Ice-cream sales correlate with drowning deaths — what can and cannot be concluded from that?
answer
- the number describes co-movement only
- at least four rival explanations
- r(x, y) equals r(y, x)
- seeing versus doing
- prediction earned, intervention not
basics
~20 sA correlation only says the two series move up and down together in the observed data. It cannot say ice cream causes drownings: something else, such as hot weather, can drive both, and correlation carries no direction while causation does.
solid answer
~40 sThe association is real — weeks with high ice-cream sales really do tend to be weeks with more drownings — but the coefficient describes co-movement in observed data and nothing more. Three explanations stay open: sales drive drownings, drownings drive sales, or some other factor such as hot weather raises both swimming exposure and dessert buying. Pearson's r is also symmetric, `r(x, y) = r(y, x)`, so the number itself picks no direction, while every causal claim has one. The honest reading is predictive: if I observe a high-sales week I should expect a higher drowning count. The claim it does not license is interventional: cutting ice-cream sales would not be expected to change drownings. Getting a causal claim needs randomised assignment or an explicit design, not a bigger r.
go deeper
Be ready to state in one sentence that the coefficient describes how two columns move together, and to name at least one concrete alternative story for the pair you are shown.
Explain the mechanics: r is symmetric and scale-free, so it cannot encode direction, and reverse causation plus a shared driver both remain compatible with any value it takes.
Show the judgment to separate prediction from intervention on real work — a feature that forecasts an outcome is not automatically a lever, and shipping a change on that confusion is a common incident.
Own the reporting standard: decide what verbs your organisation is allowed to use for observational findings, and what evidence upgrades a dashboard association into a decision.
## What the number actually is Pearson's correlation coefficient `r` is a standardised measure of how two variables co-vary in a sample. Written out, `r = cov(x, y) / (sd(x) * sd(y))`, where `cov(x, y)` is the average product of each pair's deviations from its own mean. It ranges from -1 to 1, and it answers exactly one question: **as one variable sits above its average, does the other tend to sit above its average too, in a straight-line way?** That is a statement about a table of numbers that was passively observed. It contains no information about what would happen if someone reached in and changed one of the columns. ## The classic phrasing trap Weekly ice-cream sales and weekly drowning deaths correlate strongly in most temperate countries. The correlation is not an artefact and not noise — it will replicate. What is wrong is the sentence people build on top of it: 'ice cream causes drowning'. From an observed association between two variables A and B, at least four accounts remain live: 1. **A causes B.** Plausible for some pairs, absurd here. 2. **B causes A.** Reverse causation. Often the embarrassing one: hospitals with more staff per bed showing higher mortality usually reflects sicker patients being routed there, not staff killing patients. 3. **Something else moves both.** Hot weather sends people both to the beach and to the ice-cream counter. Neither series acts on the other; they share an upstream driver. 4. **Selection or coincidence.** The sample was assembled in a way that manufactures the link, or with enough variable pairs tested, some correlate by chance alone. Nothing in `r` distinguishes among these. Two data columns and their correlation are compatible with all four worlds. ## Why symmetry matters `r(x, y)` and `r(y, x)` are the same number by construction — swap the roles and the formula is unchanged. Causation is not symmetric: 'smoking causes cancer' and 'cancer causes smoking' are different claims with different consequences. A statistic that cannot tell the two apart cannot be the evidence for either. This is the cleanest one-line argument in an interview, and it is stronger than reciting the slogan. ## Prediction versus intervention The useful distinction is between **seeing** and **doing**. Seeing a high-sales week and expecting more drownings is legitimate: the correlation supports conditional prediction, and a forecasting model may use ice-cream sales as a feature with no apology. Doing something — cutting sales in half to save lives — relies on the relationship surviving the intervention, and there is no reason it would, because the mechanism that generated it (weather) is untouched. Many production incidents come from the second sentence being written when only the first was earned: a metric that predicts churn is turned into a lever to reduce churn, and moving the lever moves nothing. ## What would license the causal claim Randomised assignment is the direct route: assign the exposure by a coin flip so the groups differ, on average, in nothing but the exposure, then the observed difference in outcomes has no other systematic explanation. Where randomisation is impossible, a causal claim needs an explicit design and stated assumptions — not a larger sample and not a larger `r`. Sample size shrinks the uncertainty around the correlation; it does nothing to the question of what generated it. ## Interview language that scores Weak candidates recite 'correlation is not causation' and stop, which reads as a memorised phrase. Strong candidates do three things: name the specific rival explanation for the case in front of them, say what the association *is* good for (prediction, screening, hypothesis generation), and state what evidence would upgrade it. In writing, keep the register honest: 'associated with', 'predicts', 'is higher among' for observational findings; reserve 'causes', 'drives' and 'reduces' for designs that support them. That discipline matters more in practice than the ability to name the fallacy.
- If ice-cream sales predict drownings, is it wrong to use them as a feature in a forecasting model?No. Prediction only needs the association to hold in data drawn the same way as the training data, and it does. The mistake would be treating that feature as a lever: reporting that its coefficient tells you how many lives a marketing campaign would cost. Use it to forecast, do not use it to justify an intervention.
- Does a very strong correlation, say r = 0.95, make a causal reading safer?No. Strength and causality are unrelated axes. A shared driver can produce an almost perfect association between two effects that never touch each other, and a genuine cause with a lot of other noise around it can show a weak correlation. Strength tells you how tightly the points hug a line, not what generated the line.
- What would you need in order to state that one of these actually causes the other?A design, not a bigger dataset. The cleanest is randomised assignment of the exposure, so the compared groups differ systematically in nothing else. Failing that, you need an explicit argument for why no other explanation survives, stated as assumptions that a reader can attack. Neither of those comes from recomputing the correlation.
A thermometer and a heating bill rise together all winter. Reading the thermometer tells you what the bill will look like; smashing the thermometer does not heat the house.
saying these in an interview costs you the question
- Says a high enough correlation proves causation
- Treats a predictive feature as an actionable lever
- Claims a bigger sample turns correlation into causation
- Recites the slogan without naming a rival explanation
- Forgets that reverse causation is also on the table
- Uses causal verbs to report an observational association