Correlation Is Not Causation

Correlation means two variables tend to move together — it's a pattern I can measure directly from data. Causation means one variable actually makes the other change, which is a much stronger claim about mechanism. The classic example is that ice-cream sales and drowning deaths are strongly correlated across the year, but ice cream plainly doesn't cause drowning — hot summer weather independently drives both. That example captures why the distinction matters: a correlation can be completely real and completely non-causal. In practice I treat a correlation as the start of an investigation, not a conclusion. It tells me two things are worth looking at together, but before I claim one causes the other I have to rule out reverse causation, a confounding third variable, and plain coincidence.

I can't conclude causation because the users who adopted the feature chose to — this is observational, self-selected data. The most likely problem is confounding: highly engaged users are both more likely to discover and use the feature and more likely to retain anyway, so engagement drives both and inflates the apparent effect. There's also possible reverse causation, where users who were already going to stick around are the ones who explore features. The correlation looks the same under all of these. To actually establish causation I'd run a randomised experiment — an A/B test — where I randomly assign some users to get the feature (or have it defaulted on) and compare retention against a control group. Randomisation balances engagement and every other confounder across the two groups on average, so a retention difference can be credibly attributed to the feature itself. Until I have that, I'd report the relationship as a correlation and explicitly flag that it's not yet causal.

There are four broad explanations. First, A genuinely causes B. Second, B causes A — reverse causation, where the arrow runs the opposite way to what I assumed. Third, a confounder C independently causes both A and B, creating a link between them with no direct connection at all. Fourth, it's coincidence — with enough variables compared, some correlate by pure chance. Telling them apart starts with reasoning: I ask whether each direction has a plausible mechanism, and I brainstorm candidate confounders — is there some third factor, like wealth, age, motivation, or seasonality, that could drive both? A strong correlation with no believable mechanism points toward coincidence or a spurious relationship. But reasoning only narrows it down; the decisive tool is a randomised controlled experiment, because randomisation balances all confounders — even ones I didn't think of — so that a remaining difference is attributable to the treatment. When an experiment isn't possible, I stay tentative and lean on careful confounder control, and I'm explicit that the causal claim is uncertain.