How to Learn Science

This sounds trivial and is the foundation of everything else. The plant bent towards the window is an inference. What you observed was that the plant is bent and the window is over there. The explanation arrived so fast it felt like part of the seeing.

Almost every argument about evidence is really an argument about which inferences a set of observations supports. Being able to state the observation without the explanation attached is the skill, and it takes deliberate practice because the brain does not naturally keep them apart.

The single most useful question to ask about any claim is what result would count against it. If the answer is that nothing would, the claim is not being tested, whatever else is happening.

This is the cleanest line between science and things shaped like science. Astrology, most conspiracy theories and a good deal of confident business advice share the property that every possible outcome confirms them. That is a defect and not a strength, though it is regularly presented as one.

Applied to your own work, the discipline is to write down what you expect to see and what would make you abandon the idea, before running anything.

Everyone learns that a fair test changes one variable and controls the rest. The part that gets skipped is that identifying all the relevant variables is the actual difficulty, and it is where real experiments go wrong.

The habit worth building is asking what else differs between the groups besides the thing you intended. In school experiments that is usually temperature or timing. In studies about people it is almost always something about who ended up in which group, which is why randomisation matters so much and why observational studies are so much weaker than they look.

A measurement without an uncertainty is an overclaim. Saying a rod is exactly three metres asserts something no instrument supports, and treating it as exact carries that false confidence into everything downstream.

Accuracy and precision are different properties and both matter. A balance that reads 5.001 kilograms every time for a three kilogram mass is beautifully precise and completely wrong. Repeatability is not correctness, which is worth remembering whenever someone cites a consistent result.

The most common error in reading science news is treating a single result as settled. Individual studies are small, often underpowered, sometimes wrong, and occasionally fraudulent. What matters is whether the result holds up when someone else tries it.

Statistical significance is the phrase most abused here. It does not mean important, large, or true. It means the result would be unlikely if nothing were going on, which is a much weaker claim than the headline usually implies, and it says nothing about effect size.

The practical question when you meet a claim is not whether a study found it. It is whether anyone has replicated it, and how large the effect actually is.

Presenting data is a set of choices, and the choices change what people conclude. A bar chart implies separate categories; a line implies the values between the points mean something. Using a line for unrelated categories asserts a relationship that does not exist.

The same applies to axes. A truncated y-axis makes a trivial difference look dramatic, which is why it is so common in advertising. Learning to notice this in other people's charts is the same skill as not doing it in your own.

This track is worth doing even if you never study a science again, and it is worth doing early rather than last. Every other subject on this site makes claims, and the units here are about how to decide which claims to accept.

If you only do one thing from this guide: when you next meet a confident claim, ask what would have to be true for it to be wrong. If nothing would, you have learned something about the claim.