Anatomy of an Experiment
A controlled experiment isolates a cause by manipulating one independent variable under random assignment while holding everything else constant, which is why correlation alone can never establish causation. · 12 min
You want to know whether a study app raises test scores. You could compare people who use it to people who do not — but the people who chose the app may already be the more motivated ones. Any difference you find might be about the students, not the app. An experiment is the tool that solves exactly this problem. It is built to let you say one hard word with confidence: caused.
Guess before you learn
Across many towns, months with higher ice-cream sales also have more drownings. The two rise and fall together, tightly. What is the most likely reason?
Summer heat sends both ice-cream sales and swimming — and therefore drownings — upward together. Neither causes the other. This is the trap the whole lesson is built to escape: two things moving together does not tell you what moves them. Keep your pencil mark.
9–12
3–5
An experiment is a fair test. You change one thing and keep everything else the same, so you know what made the difference. If you change three things at once, you cannot tell which one mattered.
6–8
An experiment answers 'does X cause Y?' by changing X on purpose while holding everything else steady. The thing you change is the independent variable; the thing you measure is the dependent variable. To be fair, you also flip a coin to decide who gets the change — random assignment — so the two groups start out alike. Then any difference at the end points back to the one thing you changed.
9–12
A controlled experiment isolates a cause. You manipulate one independent variable, measure a dependent variable, and use random assignment to sort people into groups so that, on average, the groups differ only in the variable you control. Everything else is held constant. If the groups end up different, the manipulated variable is the credible cause.
A correlation cannot do this. When two things rise and fall together, three explanations stay open: the first causes the second, the second causes the first, or some third variable drives both. Ice-cream sales and drownings correlate because summer heat lifts both. Only manipulation plus random assignment closes off the rival explanations.
K–2
Want to know if a plant grows better with music? Give two plants the same water, same sun, same soil. Only one hears music. If it grows taller, the music helped. Change one thing. Keep the rest the same.
Undergrad
Random assignment is the load-bearing idea. By distributing every other variable — measured or not, known or unknown — equally across conditions in expectation, it converts a difference in outcomes into an unbiased estimate of the manipulated variable's effect. This is what a mere correlation, however strong, can never buy: control over the assignment mechanism itself.
Correlational data leave the assignment to the world, which rarely assigns at random. Confounds — third variables tied to both cause and effect — stay live, as do reverse causation and self-selection. The experiment's power is not that it observes more, but that it engineers the one condition, comparable groups, under which a difference licenses a causal claim.
Postgrad
In the potential-outcomes framework, each unit has an outcome under treatment and one under control, only one of which is observed; the causal effect is their contrast, and the fundamental problem of causal inference is that the counterfactual is missing. Randomization solves it in expectation by making treatment assignment independent of the potential outcomes, so the difference in group means is unbiased for the average treatment effect.
Observational association carries no such guarantee: without ignorability, the estimand is confounded by any common cause of treatment and outcome, and no volume of data removes the bias. 'Correlation is not causation' is precisely the statement that association identifies a causal effect only under assumptions — randomization being the one design that makes those assumptions hold by construction.
confound
A third variable tied to both the suspected cause and the effect, which offers a rival explanation for their link. Summer heat is the confound behind ice cream and drownings.
Here is the discipline in miniature: turn a loose question into a fair test. Follow each step and watch the confounds get closed off one at a time. The move that does the real work is the coin flip — it is what makes the two groups exchangeable, so a difference at the end has only one honest explanation left.
Turn a question into a controlled experiment: does a study app raise test scores? — the steps fade as you master them
Independent variable: study app vs. no app
Dependent variable: score on the same test
Flip a coin: app group and control group
Same test, same time limit, same room
If the app group averages higher, the app is the credible cause
Why is this true?
Why can a correlation never, by itself, prove that A causes B?
Because the same correlation is equally consistent with three different worlds: A causing B, B causing A, or a third variable C causing both. Observation alone cannot tell them apart. Only manipulating A while holding everything else constant closes off the other two.
Hold onto this. From here on, whenever a study claims one thing changes another, you will ask the same two questions: what was the independent variable, and were the groups assigned by chance? Next folio turns to a different limit on experiments — not what they can prove, but what we are allowed to do to the people inside them.
Practice — new ink and old, interleaved
1.Which claim is empirical — something a study could actually test?
2.Which statement could a psychology experiment test?
3.Close the page. Name the three schools in order, and the one question all three were answering.
Introspection, then behaviorism, then cognitive science — each answering: what evidence about the mind can we trust?
How close were you? Grade yourself honestly — it sets your review date.
4.A study gives one group a caffeine pill and another a look-alike sugar pill, then times both on a puzzle. What is the dependent variable?
5.Without looking back: what makes a claim about the mind 'empirical'?
It predicts something observable that others could test and repeat, so the world could prove it wrong.
How close were you? Grade yourself honestly — it sets your review date.
6.Towns with more bookstores have higher average lifespans. What is the safest conclusion?
7.Put the steps of a controlled experiment in order, first to last.
- Assign people to groups by chance
- Give one group the treatment, the other none
- Hold all other conditions constant
- Measure the outcome and compare the groups
8.From the last folio: what one demand turned psychology into a science?
That claims about the mind be tested against controlled, repeatable evidence rather than settled by argument.
How close were you? Grade yourself honestly — it sets your review date.
9.Match each school to how it gathered its evidence.