belief & evidence · metaphor 42 of 100
People who exercise are happier — so will exercising make you happier? Watching the world and moving it are different mathematical objects: P(happy | exercise) and P(happy | do(exercise)) can be different numbers, and every self-help book quietly swaps one for the other.
The observation is real. Joiners of morning gyms do have better marriages; readers of poetry do live longer; early risers do out-earn the rest of us. Nobody fudged the spreadsheets. But observation answers exactly one question — what are people-who-do-X like? — and the people who do X were different before they ever did it. They had the health, the mornings, the discretionary energy that made doing X thinkable. Watching faithfully reports the company X keeps. It says nothing, yet, about what X does.
To learn what X would do to you, you need a stranger object: X severed from everything that usually causes it — chosen by nothing, correlated with nothing, arriving like weather. That severing is what the word do means in Judea Pearl's notation, and in general it cannot be computed from watching alone. Except when it can — and knowing exactly when a seen world contains the answer to a done question is his whole calculus.
Vitality is unobserved in this world's dataset. No formula, however clever, can adjust for a confounder you never measured — from these columns alone, the mirage and the truth are indistinguishable. This grayness is honest.
Stratify the SEE-mode data into 10 vitality bands, compare exercisers with non-exercisers inside each band, then reweight by band size: backdoor estimate — · true effect —
Honest toy: 2,000 simulated people, vitality spread evenly from 0 to 1, fixed random seed. exercise chance = 0.5 + confounding·(vitality−0.5) · happiness chance = 0.10 + 0.55·vitality + effect·(exercise). Every readout is computed live from the dots you see; the faint "expectation" is the same quantity with the sampling noise integrated out.
The two questions
P(happy | exercise) is a filter: take the world exactly as it is, keep only the people who happen to exercise, and average. P(happy | do(exercise)) is a surgery: reach into the mechanism, force exercise on everyone regardless of who they were, and let the consequences flow downstream. The filter inherits every reason people exercise; the surgery deletes those reasons. In the world above, vitality feeds both exercise and happiness — so the filtered comparison credits exercise with vitality's work, arriving through the back door.
Ordinary language cannot hear the difference. "If I exercise, will I be happier?" means both questions at once: if I turn out to be an exerciser (a fact about which kind of person you are) and if I take up exercising (an act that changes nothing about who you were yesterday). Grammar gives you one conditional where reality needs two, which is why the swap in every self-help book goes unnoticed — including by its author.
What to try
Manufacture the mirage. Set the true effect to exactly zero and confounding full positive. SEE mode swears by exercise — a gap near +18 points, real people, real arithmetic. Toggle DO: the arrow is cut, everyone flips a coin, and the same comparison collapses to noise around zero. Nothing about the effect changed. Only the question did.
Find the masked cause. Slide confounding to −1 — a world where doctors send the depleted to the gym — and set the true effect near +18 points. Watching now reads roughly nothing: a genuine cause, erased from view because the people doing it started worse off. Nudge the effect down to +10 and observation flips sign entirely — the watchers would tell you to quit the thing that is helping you.
Resurrect truth from pure watching. Stay in SEE mode, in the mirage, and switch vitality measured on. The backdoor panel stratifies the observational data by vitality, compares like with like, reweights — and recovers the true effect without a single coin flip. Then switch it off and read the gray text. Seeing can become doing, but only when you can name and measure what feeds the habit.
The arrow-cutting
Randomization is graph surgery. Your decision to exercise carries your whole biography — health, temperament, free mornings, the kind of childhood that made discipline feel like home. A coin carries nothing. Assigning by coin flip deletes every arrow pointing into exercise, so whatever difference remains between the groups can only have flowed out of it. That is the entire epistemic magic of the randomized trial, and it is why five hundred coin-flipped strangers can outweigh a million sincere testimonials: a testimonial is SEE-mode data delivered in DO-mode grammar.
Life occasionally performs the surgery for free. A job transfer you didn't seek, a gym that closes, a lottery, a policy that draws a line through a birthday — genuinely unchosen change is an arrow cut without a coin, and it is worth more evidence about yourself than a decade of chosen habits, precisely because you didn't choose it.
Everyday do-calculus
The discipline is the reflex. Every "successful people do X" is a report of P(success | X); what you are being sold is P(success | do(X)); and the entire difference between them is the set of arrows pointing into X. So interrogate the advice: what feeds the habit? If the five-a.m. run is fed by the same vitality that feeds the promotion, copying the run buys you the sweat and not the promotion. Ask of any study, any memoir, any envied acquaintance: is this a dispatch from SEE mode or from DO mode? And if SEE — what graph would I have to draw, and which of its confounders could I actually measure, before this correlation earned the right to be spoken as a cause?
The mapping
| Mathematics | Life |
|---|---|
| the causal graph | Your story of what causes what in the situation — drawn honestly, arrows and all, before any numbers arrive. |
| the confounder | Whatever made the doers different before they did it: the vitality behind the gym habit, the curiosity behind the poetry. |
| P(Y | X) | What the successful are like — a faithful portrait of the company X keeps. |
| P(Y | do(X)) | What copying them would actually do to you — the only number advice ever really promises. |
| cutting the arrow | Randomization, or any genuinely unchosen change — X arriving with none of your biography attached. |
| backdoor adjustment | How careful watching can sometimes earn the right to causal talk: name the confounders, measure them, compare like with like. |
Where the metaphor tears
The instrument hands you the diagram; life never does. Do-calculus is exact given the graph, and the graph is where the fight actually lives — does vitality drive exercise, or exercise vitality, or both? The formalism disciplines a causal argument beautifully; it does not replace having one. Two honest people with different diagrams will compute different truths from the same data, and no theorem adjudicates between their drawings.
The operator do(exercise) changes exercise and nothing else — a surgeon's incision. But deciding to exercise changes your sleep, your friends, your self-image, your calendar, your identity as someone-who-follows-through. Real resolutions are fat-hand interventions that jostle a dozen variables at once, so even the cleanest personal experiment on yourself is only an approximation of the do() it aspires to be.
You cannot randomize childhoods, marriages, religions, or the century you were born into — and these are precisely the causes we most want to know. Where the coin flip is impossible and the confounders unmeasurable, the honest options are natural experiments, triangulation, and humility. The calculus's deepest lesson may be negative: knowing exactly which questions your data cannot answer, no matter how much of it you gather.