Chapter 7: The Ladder of Why
Ground floor
Every summer, ice cream sales rise, and so do drownings. The numbers move together, year after year, as reliably as anything in the ledger. Ban ice cream, save swimmers? The child laughs, because the child can see the sun: hot days send people to both the freezer and the lake. The two numbers are married, but neither causes the other; they share a parent.
Now notice what the laugh knows. The child has, without instruction, distinguished three completely different worlds that produce identical data: ice cream causes drowning; drowning causes ice cream; something else causes both. No amount of staring at the sales-and-drownings table, no cleverness of arithmetic performed upon it, can tell those three worlds apart, because the table is the same in all three. The difference between them is not in the data. It is in what would happen if you reached in and changed something, and the table was written by a world where nobody reached in.
That is the whole scandal of this chapter, and it took statistics a century to face it: seeing and doing are different kinds of knowledge, and the second cannot be computed from the first without assumptions about the arrows. Judea Pearl, who forced the issue, calls it a ladder. Rung one: seeing, what goes with what. Rung two: doing, what happens if I intervene. Rung three: imagining, what would have happened had I acted otherwise. Correlation lives on rung one. Every question you actually care about lives on rungs two and three.
The stairs
The bookkeeping tool is embarrassingly simple: circles and arrows. Draw a node for each quantity, an arrow for each direct causal influence you are willing to assert. Sun → ice cream. Sun → drowning. No arrow between the last two. Such a picture is called a causal graph, and its power is not that it is true, you drew it, it encodes your assumptions, but that once drawn, it mechanically answers the question that intuition fumbles: which correlations in my data are causal, and what must I adjust for to isolate them?
The villain has a name: a confounder, a common cause like the sun, whose arrows create a spurious path between two variables, a “backdoor” through which correlation flows without causation. The classical fix is to adjust for it: look within hot days and within cool days separately, and the ice-cream-drowning correlation evaporates. So far, so familiar; “control for everything” is the folk wisdom.
Now watch the folk wisdom kill a patient. Here is a real dataset, from a published study of two treatments for kidney stones, and the numbers are worth reading twice. Treatment A succeeded in 93 percent of small-stone cases (81 of 87) and 73 percent of large-stone cases (192 of 263). Treatment B: 87 percent on small stones (234 of 270) and 69 percent on large (55 of 80). A wins on small stones. A wins on large stones. Now pool them: A succeeded 273 times in 350 tries, 78 percent; B succeeded 289 times in 350, 83 percent. B wins overall. A is better for every kind of patient there is, and worse for patients in general. This is Simpson’s paradox, and it is not a trick of small numbers; the totals are identical, 350 and 350.
The resolution is not in the arithmetic, which is flawless, but in the arrows. Doctors gave the harder treatment, A, to the harder cases: stone size → treatment choice, and stone size → outcome. Severity is a confounder, and pooling lets it pour its influence through the backdoor: B looks better overall only because B got the easy patients. The subgroup numbers are the causal truth; the pooled number is an artifact of who got assigned what. And here is the sentence that should chill you: the data alone cannot tell you that. Flip the story, make the third variable a consequence of treatment instead of a cause of it, say, a side effect that itself influences recovery, and with the very same table, the pooled number would be the honest one and the subgroups the artifact. Same numbers, opposite conclusion, and the difference lives entirely in the arrows, which live entirely outside the spreadsheet. Anyone who says “let the data speak” has simply chosen not to hear the assumptions doing the talking.
Two more pieces complete the toolkit. First, the meaning of a controlled experiment, which can now be stated in one clean sentence: randomization is a machine for deleting incoming arrows. When a coin assigns the treatment, nothing else, not severity, not self-selection, not the sun, can influence who gets what; every backdoor is sawn off at once, including the ones you never thought of. That is the entire magic of the randomized trial, and it is why the control group in the brain-training autopsy was not a bureaucratic nicety: it is the only tool that severs “chose to train” from “improved,” because motivated, freshly-rested, regressing-to-the-mean humans assign themselves, and self-assignment is an arrow.
Second, the trap on the far side, because the folk wisdom fails in both directions. Adjusting can create correlation as well as remove it. A variable caused by two others, a collider, arrows pointing in, acts as a dam: leave it alone and no association flows; condition on it and the dam opens. The classic instance: among hospitalized patients, unrelated diseases correlate negatively, because having either one is enough to get you through the door, so within the hospital, having one makes the other less necessary as an explanation. Select on an effect and its causes start trading off against each other in your sample while remaining strangers in the world. Every dataset gathered from people who chose to show up, every user base, every survey of volunteers, every retention cohort, is conditioned on a collider called showing up, and correlations inside it are suspects, not witnesses.
So the chapter’s law: whether to adjust for a variable is not a statistical question. It is a causal one, answered by the arrows, and the arrows are an argument you must make in the open, where it can be attacked. The spreadsheet will hold perfectly still while you get it wrong.
The locked door
Behind the door: the do-calculus, three rules that completely characterize when rung-two questions are answerable from rung-one data plus a graph, with a completeness proof; instrumental variables, nature’s accidental randomizers, and Mendelian randomization, which uses your genome as the coin flip; counterfactuals and the rung-three machinery of “would have”; the rival Rubin potential-outcomes school and the long, productive feud between the frameworks; and causal discovery, the audacious attempt to infer the arrows themselves from data. The reading order is Pearl and Mackenzie’s The Book of Why for the full story told with a grudge, then Hernán and Robins’ Causal Inference: What If, free online, when the machinery is wanted whole.