the Rubin causal model and the maxim "no causation without manipulation": causal effects are defined relative to interventions, not to passively observed associations. ↩